EmbodiedIntentBench: Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions

EmbodiedIntentBench: Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions Rajat Bhattacharjya1,* , Changwoon Lee2 , Hyeonjong Ju2 , Sing-Yao Wu1 , Hamidreza Alikhani1 , Zhuoyi Yang1 , Ian Harris1 , Nalini Venkatasubramanian1 , Nikil Dutt1 1 University of California, Irvine, USA

2 Yonsei University, South Korea

arXiv:2611.10498v1 [cs.RO] 7 Nov 2026

Abstract Mobile robotic sensing systems must adapt to operator requirements, limited resources, changing connectivity, and observation opportunities shaped by physical motion. Evaluating runtime strategies therefore requires assessing whether their decisions fulfill mission objectives, beyond isolated perception accuracy or execution efficiency. We present EmbodiedIntentBench, a benchmark for intent-driven runtime adaptation in mobile robotic missions. It represents mission intent through structured contracts for coverage, temporal evidence, reporting, urgency, and resources, and couples simulated unmanned aerial vehicle (UAV) workloads with causal, resource-accounted execution. Computation decisions affect mission time, resource use, and which observations can contribute evidence. We evaluate fixed and adaptive perception strategies in multi-phase missions and under changing urgency, connectivity, and battery conditions. In the multi-phase Depot study, a configuration with 87.5% full-bank target recall across visible target observations completes only 1 of 16 missions, whereas a lower-recall configuration completes 14. In an ablation, ignoring executioninduced observation loss increases aggregate mission success and reverses the ordering of two runtime strategies. Replays using Jetson AGX Orin CPU and GPU profiles change the relative success of local and remote strategies. In selected Jetson GPU-profile replays, intent-driven adaptation reduces modeled task energy by up to 28% with unchanged mission success. EmbodiedIntentBench provides an extensible framework for evaluating runtime decisions against mission fulfillment under physical and resource constraints.

Figure 1: Motivation and overview of EmbodiedIntentBench. (a) Operatorspecified missions impose requirements for target coverage, evidence verification, timeliness, and resource use. (b) Local, split, and remote execution offer different computation and communication tradeoffs; the current evaluation considers whole-model local and remote execution. (c) Perception quality and execution latency influence mission evidence, as longer processing times can prevent subsequent observations from being processed. (d) EmbodiedIntentBench connects structured mission requirements with physically situated, resource-accounted execution and mission-level evaluation.

Consider a UAV deployed to locate stranded individuals following a flood, where an operator requests verified locations to support rescue operations (Fig. 1(a)). An initial search may prioritize broad target coverage, while subsequent rescue operations may require corroborated findings or more timely updates. Although these requests concern the same environment, they impose different evidence, timing, and resource requirements [5]. Mission intent describes an operator’s desired outcome and is represented through a structured contract, which can be supplied directly or constructed from a supported natural-language request. Fulfilling such requirements cannot be assessed through perception accuracy or execution efficiency alone. A correct prediction may arrive too late, while a timely detection may lack sufficient evidence for confirmation. Even with sufficient evidence, reporting errors or resource violations can prevent mission completion. Consequently, successful individual computations do not necessarily translate into successful mission execution [6, 7]. These distinctions become particularly important for mobile robotic systems, where physical motion creates time-sensitive observation opportunities. As a UAV traverses its environment, useful views emerge and disappear with changing viewpoints, occlusions, and target visibility. A computationally demanding perception configuration may improve an individual prediction while preventing subsequent observations from being processed (Fig. 1(c)) [8, 9]. Such

Keywords embodied robotic missions, intent-driven adaptation, runtime computing systems, mission-level benchmarking, cyber-physical systems

1

Introduction

Mobile robotic systems, particularly unmanned aerial vehicles (UAVs), are increasingly employed in mission-critical applications such as search and rescue, disaster response, and infrastructure inspection [1–3]. Advances in language-guided robotics and intentdriven embodied intelligence are enabling operators to specify what a mission must accomplish, rather than prescribing how individual sensing and computation tasks should be executed [1–4]. However, translating such high-level natural-language intent into effective runtime behavior remains challenging under constrained onboard computation, limited energy, and variable network connectivity [1]. ∗ Correspondence: rajatb1@uci.edu.

Paper is currently under review. Authors’ version posted for personal use and not for redistribution. Copyright may be transferred without notice, after which this version may no longer be accessible. 1

Bhattacharjya et al.

• Intent-driven mission specification. We introduce structured mission contracts expressing operator objectives through explicit requirements for target coverage, temporal evidence, verification, reporting, urgency, and resource constraints. • Causal, resource-accounted execution. We develop an evaluation methodology coupling runtime computation with physically situated observation opportunities, accounting for how processing delays, communication, and onboard resource consumption affect mission evidence. • Mission-level evaluation and diagnostics. We connect runtime decisions to evidence acquisition, confirmation, reporting, and mission fulfillment, enabling analysis of the requirements and execution conditions underlying mission outcomes. • Extensible benchmark and empirical characterization. We provide configurable workloads, reference strategies, and extension interfaces, demonstrating discrepancies between perception performance and mission success, the effects of causal execution, and the consequences of changing resource conditions.1

tradeoffs also depend on whether computation executes onboard or remotely (Fig. 1(b)), with different implications for latency, communication, and energy consumption [10–12]. Runtime decisions must therefore be evaluated through their contribution to mission fulfillment, rather than isolated measures of prediction quality or execution efficiency. Prior work has established the importance of evaluating robotic computation in the context of physical execution and downstream task performance [13]. MAVBench [14] characterizes interactions among computation, physical dynamics, and resource consumption, while RobotPerf [15] and A2Perf [16] provide robotic computing benchmarks. Streaming perception [17, 18] and control-aware prediction [19] highlight the importance of prediction timing and downstream consequences. More recently, DEEP [20] and LIBERODynamic [21] examine decision-level computation value and inference latency. However, these approaches do not jointly evaluate runtime computation against explicit mission requirements involving accumulated evidence, temporal verification, reporting, and resource constraints. This motivates a fundamental question: How can runtime sensing and computation strategies be systematically evaluated based on their ability to fulfill operator-specified mission requirements under changing physical and resource conditions? Addressing this requires explicit mission requirements, causal accounting of computation and observation opportunities, and evaluation of whether accumulated evidence and reports satisfy the intended mission. To this end, we present EmbodiedIntentBench, a benchmark for evaluating intent-driven runtime adaptation in mobile robotic missions (Fig. 1(d)). EmbodiedIntentBench operationalizes mission intent through structured mission contracts specifying target coverage, temporal evidence, verification, reporting, urgency, and resource constraints. It combines Gazebo-based UAV workloads [22] with causal, resource-accounted execution; using shared recorded trajectories, it measures how runtime computation changes observation availability, resource use, and mission fulfillment. The current evaluation isolates computation decisions over fixed recorded UAV trajectories; navigation and physical control are outside its scope. We demonstrate EmbodiedIntentBench through two complementary case studies. The first examines perception choices in multi-phase Lake and Depot missions (lake and warehouse) involving coverage, repeated confirmation, timely reporting, and fresh revisits. The second investigates runtime adaptation under changing urgency, connectivity, and battery conditions using whole-model local and remote execution. We also examine how mission requirements can guide when perception stops, using Jetson-derived profiles and measuring energy for selected schedules on the device. Our evaluation reveals that a configuration achieving 87.5% full-bank target recall (across visible target observations) completes only 1 of 16 missions, whereas another with lower recall completes 14. Further validation shows that ignoring execution-induced observation loss can overestimate mission success and reverse the relative ranking of runtime strategies. The main contributions of this paper are as follows:

2

EmbodiedIntentBench: Design and Architecture 2.1 Architecture Overview EmbodiedIntentBench connects operator-specified mission requirements to runtime computation and mission-level outcomes, as illustrated in Fig. 2. A structured mission contract defines successful execution, while a runtime policy selects computation configurations based on mission requirements and current system state. Causal, resource-accounted execution determines result availability, resource consumption, and subsequent observation opportunities. Admitted outputs contribute to mission evidence and reporting, which are evaluated against the contract. This separation enables controlled comparisons of runtime strategies without prescribing a particular perception model or adaptation algorithm.

2.2

Intent Specification and Mission Contracts

Intent specification ( 1 – 2 ). The first component of EmbodiedIntentBench establishes what constitutes successful mission execution. As illustrated in Fig. 2(a), an operator-facing request is represented through a structured mission contract that makes the intended sensing objectives and execution requirements explicit. For example, a request to locate stranded individuals and report verified positions can specify the required target coverage, corroborating observations, and reporting constraints. The contract provides a common interface between operator intent, runtime policies, and mission evaluation. Requirements are expressed through a validated JSON-based specification, which can be supplied directly or constructed from supported operatorfacing descriptions. The benchmark operates on the structured requirements rather than assuming unrestricted natural-language interpretation during execution. 1 A public artifact provides code, inputs, and reference results for the dynamic study and

the Jetson and intent-driven analyses; Appendix A describes its scope. The complete benchmark artifact, including the multi-phase workloads and prediction banks, will be released upon acceptance. Artifact: https://github.com/rajatbh21/EmbodiedIntentB ench-artifact 2

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions

Figure 2: Architecture of EmbodiedIntentBench. (a) Intent specification: An operator request ( 1 ) is represented through a structured mission contract ( 2 ) defining evidence, temporal, reporting, and resource requirements. (b) Mission workloads and runtime decisions: Physically situated observations ( 3 ) and current system state inform a runtime policy ( 4 ) that selects a computation configuration. (c) Causal execution and resource accounting: A resource-accounted executor ( 5 ) applies the selected computation and its associated costs, while the execution clock and resource state ( 6 ) determine result availability and subsequent observation opportunities. (d) Mission evidence and evaluation: Admitted results contribute to evidence accumulation and reporting ( 7 ), while a mission evaluator ( 8 ) assesses fulfillment against the contract. Solid arrows indicate runtime execution and information flow, including state feedback; red dashed arrows indicate private ground truth accessible only to the evaluator.

We represent a mission contract as C = (G, D, Q, B),

limits on incorrect or duplicate reports. These requirements distinguish identifying a target in an observation from producing a valid mission report. Resource constraints (B). The contract defines applicable resource limits, such as onboard energy budgets, required battery reserves, and communication constraints. These limits are distinct from the instantaneous resource state observed during execution. The mission contract thus defines both the requirements exposed to a runtime policy and the criteria used to evaluate its execution. However, the actions available to a policy also depend on the observations and operating conditions encountered during the mission.

(1)

where G defines mission goals and evidence requirements, D specifies temporal constraints, Q captures reporting obligations, and B represents resource constraints. Mission goals and evidence (G). The mission specification identifies the targets or regions of interest and the evidence required for their successful identification. Requirements may include target coverage, repeated observations, minimum temporal separation between confirmations, and evidence collected during distinct mission phases. For multi-phase missions, a later phase may require fresh observations, preventing previously collected evidence from automatically satisfying a new requirement. Temporal requirements (D). Temporal constraints operate at different levels, including individual perception requests, mission phases, and final reporting. For a perception request released at time 𝑟𝑖 , an execution allowance Δ𝑖 defines its deadline: 𝑑𝑖 = 𝑟𝑖 + Δ𝑖 .

2.3

Mission Workloads and Runtime Decisions

Physically situated workloads ( 3 ). EmbodiedIntentBench represents mobile robotic missions through time-indexed observations associated with physical execution. A workload specifies the observation opportunities presented to the runtime, together with the physical context and evaluation annotations required to determine mission outcomes. In the current implementation, Gazebo-based UAV missions [22] provide recorded trajectories with synchronized RGB observations, vehicle poses, and acquisition timestamps. A workload is represented as

(2)

Urgency may evolve during execution, changing the allowance assigned to newly released requests without necessarily changing the underlying sensing objective. Phase and reporting deadlines independently constrain when the required evidence and findings must become available. Reporting obligations (Q). Reporting requirements specify which findings must be delivered, their temporal validity, and admissible

𝑁 W = {(𝐼𝑖 , p𝑖 , 𝑡𝑖 )}𝑖=1 ,

(3)

where 𝐼𝑖 is the image acquired at time 𝑡𝑖 and p𝑖 denotes the corresponding vehicle pose. Ground-truth annotations are retained separately for evaluation. 3

Bhattacharjya et al.

These observations define a common set of potential observation opportunities across competing strategies. Their actual processed subsets, however, depend on the execution decisions made during the mission. Holding the physical trajectory fixed within a comparison isolates the effect of runtime computation on observation availability without conflating it with changes in navigation behavior. To support reproducible comparisons, the implementation constructs a bank of predictions by evaluating candidate models on recorded observations. During execution, only the prediction associated with the selected model and observation can enter downstream processing, subject to its execution and timing requirements. Unselected predictions remain unavailable to the runtime policy. Runtime policy and computation selection ( 4 ). At each eligible decision point, a runtime policy selects from the computation configurations supported by the workload. Let 𝑠𝑖 represent the public operating state and A the available action space. A policy 𝜋 selects 𝑎𝑖 = 𝜋 (𝐼𝑖 , C, 𝑠𝑖 ) ∈ A,

Execution times, transferred data, and onboard energy are charged according to the declared execution and resource profiles. Importantly, resource accounting is not restricted to successful inference. A late computation retains its incurred processing cost even when its result cannot contribute timely evidence. This ensures that different runtime strategies are compared against the resources consumed by their executed decisions, rather than only the costs of successful outputs. Causal execution and observation availability ( 6 ). The benchmark maintains a logical execution clock that advances with the operations performed by the runtime. Execution time determines when a selected result becomes available and, under the serial processing rule, when the next observation can be admitted. Let 𝑓𝑖 denote the completion time of the blocking processing interval initiated for observation 𝑖. The next eligible observation is 𝑗 = min{𝑘 > 𝑖 | 𝑡𝑘 ≥ 𝑓𝑖 }.

If no such observation remains within the workload, processing reaches the end of the available observation stream. Captures acquired during the intervening interval are recorded as unprocessed opportunities rather than queued for later execution. This serial execution model represents blocking perception pipelines, while alternative buffering and asynchronous processing can be incorporated through corresponding execution backends. Result admission is determined separately by its applicable timing requirements. A prediction that becomes available after its request deadline cannot contribute as a timely result, although the work already performed remains accounted for. Thus, the execution framework connects the latency of a selected configuration to both its immediate result and the subsequent observation opportunities available to the mission. Resource-state evolution. In addition to execution time, the benchmark tracks the resources consumed throughout a mission. Onboard energy over a mission horizon 𝑇 is modeled as ∫ 𝑇 𝐸 onboard (𝑇 ) = 𝑃flight (𝑡) 𝑑𝑡 (6) 0 + 𝐸 compute + 𝐸 comm + 𝐸 support,

(4)

where the action may specify a perception model, execution settings, computation placement, or other supported processing choices. The operating state may include current network conditions, available onboard resources, request deadlines, and declared execution profiles. Where supported, information derived from previously completed, selected actions may also inform subsequent decisions. Ground-truth target identities and correctness labels are reserved for the evaluator. Future observations and outputs of unselected perception configurations are likewise inaccessible through the runtime policy’s decision interface. For the evaluated perception-routing configurations, an execution action is instantiated as 𝑎𝑖 = (𝑚𝑖 , ℓ𝑖 ), where 𝑚𝑖 identifies a perception configuration and ℓ𝑖 denotes local or remote execution. The action space also permits abstention, denoted by 𝑎𝑖 = ⊥. Other execution arrangements, including partitioned inference, can be incorporated through configurations with their own processing stages and execution profiles. This interface separates the decision strategy from the consequences of its actions. A policy may optimize predicted quality, timeliness, resource feasibility, mission progress, or a combination of these objectives. Whether its selected computation contributes useful evidence, however, depends on how that computation executes over time.

2.4

(5)

where the terms represent flight, onboard computation, communication, and supporting operations, respectively. Supporting work may include preprocessing, routing, loading, decoding, association, and verification, according to the execution profile. Costs are charged at their declared execution boundaries to avoid duplicate accounting. The recorded flight duration is charged in full, and the framework tracks remaining energy, required reserve, and communication consumption as execution proceeds. The resulting time and resource state feeds back into subsequent policy decisions. Consequently, the selected action affects not only the current prediction but also the computation and observation opportunities available later in the mission. The outputs that remain usable are then passed to the mission evidence pipeline.

Causal Execution and Resource Accounting

Resource-accounted execution ( 5 ). The execution framework applies the time and resource costs associated with a selected action, as illustrated in Fig. 2(c). Each execution configuration specifies the processing operations and associated costs required to produce its output. In the current implementation, local execution incurs the configured onboard inference latency and energy consumption, together with applicable preprocessing, model loading, switching, and supporting operations. Remote execution additionally accounts for input preparation, upload, server-side processing, response transfer, and decoding. Communication is evaluated against the applicable network profile, including bandwidth changes and outage intervals.

2.5

Mission Evidence and Evaluation

Evidence accumulation and reporting ( 7 ). The evidence pipeline connects individual perception results to the requirements of a complete mission, as illustrated in Fig. 2(d). Admitted predictions are processed through the applicable downstream stages, 4

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions

3

which may include target association, evidence verification, and report generation. Association connects observations corresponding to the same target. Runtime association uses projected detection centroids and camera poses, without ground-truth identities; evaluator matching is performed separately. Verification determines whether the accumulated observations satisfy the mission’s evidence requirements, including repeated confirmations, temporal separation, and freshness across phases. Reporting produces findings subject to the contract’s validity and delivery requirements. The processing stages and their requirements depend on the mission configuration. A preliminary survey may require a different evidence threshold from a confirmation or revisit task, even when the underlying observations and perception models are shared. This allows different mission objectives to be evaluated through a common execution interface without imposing a single evidence criterion on every workload. Mission-level evaluation ( 8 ). The evaluator determines whether the executed sequence satisfies the mission contract. Let H𝑇 denote the mission execution history over horizon 𝑇 , including selected actions, execution events, admitted predictions, accumulated evidence, generated reports, and resource consumption. Each contract defines a set of requirements R (C). Complete mission fulfillment is expressed as

 Û     𝑆 (H𝑇 , C) = I  𝑟 (H𝑇 )  , 𝑟 ∈ R ( C)   

EmbodiedIntentBench: Implementation and Extensibility

We implement EmbodiedIntentBench as a configurable benchmark suite with separate components for workload preparation, runtime execution, and mission evaluation. The implementation supports versioned study configurations, interchangeable runtime policies and processing components, and reproducible execution over recorded sensing workloads.

3.1

Study Assembly and Input Validation

A benchmark study combines a structured mission contract, timeindexed observations, a catalog of perception configurations, execution and resource profiles, and a set of runtime policies. Each study configuration binds these inputs to the applicable mission evaluator and operating conditions. The implementation separates physical workload acquisition from runtime evaluation. Recorded RGB images, vehicle poses, and timestamps define the potential observation stream. Candidate perception configurations are evaluated on these observations to construct a prediction bank indexed by capture and configuration. During replay, this bank is accessed only for the selected computation; precomputation does not eliminate its configured execution costs. This separation enables CPU-only benchmark replay using prepared prediction banks and execution profiles, without requiring repeated neural inference or robotic simulation. GPU acceleration may be used to generate predictions for new perception configurations, but is not required for evaluating runtime strategies over existing workloads. Study manifests identify the observation sources, prediction artifacts, model configurations, mission contracts, and execution profiles. Input checks verify the required capture and configuration coverage, while source and configuration hashes support consistency checks across repeated executions. Execution costs retain their provenance as measurements, characterized profiles, or declared assumptions rather than being treated as equivalent hardware measurements. Importantly, the runtime receives only the public projection of the mission contract, the current eligible observation and system state, and prediction-derived history from completed selected actions. Ground-truth identities, correctness labels, future observations, and unselected model outputs remain inaccessible through the policy interface. These restrictions provide a controlled evaluation boundary for trusted policy implementations.

(7)

where each requirement evaluates an applicable condition on the executed mission history. These conditions may include evidence coverage, confirmation, temporal validity, reporting correctness and timeliness, and resource compliance. Request-level deadlines determine whether individual outputs are admitted as timely evidence, but a missed request does not necessarily imply complete mission failure. Mission success depends on whether the resulting history satisfies all applicable requirements of the contract. The evaluator uses private ground truth to assess prediction correctness, target identities, and evidence validity. This information is excluded from the runtime policy’s decision interface. Beyond complete mission success, EmbodiedIntentBench records diagnostic measures that characterize progress toward individual requirements and explain execution outcomes. These include target coverage, confirmation, timely and unique reports, incorrect or duplicate findings, request deadline misses, processed and skipped observations, and resource consumption. Together, these measures make it possible to distinguish whether a strategy fails because of insufficient evidence, untimely execution, reporting deficiencies, or resource constraints. They also allow improvements in individual systems-level metrics to be examined separately from their contribution to complete mission fulfillment. The following section describes how these architectural components are instantiated in the current benchmark implementation, together with the interfaces for extending its workloads and runtime configurations.

3.2

Runtime Interfaces and Extension Points

EmbodiedIntentBench exposes interfaces for mission specification, physical workloads, perception configurations, runtime policies, execution backends, and evidence processing. Table 1 summarizes their inputs and extension mechanisms. The current studies instantiate perception selection using wholemodel local execution and modeled remote execution. Separate development experiments also exercise replaceable association and evidence- verification stages. These mechanisms are implemented through defined component interfaces rather than unrestricted robotic task execution. Additional arrangements, such as split inference [23, 24], require implementations of their processing stages, intermediate transfers, and cost profiles. Similarly, new mission 5

Bhattacharjya et al. Table 1: Implementation interfaces and extension mechanisms in EmbodiedIntentBench. Components can be replaced independently, subject to their input requirements and corresponding execution or evaluation rules. Component

Required inputs

Extension mechanism

Mission specification

Structured goals, evidence, timing, reporting, and resource requirements. Timestamped observations, vehicle poses, and private evaluation labels. Prediction outputs, preprocessing metadata, and execution profiles. Public contract, eligible observation, resource state, and selected history. Selected actions, processing stages, network conditions, and cost profiles. Completed predictions, evidence history, mission contract, and evaluator truth.

Add mission contracts and corresponding evaluation rules.

Physical workload Perception configuration Runtime policy Execution backend Evidence and evaluation

objectives require compatible evidence-processing and evaluation logic; arbitrary natural-language mission interpretation is not assumed.

3.3

Controlled Replay and Reproducibility

Experimental Methodology

We evaluate EmbodiedIntentBench through two complementary studies. The first examines perception configuration selection in multi-phase UAV sensing missions, while the second investigates runtime adaptation under changing request urgency, connectivity, and battery conditions. We additionally conduct validation experiments examining mission-contract sensitivity, execution-dependent observation availability, mission feasibility, and failure attribution. All studies use controlled replay over recorded simulated workloads, enabling comparisons under matched observation opportunities and operating conditions.

4.1

Implement alternative execution and resource-accounting mechanisms. Replace association and verification stages or introduce task-specific scorers.

The UAV follows a fixed flight trajectory covering the sectors before revisiting the first. The Lake and Depot cohorts contain six and four simulated recordings, respectively. Each recording provides 60 RGB observations at 960 × 720 resolution, sampled every two seconds over a 120-second mission. The mission contracts are motivated by search-and-rescue requirements for coverage, verification, reporting, and continued operation [31], while their numerical thresholds are explicitly defined benchmark parameters. A rapid-survey mission requires evidence for at least half of the designated targets in each phase, supported by two observations spanning at least two seconds. A verified-inspection mission requires all designated targets, with three observations spanning at least four seconds. Both contracts comprise four 30-second phases, including a revisit requiring fresh evidence. The respective minimum uniquetarget report recall requirements are 0.5 and 0.75, with reports due by 120 seconds. Each phase allows at most two false detection components, while the mission permits at most two incorrect and two duplicate reports. An outer execution deadline of 130 seconds is evaluated separately from the phase and reporting deadlines. Dynamic-urgency missions. The second study reuses six 42second Lake recordings covering three route families and two focal sectors. These recordings use a separate wide-area Lake layout containing 24 people across eight sectors (Fig. 3(a,d)). Each local pass targets three people in one focal sector; detections of non-focal people neither satisfy target coverage nor count as false findings. Each recording contains a 36-second sensing pass, followed by an execution and reporting allowance. The mission requires confirmation and reporting of at least two of three focal people using two observations separated by at least two seconds, with reports due by 44 seconds. Global limits permit at most two false detection components, two incorrect reports, and two duplicate reports. The semantic sensing objective remains fixed; changing urgency modifies the execution allowance for newly released perception requests. We consider three schedules: a constant eight-second allowance, a reduction from eight to two seconds at mission time 12 seconds, and a reduction followed by restoration to eight seconds at 24 seconds. Table 2 summarizes the evaluation cohorts. The multi-phase studies cross each recording with two mission contracts and two initial battery conditions. The dynamic study crosses six recordings with three urgency schedules, two network conditions, and two battery conditions. These combinations reuse simulated recordings rather than constituting independent missions.

Each replay initializes fresh policy and evidence-tracking state under a fixed recording, mission contract, and operating condition. Competing policies share the same potential observations and external conditions, while their execution decisions determine which observations are processed and contribute to mission fulfillment. Execution traces record selected actions, prediction accesses, completion times, skipped observations, resource consumption, evidence updates, reports, and requirement outcomes, enabling reconstruction of individual executions. Validation checks reconstruct execution timing, resource accounting, and mission outcomes from recorded inputs, while verifying policy-input isolation and deterministic regeneration of evaluation artifacts. These checks establish replay consistency rather than independently reproducing Gazebo simulations, neural inference, or hardware measurements. Thus, repeated executions provide controlled comparisons over shared simulated workloads, not independent robotic deployments.

4

Import new recordings, environments, and trajectories. Register new models and configuration-specific costs. Implement fixed or adaptive computation-selection strategies.

Physical Workloads and Mission Requirements

Multi-phase sensing missions. We construct UAV sensing workloads in Gazebo Harmonic [22, 25] using two environments: a Lake scene comprising terrain, vegetation, water, and residential structures, and a textured warehouse environment based on the Depot model. The scenes incorporate assets from Gazebo Fuel [26–30]. Fig. 3 illustrates their physical settings and observation opportunities. Each multi-phase scenario designates six static human targets across three sectors. The Lake recordings use separate outbound– return corridor layouts, with all six people designated as targets. 6

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions (a) Lake environment

(b) Depot environment

8 groups / 24 people

6 targets along the aisle

(c) UAV sensing observations Roadside slope Rocky ridge

(d) Flight and mission phases First circuit

Revisit circuit

N

Start / end S8 S5

S1 Open grass

S4

Warehouse aisle

S7 S6 S2

S3

Figure 3: Physical workloads and mission scenarios in EmbodiedIntentBench. (a) Wide-area Lake environment containing 24 people across eight sectors, used for the dynamic-urgency study. (b) Depot warehouse environment. (c) Recorded UAV observations illustrating variations in target visibility. (d) Illustrative wide-area sector circuit, shown for spatial context rather than as an evaluated trajectory. The multi-phase Lake study uses separate six-target corridor layouts, whereas dynamic-urgency recordings comprise local passes near focal sectors. Recorded trajectories remain fixed within each runtime-policy comparison. Table 2: Experimental cohorts and controlled factors. A condition combines one simulated recording with the specified mission and operating settings. Every policy is evaluated once per condition; the three cohorts are analyzed separately. Cohort Lake multi-phase Depot multi-phase Lake dynamic urgency

4.2

Sim. flights

Duration

Conditions

Policies

Runs

Varied factors

6 4 6

120 s 120 s 42 s

24 16 72

21 22 9

504 352 648

Mission contract and initial battery; local perception selection. Mission contract and initial battery; local perception selection. Request urgency, modeled connectivity, and battery; local/remote model routing.

Perception Configurations and Resource Profiles

union cost sums separately timed inference passes and mask fusion. We assume a 30 Wh battery, consistent with the capacity class of lightweight commercial UAVs [38], and a constant 120 W flight-power budget, comparable to experimentally characterized rotary-wing UAV power consumption [39]. Onboard inference is modeled at 15 W, corresponding to a supported power budget of embedded AI platforms such as the NVIDIA Jetson TX2 [40]. Model loading incurs 0.3 seconds and 2 J, routing incurs 0.05 seconds and 0.75 J per decision, and tracking incurs 0.1 seconds and 1 J. Initial battery charge is either 34% or 65%, with a required final reserve of 20% [41], leaving approximately 0.72 kJ and 34.2 kJ, respectively, for task execution after accounting for flight energy and reserve. All configurations execute locally in the primary cohorts. Dynamic resource profiles. The dynamic-urgency study uses a separately configured 120 Wh battery model, consistent with the capacity of commercial UAV batteries such as the Autel EVO Max [42]. Its ample-energy condition begins at 95% charge, whereas the constrained condition is configured to leave 250 J for task execution after accounting for the flight-energy baseline and the required 20% reserve.

The multi-phase studies evaluate nine perception configurations derived from four pretrained model checkpoints: Mask R-CNN with a ResNet-50-FPN v2 backbone [32, 33], Ultralytics YOLO11s and YOLO11n [34], and CABiNet [35] using a checkpoint trained on UAVid [36, 37]. Mask R-CNN contributes standard, 90-degree rotated, and twoview union configurations. The YOLO11 segmentation models contribute full-frame inference and three tiled YOLO11s variants using standard, rotated, or gamma-adjusted inputs. The tiled configurations use a 2 × 2 grid with overlapping regions; the gammaadjusted variant uses 𝛾 = 0.5. CABiNet contributes a semanticsegmentation configuration. All outputs are converted to a common binary person-mask representation. Thus, the nine configurations represent distinct inference pipelines rather than nine independently trained models. Primary resource profiles. The multi-phase cohorts use frozen profiles derived from development-host CPU timings, with configured inference latencies from 0.05 to 6.91 seconds. The two-view 7

Bhattacharjya et al.

We consider two synthetic network regimes. The favorable link provides 50 Mbps uplink, 100 Mbps downlink, and 30 ms round-trip time, while the degraded link provides 2 Mbps uplink, 10 Mbps downlink, and 150 ms round-trip time. The dynamic study uses Mask R-CNN, YOLO11s, and CABiNet, each available through whole-model local or remote execution. Its profiles account for input preparation, communication, server-side processing, result return, and onboard energy costs. Remote service times and network behavior are modeled; the experiments do not execute inference on a deployed remote server. For all cohorts, candidate predictions are prepared on the recorded observation bank. During replay, only the selected configuration’s output becomes available after its charged completion time. The physical trajectory and candidate predictions remain fixed within each matched comparison, while execution delays determine which subsequent observations can be processed.

4.3

Within every condition, policies share the same potential observations, candidate configurations, resource profiles, and evaluation rules. Only public system state and outputs of completed selected actions may inform subsequent decisions; future observations, unselected predictions, and ground-truth identities remain inaccessible to the policy.

4.4

Evaluation Metrics and Validation Protocol

Our primary endpoint is complete mission fulfillment, defined by the conjunction of applicable requirements in Sec. 2. We additionally measure phase completion, target coverage, repeated-evidence confirmation, timely unique reports, incorrect and duplicate findings, request deadline misses, processed and skipped observations, and resource consumption. Ground-truth identities are used for evaluation rather than runtime policy selection. For the multi-phase cohorts, predicted person masks are matched one-to-one against capture-time ground-truth masks using an Intersection over Union (IoU) threshold of 0.3. Evidence must become available within the applicable phase and reporting deadlines to satisfy temporal requirements. Predictions completing after their deadlines retain their incurred execution costs and applicable falsecomponent penalties. We perform four additional benchmark validation analyses using the dynamic study’s recorded observations, predictions, and reference executions; resource-accounted hindsight searches are also applied to the multi-phase cohorts. Contract sensitivity. We evaluate alternative coverage, confirmation, revisit, reporting, and timing requirements through contract-specific rescoring and controlled policy reruns. Rescoring holds execution histories fixed, whereas reruns expose the modified contract to the policy and applicable processing stages. A single-observation survey relaxation is evaluated only through rescoring because the current reporting pipeline requires at least two observations. Mission feasibility. We compute optimistic physical- opportunity and candidate-model bounds to identify conditions that cannot satisfy the required evidence criteria. We additionally search for successful hindsight execution schedules under the original resource-accounted runtime. A successful witness establishes feasibility under the modeled conditions, whereas an unsuccessful bounded search remains inconclusive. Causal execution ablation. We compare the original serial, causal replay with a noncausal no-skip variant that makes otherwise blocked captures available. Configured inference/service latency is varied by factors of 0.5, 1, 2, and 4, while model predictions remain fixed. Local compute energy scales with latency under the declared power model. The no-skip protocol is an analytical ablation rather than a realizable parallel-execution implementation. Failure attribution. We examine unmet evidence, reporting, timing, and resource requirements for unsuccessful executions. Counterfactual rescoring relaxes individual requirements without changing the execution history, distinguishing diagnostic changes in mission scoring from improvements produced by a policy. All comparisons report the number of matched policy–condition executions, while retaining simulated recordings as the unit of workload diversity. The dynamic study uses two recordings for development and four for comparative evaluation; all recordings

Runtime Strategies and Experimental Controls

The multi-phase cohorts compare fixed perception configurations with adaptive reference strategies. Fixed policies repeatedly select one configuration, including when its incurred resource costs violate the mission constraints. The adaptive pool includes round-robin and predicted-quality selection, mission- and phase-survival objectives, temporal evidence-aware selection, and coverage-oriented policies with matched ablations. The Depot cohort additionally evaluates an adaptation of the quality-under-latency scheduling objective from ApproxDet [43]. The comparator uses the benchmark’s perception pool, quality estimates, and execution profiles; it is not a reproduction of the complete ApproxDet detection and scheduling system. For the multi-phase studies, quality-aware policies use frozen, per-configuration logistic recall estimators trained on 596 captures from separate simulated Lake development flights. Estimates depend on 12 image statistics extracted from the current capture, without accessing candidate predictions or evaluation labels. The dynamic study evaluates nine policies: fixed local Mask R-CNN, YOLO11s, and CABiNet; fixed remote Mask R-CNN; a fastestfeasible selector; and four adaptive quality-oriented routers. The adaptive variants differ in whether they account for estimated false-detection risk and whether they observe updated request deadlines. Deadline-aware and deadline-blind variants share the same actual mission requirements and operating conditions. The latter continue selecting configurations using the initial request allowance, allowing the effect of responding to urgency updates to be examined through matched comparisons. The dynamic study instead uses fixed per-model quality and falsedetection estimates calibrated on 162 earlier Lake captures, disjoint from evaluation images. Local and remote variants share the same model-quality estimates. To examine how mission requirements guide execution, we also evaluate six reference policies with and without an intent-driven stopping rule. Within each phase, the rule skips later perception once completed reports contain the contractrequired number of distinct runtime labels. It uses the public mission contract and the runtime’s own reports. 8

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions Table 3: Perception quality and mission fulfillment in Depot. Full-bank recall and mission success are reported alongside completed mission phases and battery-reserve compliance for representative fixed configurations. Recall uses 680 visible target observations across 240 potential captures; mission outcomes use 16 conditions, each comprising four required phases.

were previously available and do not constitute unseen-map generalization. The validation analyses reuse these recordings and do not introduce additional simulated flights.

5

Evaluation

We evaluate EmbodiedIntentBench through two complementary studies of runtime perception and computation strategies in mobile robotic missions. Our evaluation examines how perception quality, execution timing, and resource conditions influence mission fulfillment, and how causal execution and mission-level diagnostics characterize the effectiveness of runtime decisions. We first examine the multi-phase Lake and Depot missions, followed by the dynamic-urgency study and additional benchmark validation experiments.

5.1

Fixed configuration

Recall

Success

Phases

Reserve

Mask R-CNN union YOLO11s, full frame YOLO11s, tiled YOLO11s, tiled + 90◦

0.875 0.818 0.807 0.785

1/16 12/16 14/16 14/16

22/64 58/64 60/64 62/64

8/16 16/16 16/16 16/16

Comparison of runtime strategies. Across Lake, the evaluated adaptive policies complete 6–13 of 24 missions, while those in Depot complete 2–7 of 16. The coverage-oriented adaptive reference completes 9/24 and 6/16 missions, respectively, whereas the strongest fixed configurations complete 16/24 and 14/16. The ApproxDet-inspired comparator completes 2/16 Depot missions. Mission feasibility and runtime effectiveness. Resourceaccounted, evaluator-guided search identifies successful execution schedules for all 24 Lake and 16 Depot conditions. These results establish achievable mission outcomes under the evaluated observation opportunities, perception configurations, and resource constraints. These feasible execution references quantify the headroom between the evaluated strategies and achievable mission outcomes under the modeled conditions.

Perception Accuracy and Mission Fulfillment

Perception quality versus mission success. Our first experiment reveals a substantial discrepancy between frame-level perception accuracy and mission fulfillment. Fig. 4(a) compares full-bank target recall and mission success across nine fixed perception configurations in Depot, while Table 3 provides phase-completion and resource-compliance results for representative configurations. Fullbank recall is computed over 680 visible target observations across 240 potential RGB captures, whereas mission success is evaluated over 16 resource-accounted execution conditions. The two-view Mask R-CNN union achieves the highest full-bank recall of 0.875, yet completes only 1 of 16 missions. In contrast, tiled YOLO11s achieves a recall of 0.807 and completes 14 of 16 missions. Its rotated tiled variant likewise completes 14 missions with a recall of 0.785. The gamma-adjusted tiled variant achieves comparable recall (0.797 versus 0.807) but completes only 6 rather than 14 missions, primarily due to violations of the phase-level false-detection limit. The difference highlights how perception configurations with similar individual-frame accuracy can contribute very differently to complete robotic missions. Perception quality estimation. Fig. 4(b) compares frozen predicted and observed target recall across the nine configurations. The estimates exhibit substantial configuration-dependent discrepancies. For example, rotated tiled YOLO11s achieves an observed recall of 0.785 against a predicted recall of 0.007, while CABiNet has a predicted recall of 0.548 and zero observed recall. These discrepancies characterize the uncertainty in quality estimates available to adaptive computation-selection strategies. The benchmark further distinguishes the contributions of perception, evidence accumulation, reporting, and resource compliance. The Mask R-CNN union achieves a mean timely report recall of 1.000, while completing 22 of 64 required mission phases and satisfying the battery reserve in 8 of 16 conditions. With 6.91 seconds of configured inference latency and two-second capture intervals, Mask R-CNN union processes approximately one in four observation opportunities, limiting the evidence available for phase completion. In comparison, standard tiled YOLO11s completes 60 phases and satisfies the reserve in all 16 conditions. These results demonstrate how mission fulfillment emerges from the combined effects of observation availability, temporal evidence, reporting, and resource constraints.

5.2

Runtime Adaptation Under Changing Conditions

We next examine how request urgency, connectivity, and battery availability influence runtime decisions and mission outcomes. The dynamic-urgency study evaluates nine strategies over 72 conditions; the four-recording comparative subset contains 48 matched conditions per strategy. Fig. 5 summarizes the effects of these changing conditions on mission success, execution timeliness, and resource consumption. Adaptation to changing urgency. As shown in Fig. 5(a), the urgency-aware quality- and risk-oriented router completes 25/48 missions, compared with 24/48 for its urgency-blind counterpart and 22/48 for fixed remote Mask R-CNN. This corresponds to three additional successful conditions, or 6.25 percentage points, relative to the fixed remote reference. The influence of urgency awareness is particularly evident in execution timeliness. By incorporating updated request deadlines, the quality- and risk-oriented router reduces late perception requests from 90 to zero across the 48 evaluation conditions. Its mean modeled onboard task energy also decreases from 198 J to 168 J, a reduction of approximately 15%, excluding flight energy. The corresponding mission success changes from 24/48 to 25/48. The benchmark thus distinguishes improvements in computational timeliness and resource efficiency from their ultimate contribution to mission evidence and fulfillment. Connectivity-dependent execution. Fig. 5(b) shows the influence of modeled connectivity on mission fulfillment. Fixed remote Mask R-CNN completes 18/24 missions under favorable connectivity and 4/24 under degraded connectivity. Under the degraded link, the urgency-aware quality- and risk-oriented router completes 7/24 missions. These results illustrate how network conditions affect 9

Bhattacharjya et al. (a) Full-bank recall and mission success Observed recall

(b) Predicted versus observed recall

Missions completed 0.875

Mask R-CNN union Mask R-CNN

0.840

YOLO11s full

0.818

4/16 12/16

YOLO11s tiled

0.807 0.801

5/16

YOLO11s tiled γ=0.5

0.797

6/16

14/16

0.785

YOLO11n full

14/16

0.696

CABiNet (UAVid)

Predicted

1/16

Mask R-CNN +90°

YOLO11s tiled +90°

Observed

8/16

0.000

0/16

0

0.5

1

0

Recall

8

16

Successes / 16

0

0.5

1

Full-bank recall

Figure 4: Perception quality and mission fulfillment across nine configurations in Depot. (a) Full-bank target recall and completed missions for fixed perception configurations, highlighting the discrepancy between Mask R-CNN union (0.875 recall; 1/16 missions) and tiled YOLO11s (0.807 recall; 14/16 missions). (b) Frozen predicted and observed full-bank target recall across configurations, illustrating variation in quality-estimation accuracy. Predicted recall is averaged over captures with visible-target counts as weights for offline analysis. Recall is evaluated over 680 visible target observations from 240 potential captures; mission success uses 16 resource-accounted execution conditions. (a) Urgency adaptation

Successes (48)

Aware

Blind

25

24

Late requests

0

90

Task energy (J)

168

198

(b) Connectivity Fixed remote

(c) Battery Aware

Fixed local

5.3

Aware

24

18 18

12 7 6

4

0

Successful executions (of 24)

Successful executions (of 24)

24

18

Embedded hardware sensitivity. In replay with Jetson AGX Orin profiles, each adaptive router succeeds in 26/72 conditions with CPU costs and 60/72 with GPU costs. Fixed-local Mask R-CNN succeeds in 0/72 and 60/72, respectively, while fixed-remote Mask R-CNN succeeds in 38/72. Device and replay capture schedules agree in 10/12 fixed-local runs (Appendix A.5).

13

6

Degraded

12

6

0

0 Good

A key feature of EmbodiedIntentBench is that computation affects the observation opportunities available during physical motion. We evaluate this relationship through a controlled comparison between causal execution, where processing blocks intervening captures, and a full-capture counterfactual that retains these observations. The no-skip protocol permits overlapping inferences as an analytical ablation; it is not an implementation of serial execution. Fig. 6(a) compares mission success under causal and no-skip execution across four inference-latency multipliers. At the original latency, causal execution yields 218 successful executions, compared with 255 under no-skip execution. The no-skip counterfactual yields 37 additional successes under these modeled conditions. Impact on runtime strategy rankings. Fig. 6(b) shows how causal observation accounting changes the relative effectiveness of runtime strategies. At the original latency, urgency-aware quality routing completes 41/72 conditions, compared with 38/72 for fixed remote Mask R-CNN. Under no-skip execution, the relative ordering of these two strategies reverses, with 45/72 and 48/72 completed missions, respectively. Across all four adaptive routers, causal execution yields 41–43 successful missions, compared with 38 for fixed remote execution. Under no-skip execution, the corresponding results are 45–48 versus 48, showing that causal observation accounting influences strategy comparisons beyond the representative pair in Fig. 6 (b). Sensitivity to execution latency. Doubling configured inference latency reduces causal mission success from 218 to 186 executions, while full-capture success remains at 255. Among executions

18

12

Ample

Causal Execution and Observation Availability

Constrained

Figure 5: Runtime adaptation under changing urgency and resource conditions. (a) Mission success, late perception requests, and mean onboard task energy for urgency-aware and urgency-blind quality-and-risk routing across 48 matched conditions. (b) Mission success under favorable and degraded connectivity, comparing fixed remote execution with urgency-aware routing. (c) Mission success under ample and constrained battery conditions, comparing fixed local execution with urgency-aware routing. Connectivity and battery comparisons use 24 conditions per operating setting. Energy excludes flight consumption.

mission fulfillment through communication latency and evidence availability. Energy-constrained execution. Fig. 5(c) illustrates the impact of battery availability on mission fulfillment. Fixed local Mask RCNN completes 6/24 missions under the ample-energy condition, compared with 0/24 under the constrained condition, where all executions violate the specified reserve. This contrast illustrates how resource obligations influence mission fulfillment alongside perception and reporting outcomes. 10

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions (a) Latency sensitivity

(b) Policy success by protocol

Causal

No-skip

Quality-aware

300 284

60 255

255

200

218 186

100

81

Successes / 72

272 Successes / 648

Feasibility characterization. We analyze 72 dynamic-urgency conditions using physical observation bounds, candidate-model evidence bounds, and resource-accounted hindsight searches. All 72 satisfy the physical visibility bound, while 12 are certified infeasible for the evaluated model pool due to insufficient repeated confirmations. Reference policies complete 43 conditions, and hindsight search identifies successful schedules for eight more; nine remain unresolved. These resource-accounted witnesses identify conditions where successful schedules exist under the modeled execution and quantify opportunities beyond the evaluated policies. Tracing mission evidence. Across 216 person–condition instances, the urgency-aware quality-and-risk router obtains matched detections for 172 instances, confirms 139, and generates 129 timely reports. Fixed remote Mask R-CNN achieves 146 matched detections, 114 confirmations, and 112 timely reports. This progression identifies where evidence is lost between observation, detection, verification, and reporting. Requirement-level diagnosis. Within the 48-condition comparative subset, the urgency-aware router completes 25 missions. Of the remaining 23, all fail the report-recall requirement, while 18 also lack sufficient confirmation. Thus, eliminating late requests does not necessarily provide the evidence required for mission fulfillment.

Fixed remote

40

41

45

48

38

20

53 0

0 0.5×

1× 2× Latency multiplier

4×

Causal No-skip Lead among shown policies: quality → remote

(c) Capture availability over time Fixed local Mask R-CNN Processed

Skipped (busy)

Person 3 matched

P3 report 31.46 s

Causal 27.46 s No-skip 0

8

16

24

32

Mission time (s) Processed: 9/18 causal, 18/18 no-skip

Figure 6: Impact of causal execution on mission outcomes and runtime strategy comparisons. (a) Mission successes under causal and no-skip execution across four inference-latency multipliers, using 648 matched policy–condition executions per setting. (b) Mission successes for urgency-aware quality routing and fixed remote Mask R-CNN under causal and no-skip execution. The relative advantage reverses from 41/72 versus 38/72 under causal execution to 45/72 versus 48/72 under no-skip execution. (c) Observation availability in an illustrative fixed-local Mask R-CNN execution. Causal processing admits 9 of 18 captures, compared with 18 under no-skip execution. Markers show capture processing, person-3 matches, and person-level report times.

5.5

that transition from success to failure under causal replay, every one loses at least one observation previously contributing to target confirmation. For fixed-local Mask R-CNN, the mean number of processed observations decreases from nine to five as latency doubles. These results reveal a direct connection between computation latency, available sensing opportunities, accumulated evidence, and complete mission fulfillment. Execution-induced observation loss. Fig. 6(c) illustrates the effect of blocking computation on observation availability during an illustrative fixed-local Mask R-CNN execution. Causal processing admits nine of 18 potential captures, whereas no-skip execution processes all 18. Person-3 reports become available at 31.46 and 27.46 seconds, respectively. This trace illustrates how computation latency shapes both the observations processed and the timing of accumulated mission evidence. The evaluation establishes causal observation accounting as an important component of benchmarking runtime computation in mobile robotic systems. It captures mission-level consequences that arise from the interaction between physical motion and computation.

5.4

Sensitivity to Mission Requirements

Finally, we examine how mission requirements affect runtime strategy comparisons. We rescore the dynamic-study executions under 144 contract settings varying target coverage, observation requirements, false-detection limits, reporting obligations, and deadlines. For each setting, we compare the retrospectively best adaptive policy against the accurate fixed-reference class (local and remote Mask R-CNN). A material advantage requires at least three additional successful conditions and no recording favoring the comparison class. Across the 144 settings, the retrospectively selected best adaptive policy achieves a material advantage over both fixed Mask R-CNN references in 128 settings; the reverse advantage is not observed. The grid was specified before these settings were scored, after the five-contract results and diagnostics were available. In separate contract-conditioned reruns of the original policies, executed actions changed from the reference schedules in 54 of 1,944 cases, and one mission outcome differed from its fixed-history rescore. While rescoring examines how requirements affect the assessment of existing executions, the following comparison examines how requirements can guide runtime decisions. Intent-driven execution. We compare each policy with and without the rule across four contracts and three cost profiles. With the Jetson GPU profile, fixed-local Mask R-CNN uses 28% and 20% less estimated task energy under two- and three-observation confirmation, respectively, while completing 60/72 conditions in both cases with and without the rule. Across 5,184 matched pairs of replays, all previously successful executions remain successful. Executing replay-selected schedules on the Jetson reduces mean whole-pass module energy by 7.7% for two-observation confirmation across six recordings (Appendix A.6). Together, these experiments demonstrate how EmbodiedIntentBench connects runtime

Mission Feasibility and Evidence-Level Diagnostics

We further evaluate how EmbodiedIntentBench characterizes mission feasibility and identifies opportunities for improving runtime strategies. 11

Bhattacharjya et al. Table 4: Comparison with representative UAV, embodied, and robotic computing benchmarks. Motion: task-level evaluation involving simulated or physical robot motion; Adapt.: runtime inference scheduling or inference-configuration selection; View: inference timing affects observation availability or freshness; Link: network-aware computation placement; Energy: energy or power accounting; Evidence: temporal confirmation and reporting requirements; Mission: conjunctive evaluation of evidence, timing, reporting, and resource requirements. Checkmarks (✓ ) denote evaluated capabilities; dashes (–) indicate features outside the reported evaluation scope. Benchmark

Motion

Adapt.

View

Link

Energy

Evidence

Mission

MAVBench [14] Air Learning [44] OpenUAV [45] BEDI [46] UAVBench [47] FloodReasonBench [5]

✓ ✓ ✓ ✓

– – – – – –

✓ ✓

– – – – – –

✓ ✓

– – – – – –

– – – – – –

RobotPerf [15] A2Perf [16] EmbodiedBench [48] LIBERO-Dynamic [21]

–

✓ ✓ ✓

– – – –

– – –

– – – –

– – – –

Streaming Perception [17] ApproxDet [43] DEEP [20]

– – –

✓ ✓ ✓

✓

– – –

EmbodiedIntentBench (Ours)

✓

✓

✓

– –

– – – –

✓

– – – –

✓ ✓

– –

– – –

– –

✓

– – –

✓

✓

✓

✓

✓

computation decisions to mission evidence and fulfillment under changing physical and resource conditions.

6

– – –

– –

missed during processing, introducing streaming accuracy as an online perception metric. ApproxDet [43] adapts perception configurations to changing content, latency requirements, and resource contention, while Chanakya [52] learns runtime configuration decisions for real-time perception. DEEP [20] evaluates perceptioncomputation allocation through its downstream decision value under resource budgets. ACCESS-AV [53] adapts localization computation to communication and motion conditions to balance accuracy and energy in autonomous vehicles. MemoGuard [54] validates retrieved actions against topology, resource, and outcome constraints before memory reuse in communication-limited robot navigation. These approaches demonstrate the importance of runtime computation selection and execution timing, extending evaluation beyond isolated prediction accuracy. Mission-aware runtime evaluation. Table 4 compares representative benchmarks across robotic execution, runtime adaptation, and mission-level evaluation. EmbodiedIntentBench evaluates runtime computation strategies over physically simulated UAV sensing workloads, using recorded trajectories and common potential observation opportunities. Structured mission contracts specify coverage, repeated confirmation, fresh revisits, reporting deadlines, and resource constraints. Causal, resource-accounted execution captures how model selection, local or remote computation, network conditions, and processing delays affect the observations and results available to each mission. Complete mission fulfillment is evaluated against these requirements, with evidence-level diagnostics and resource-accounted feasibility witnesses supporting the analysis of runtime strategy effectiveness.

Related Work

We now position EmbodiedIntentBench relative to prior work in robotic systems benchmarking, embodied task evaluation, and runtime-aware computation. UAV and robotic systems benchmarks. MAVBench [14] and Air Learning [44] evaluate UAV computing systems through their interactions with physical execution, task performance, and energy consumption. The UAV Roofline model [49] further characterizes the relationship between sensing, computation, and vehicle dynamics. RobotPerf [15] provides standardized performance evaluation of robotic computing pipelines across heterogeneous platforms, while A2Perf [16] evaluates autonomous agents through task performance, generalization, reliability, and system resource costs. System-level characterization of generative embodied agents [50] further identifies computation and communication bottlenecks in long-horizon embodied tasks. These studies establish the importance of relating computational performance and resource efficiency to the operation of robotic systems. Embodied task and intent benchmarks. OpenUAV [45] provides realistic simulated environments for language-guided UAV navigation. BEDI [46] evaluates UAV agents through perception, decision-making, and action, using hierarchical metrics across virtual and real-world scenarios. UAVBench [47] introduces structured UAV mission scenarios for evaluating agentic reasoning, while AirCopBench [51] examines collaborative aerial perception under challenging sensing conditions. FloodReasonBench [5] evaluates VLM reasoning segmentation and edge execution tradeoffs for flood response. Beyond UAVs, EmbodiedBench [48] evaluates vision-driven embodied agents across navigation and manipulation tasks. LIBERO-Dynamic [21] incorporates inference latency into dynamic manipulation, allowing the simulated environment to evolve while computation executes. These benchmarks evaluate complementary aspects of embodied task execution, instruction following, perception, and reasoning. Runtime-aware perception and computation. Streaming Perception [17] accounts for inference latency and observations

7

Conclusion and Future Work

We presented EmbodiedIntentBench, a benchmark for intentdriven runtime adaptation in mobile robotic missions. It couples mission contracts with physically situated sensing and causal, resource-accounted execution to connect computation decisions with observation availability, evidence, and mission fulfillment. Our evaluation shows how perception quality, observation loss, and resource constraints shape mission outcomes and strategy effectiveness. Requirement-level diagnostics and feasibility analysis identify opportunities for improved adaptation. 12

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions

The evaluation uses Gazebo-based UAV workloads with recorded trajectories and modeled execution and resource profiles. Future work will extend EmbodiedIntentBench to closed-loop execution and measured edge–cloud deployments, capturing coupled sensing, motion, and communication dynamics.

[16] Ikechukwu Uchendu, Jason Jabbour, Korneel Van den Berghe, Joel Runevic, Matthew Stewart, Jeffrey Ma, Srivatsan Krishnan, Izzeddin Gur, Austin Huang, Colton Bishop, Paige Bailey, Joe Wenjie Jiang, Ebrahim M. Songhori, Sergio Guadarrama, Jie Tan, Jordan K. Terry, Aleksandra Faust, and Vijay Janapa Reddi. A2Perf: Benchmarking autonomous agents end-to-end in realistic domains. In ICML Workshop on Championing Open-source Development in Machine Learning (CODEML), 2025. [17] Mengtian Li, Yu-Xiong Wang, and Deva Ramanan. Towards streaming perception. In European conference on computer vision, pages 473–488. Springer, 2020. [18] Jinrong Yang, Songtao Liu, Zeming Li, Xiaoping Li, and Jian Sun. Real-time object detection for streaming perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5385–5395, 2022. [19] Rowan McAllister, Blake Wulfe, Jean Mercat, Logan Ellis, Sergey Levine, and Adrien Gaidon. Control-aware prediction objectives for autonomous driving. In 2022 International Conference on Robotics and Automation (ICRA). IEEE, 2022. [20] Hoang Pham Cong and Ho Viet Duc Luong. The decision value of perception compute. arXiv preprint arXiv:2609.35910, 2026. [21] Shaojie Ren, Jingwen Sun, Zezhi Liu, Hongjin Chen, Xin Jin, Wei Zou, and Zhibo Chen. LIBERO-Dynamic: Benchmarking robotic manipulation in dynamic environments. arXiv preprint arXiv:2610.24579, 2026. [22] Nathan Koenig and Andrew Howard. Design and use paradigms for gazebo, an open-source multi-robot simulator. In 2004 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE Cat. No. 04CH37566), volume 3, pages 2149–2154. Ieee, 2004. [23] Yoshitomo Matsubara, Marco Levorato, and Francesco Restuccia. Split computing and early exiting for deep learning applications: Survey and research challenges. ACM Computing Surveys, 55(5):1–30, 2022. [24] Rajat Bhattacharjya, Minwoo Kim, Arnab Sarkar, Tamoghno Das, Sing-Yao Wu, Eli Bozorgzadeh, Marco Levorato, and Nikil Dutt. Freshlatent: Channel-aware latent adaptation for resource-constrained embodied vlm perception. arXiv preprint arXiv:2609.30629, 2026. [25] Open Robotics. Gazebo Harmonic documentation. https://gazebosim.org/docs/h armonic/. Accessed: November 3, 2026. [26] Rafi Abdullah and Louise Poubel. Depot. https://fuel.gazebosim.org/1.0/OpenR obotics/models/Depot. Accessed: November 3, 2026. [27] Cole Biesemeyer. Harmonic Terrain. https://fuel.gazebosim.org/1.0/OpenRobot ics/models/Harmonic%20Terrain. Accessed: November 3, 2026. [28] Cole Biesemeyer. Harmonic Terrain Objects. https://fuel.gazebosim.org/1.0/O penRobotics/models/Harmonic%20Terrain%20Objects. Accessed: November 3, 2026. [29] Cole Biesemeyer. Lake House. https://fuel.gazebosim.org/1.0/OpenRobotics/mo dels/Lake%20House. Accessed: November 3, 2026. [30] OpenRobotics. Coast Waves 2. https://fuel.gazebosim.org/1.0/OpenRobotics/mo dels/Coast%20Waves%202. Accessed: November 3, 2026. [31] National Search and Rescue Council. National Search and Rescue Manual: Volume 2—Search and Rescue Operations. Australian Maritime Safety Authority, 2026. Accessed: November 3, 2026. [32] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In Proceedings of the IEEE International Conference on Computer Vision, pages 2961–2969, 2017. [33] TorchVision Contributors. TorchVision: Mask r-cnn with resnet-50-fpn v2. https://docs.pytorch.org/vision/main/models/generated/torchvision.models.d etection.maskrcnn_resnet50_fpn_v2.html. Accessed: November 3, 2026. [34] Glenn Jocher and Jing Qiu. Ultralytics YOLO11. docs.ultralytics.com/models/yolo11, 2024. Accessed: November 3, 2026. [35] Saumya Kumaar, Ye Lyu, Francesco Nex, and Michael Ying Yang. CABiNet: Efficient context aggregation network for low-latency semantic segmentation. In 2021 IEEE International Conference on Robotics and Automation, pages 13517– 13524, 2021. [36] Ye Lyu, George Vosselman, Gui-Song Xia, Alper Yilmaz, and Michael Ying Yang. UAVid: A semantic segmentation dataset for UAV imagery. ISPRS Journal of Photogrammetry and Remote Sensing, 165:108–119, 2020. [37] dronefreak. CABiNet (MobileNetV3-Small) Finetuned on UAVid. https://huggin gface.co/dronefreak/cabinet-mobilenetv3-small-uavid. Accessed: November 3, 2026. [38] DJI. DJI Mini 4 Pro: Specifications. https://www.dji.com/mini-4-pro/specs. Accessed: November 3, 2026. [39] Krzysztof Falkowski and Michał Duda. Dynamic models identification for kinematics and energy consumption of rotary-wing uavs during different flight states. Sensors, 23(23):9378, 2023. Accessed: November 3, 2026. [40] NVIDIA. Jetson TX2 Module: Technical specifications. https://developer.nvidia .com/embedded/jetson-tx2. Accessed: November 3, 2026. [41] New Jersey Department of Environmental Protection. sUAS Flight Operations Manual. New Jersey Department of Environmental Protection, 2020. [42] Autel Robotics. EVO Max Smart Flight Battery. https://shop.autelrobotics.com /products/evo-max-smart-flight-battery. Accessed: November 3, 2026.

References [1] Rajat Bhattacharjya, Sing-Yao Wu, Hyunwoo Oh, Chaewon Nam, Suyeon Koo, Mohsen Imani, Elaheh Bozorgzadeh, and Nikil Dutt. AVERY: Intent-driven adaptive VLM split computing via embodied self-awareness for efficient disaster response systems. arXiv preprint arXiv:2511.18151, 2025. [2] Fernando Cladera, Zachary Ravichandran, Jason Hughes, Varun Murali, Carlos Nieto-Granda, M Ani Hsieh, George J Pappas, Camillo J Taylor, and Vijay Kumar. Air-ground collaboration for language-specified missions in unknown environments. IEEE Transactions on Field Robotics, 2:626–642, 2025. [3] Reza Farahani, Zoha Azimi, Ilir Murturi, Arda Goknil, Sagar Sen, Christian Timmerer, and Schahram Dustdar. EVLM: Intent-driven edge vision language model for UAV-based power line inspection. In 2026 IEEE International Conference on Edge Computing and Communications (EDGE), pages 32–42. IEEE, 2026. [4] Minghui Zhao, Junxi Xia, Kaiyuan Hou, Yanchen Liu, Stephen Xia, and Xiaofan Jiang. FlexiFly: Interfacing the physical world with foundation models empowered by reconfigurable drone systems. In Proceedings of the 23rd ACM Conference on Embedded Networked Sensor Systems, SenSys ’25, page 463–476, New York, NY, USA, 2025. Association for Computing Machinery. [5] Rajat Bhattacharjya, Yoomee Jung, Minwoo Kim, Sing-Yao Wu, Eli Bozorgzadeh, Nalini Venkatasubramanian, and Nikil Dutt. FloodReasonBench: Benchmarking VLM reasoning segmentation for embodied flood response at the edge. arXiv preprint arXiv:2608.15410, 2026. [6] Qitao Weng and Heechul Yun. Multi-resolution end-to-end deep neural network for optimizing latency-accuracy tradeoff in autonomous driving. In 2026 ACM/IEEE 17th International Conference on Cyber-Physical Systems (ICCPS), pages 12–18. IEEE, 2026. [7] Jingxian Chen, Yukai Song, Shengjie Xu, Kyle Kessler, Prateek Ganguli, Arkaprava Gupta, Bineet Ghosh, Parasara Sridhar Duggirala, Tianlong Chen, Samarjit Chakraborty, et al. Controller-aware perception neural network pruning for cps safety optimization. ACM Transactions on Cyber-Physical Systems, 2026. [8] Ahmet Soyyigit, Shuochao Yao, and Heechul Yun. MURAL: A multi-resolution anytime framework for LiDAR object detection deep neural networks. In 2025 IEEE 31st International Conference on Embedded and Real-Time Computing Systems and Applications (RTCSA), pages 33–43. IEEE, 2025. [9] Jingfei Xia, Yuze He, Chen Bian, Zhenyu Yan, and Guoliang Xing. EventEye: Towards high-frequency perception enhancement for autonomous vehicles using infrastructure-mounted event cameras. In Proceedings of the 2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems, SenSys ’26, page 803–817, New York, NY, USA, 2026. Association for Computing Machinery. [10] Hyunseung Kim, Sheetal Prasanna, Yin Li, Somali Chaterji, and Saurabh Bagchi. ApproxBit: Efficient video analytics through latency-aware offloading with learned binary codes. In Proceedings of the 2026 ACM/IEEE International Conference on Embedded Artificial Intelligence and Sensing Systems, pages 1179–1193, 2026. [11] Zoha Azimi, Reza Farahani, Schahram Dustdar, and Christian Timmerer. Cloud, edge, or split? profiling onboard and split vision-language model deployment for drone ai. arXiv preprint arXiv:2609.25415, 2026. [12] David Candal-Ventureira, Francisco Javier González-Castaño, Felipe GilCastiñeira, and Pablo Fondo-Ferreiro. Is the edge really necessary for drone computing offloading? an experimental assessment in carrier-grade 5G operator networks. Software: Practice and Experience, 53(3):579–599, 2023. [13] Mihir Bala, Aditya Chanana, Xiangliang Chen, Qifei Dong, Thomas Eiszler, Jingao Xu, Padmanabhan Pillai, and Mahadev Satyanarayanan. The OODA loop of cloudlet-based autonomous drones. In 2024 IEEE/ACM Symposium on Edge Computing (SEC), pages 178–190, 2024. [14] Behzad Boroujerdian, Hasan Genc, Srivatsan Krishnan, Wenzhi Cui, Aleksandra Faust, and Vijay Reddi. MAVBench: Micro aerial vehicle benchmarking. In 2018 51st annual IEEE/ACM international symposium on microarchitecture (MICRO), pages 894–907. IEEE, 2018. [15] Víctor Mayoral-Vilches, Jason Jabbour, Yu-Shun Hsiao, Zishen Wan, Martiño Crespo-Álvarez, Matthew Stewart, Juan Manuel Reina-Muñoz, Prateek Nagras, Gaurav Vikhe, Mohammad Bakhshalipour, Martin Pinzger, Stefan Rass, Smruti Panigrahi, Giulio Corradi, Niladri Roy, Phillip B. Gibbons, Sabrina M. Neuman, Brian Plancher, and Vijay Janapa Reddi. RobotPerf: An open-source, vendoragnostic, benchmarking suite for evaluating robotics computing system performance. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 8288–8297, 2024. 13

Bhattacharjya et al.

[43] Ran Xu, Chen lin Zhang, Pengcheng Wang, Jayoung Lee, Subrata Mitra, Somali Chaterji, Yin Li, and Saurabh Bagchi. ApproxDet: Content and contentionaware approximate object detection for mobiles. In Proceedings of the 18th ACM Conference on Embedded Networked Sensor Systems. ACM, 2020. [44] Srivatsan Krishnan, Behzad Boroujerdian, William Fu, Aleksandra Faust, and Vijay Janapa Reddi. Air Learning: A Deep Reinforcement Learning Gym for Autonomous Aerial Robot Visual Navigation. Machine Learning, 110:2501–2540, 2021. [45] Xiangyu Wang, Donglin Yang, Ziqin Wang, Hohin Kwan, Jinyu Chen, Wenjun Wu, Hongsheng Li, Yue Liao, and Si Liu. Towards Realistic UAV Vision-Language Navigation: Platform, Benchmark, and Methodology. arXiv:2410.07087, 2024. [46] Mingning Guo, Mengwei Wu, Jiarun He, Shaoxian Li, Haifeng Li, and Chao Tao. BEDI: A Comprehensive Benchmark for Evaluating Embodied Agents on UAVs. ISPRS Journal of Photogrammetry and Remote Sensing, 232:910–936, 2026. [47] Mohamed Amine Ferrag, Abderrahmane Lakas, and Mérouane Debbah. UAVBench: An Open Benchmark Dataset for Autonomous and Agentic AI UAV Systems via LLM-Generated Flight Scenarios. IEEE Open Journal of Vehicular Technology, 7:2542–2558, 2026. [48] Rui Yang, Hanyang Chen, Junyu Zhang, Mark Zhao, Cheng Qian, Kangrui Wang, Qineng Wang, Teja Venkat Koripella, Marziyeh Movahedi, Manling Li, et al. EmbodiedBench: Comprehensive benchmarking multi-modal large language models for vision-driven embodied agents. arXiv preprint arXiv:2502.09560, 2025. [49] Srivatsan Krishnan, Zishen Wan, Kshitij Bhardwaj, Aleksandra Faust, and Vijay Janapa Reddi. Roofline model for uavs: A bottleneck analysis tool for designing compute systems for autonomous drones. arXiv preprint arXiv:2111.03792, 2021. [50] Zishen Wan, Jiayi Qian, Yuhang Du, Jason Jabbour, Yilun Du, Yang Zhao, Arijit Raychowdhury, Tushar Krishna, and Vijay Janapa Reddi. Generative AI in embodied systems: System-level analysis of performance, efficiency and scalability. In 2025 IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS), pages 26–37, 2025. [51] Jirong Zha, Yuxuan Fan, Tianyu Zhang, Geng Chen, Yingfeng Chen, Chen Gao, and Xinlei Chen. AirCopBench: A benchmark for multi-drone collaborative embodied perception and reasoning. Proceedings of the AAAI Conference on Artificial Intelligence, 40(2):1507–1515, 2026. [52] Anurag Ghosh, Vaibhav Balloli, Akshay Nambi, Aditya Singh, and Tanuja Ganu. Chanakya: Learning runtime decisions for adaptive real-time perception. Advances in Neural Information Processing Systems, 36:55668–55680, 2023. [53] Rajat Bhattacharjya, Arnab Sarkar, Ish Kool, Sabur Baidya, and Nikil Dutt. Accessav: Adaptive communication-computation codesign for sustainable autonomous vehicle localization in smart factories. ACM Trans. Embed. Comput. Syst., 25(1), January 2026. [54] Rajat Bhattacharjya, Hyeonjong Ju, Sing-Yao Wu, Eli Bozorgzadeh, and Nikil Dutt. MemoGuard: An adaptive runtime for guarding against memory traps in communication-limited robot navigation. IEEE Embedded Systems Letters, 18(5):447–451, 2026.

14

EmbodiedIntentBench : Benchmarking Intent-Driven Runtime Adaptation for Mobile Robotic Missions

A

Appendix: Reproducibility and Validation Details

successes among the 72 conditions and at least as many successes on every recording. This comparison characterizes the evaluated histories and thresholds.

Public artifact: https://github.com/rajatbh21/EmbodiedInten tBench-artifact This appendix provides additional details for the evaluation in Secs. 4 and 5, including workload validation, mission-contract analyses, causal execution, and Jetson measurements. Several analyses use the dynamic-urgency study, which evaluates adaptation under changing urgency, connectivity, and battery conditions using recorded observations and stored predictions. The accompanying artifact provides code, inputs, and reference results for this study and the Jetson and intent-driven analyses. The complete benchmark artifact, including the multi-phase workloads and prediction banks, will be released upon acceptance.

A.1

A.3

The reference confirmation contract requires reports for at least two of three people, supported by two observations spanning at least two seconds, by 44 seconds. Four variants relax the observation count to one (survey), require three observations spanning four seconds (strict confirmation), require fresh confirmation in both [0, 18) and [18, 36) seconds (revisit), or move the phase and reporting deadlines to 36 seconds (time-critical). The single-observation survey is distinct from the multi-phase rapid-survey mission, which requires two observations. Table 5 illustrates how these contracts affect selected reference strategies. The model-union bound pools stored predictions from all candidate models at each capture to provide an optimistic evidence test. For reruns, the denominator includes conditions not certified infeasible by this bound; such conditions can remain unresolved. The survey column is rescored rather than rerun. Across 1,944 eligible contract-conditioned rerun cells, executed actions changed in 54, and one mission outcome differed from fixed-history rescoring. All 12 pairwise rerun ranking reversals were one-condition changes between adaptive variants.

Workloads and replay validation

The evaluation comprises multi-phase sensing missions in the Lake and Depot environments and a separate dynamic-urgency study in Lake. The replay validation and contract analyses below focus on the latter study, which varies request urgency and operating conditions over shared recordings. The dynamic-urgency study uses six 42-second Lake recordings with precomputed prediction banks and execution profiles. Its 72 matched conditions combine these recordings with urgency, network, and battery settings. Replay recomputes runtime decisions, evidence, and mission outcomes from stored predictions using modeled execution and network profiles. The accompanying artifact includes the dynamic-study recordings, prediction banks, execution profiles, nine evaluated strategies, reference outcomes, and validation code. The multi-phase recording and prediction bundles are outside its scope; numerical Depot case and frame extracts support verification of the data underlying Fig. 4. Validation checks compare regenerated dynamic-study results with reference outputs. Instrumented replay also checked that the evaluated policies accessed predictions only for the selected action and current capture.

A.2

Contract-conditioned comparisons

Table 5: Selected successes under five contracts. Survey uses rescoring; other columns use reruns. Denominators for reruns exclude only model-unioncertified infeasible conditions. Strategy

Survey Confirm Strict Revisit Time

Local Mask 12/72 Remote Mask 38/72 Aware quality+risk 43/72

12/60 38/60 43/60

8/60 0/36 12/60 34/60 12/36 38/60 36/60 13/36 43/60

The 144-setting grid was specified before its settings were scored, after the five-contract results and descriptive diagnostics were available. The resulting analysis characterizes sensitivity on these inspected recordings.

A.4

Contract sensitivity protocol

Feasibility and causal execution

The feasibility analyses examine whether the recorded observations and candidate predictions can support the mission requirements, while the causal-execution ablation examines how processing delays affect observation availability. The physical-opportunity bounds assume every capture is available at zero latency. The model-union bound additionally allows all candidate models’ stored predictions at every capture, with false-component and report caps relaxed. Failure of this optimistic bound certifies infeasibility for the fixed prediction pool under its assumptions; passing it indicates evidence availability under those optimistic assumptions. The ground-truth-aware hindsight search evaluates action scripts through the resource-accounted causal runtime, starting from evaluated policy and constant schedules. It permits up to 600 candidate evaluations and three sweeps per condition. A successful schedule is a feasibility witness, while an unsuccessful bounded search leaves feasibility unresolved. In the no-skip counterfactual, captures are made available even when inference overlaps; completed predictions enter a serial downstream tracker. Concurrent memory is not charged, and the protocol

The 144-setting sensitivity analysis rescores the same 648 causal reference histories. These histories correspond to nine evaluated strategies across the 72 operating conditions described above. The grid crosses required coverage of 1, 2, or 3 of the focal people; 1, 2, or 3 observations per person; raw false-component caps of 0, 1, 2, or no cap; required reporting versus sightings-only scoring; and deadlines of 36 or 44 seconds. For ℎ required observations, the minimum observation span is 2(ℎ − 1) seconds. This grid changes evaluation rules on fixed histories; it does not rerun policies under all 144 contracts. The one-observation survey relaxation is likewise rescoring-only because the reporting pipeline requires at least two observations. Separate contract-specific reruns expose supported contracts to the policy and processing pipeline. For each setting, the adaptive policy with the most successful conditions is selected retrospectively; ties follow the declared policy order. It is compared with each fixed local and fixed remote Mask R-CNN reference. A material advantage requires at least three more 15

Bhattacharjya et al.

is an analytical ablation rather than a deployable serial implementation. The latency multipliers in Fig. 6 scale configured inference or service latency and the associated local compute energy under the declared power model. Predictions, network transfers, loading costs, and downstream budgets remain fixed.

4/6 CPU runs. Device-generated masks were not retained; scoring the observed schedules with cached predictions and an assumed latency is an offline comparison, not an end-to-end device mission test.

A.6 A.5

Embedded hardware sensitivity

We profiled Mask R-CNN, YOLO11s, and CABiNet using the benchmark perception pipeline on a Jetson AGX Orin in MODE_15W (JetPack 6.1, FP32), with four CPU threads or GPU execution. After five warm-up calls, each model processed 24 captures from two recordings; clocks were not pinned. The local replay latency is the empirical 95th percentile. Power is the sum of the monitored GPU/SOC, CPU/CV, and system 5 V rails; the DDR rail is excluded because it is included in the 5 V rail. The sampler targeted 100 Hz and achieved approximately 60–62 Hz. The replay charges mean monitored power during the profiling block, including idle draw, multiplied by the empirical 95th-percentile latency for each local inference. Idle draw outside inference is not separately charged in this replay; the modeled flight baseline and other costs remain unchanged. These charges are estimates, not directly measured per-call energies. The dynamic-study replay retains the stored predictions, the original contracts and policies, and modeled remote service and network costs. On the 24 sampled captures, CPU masks for all three models matched the stored masks exactly; on the GPU, Mask RCNN matched 21/24 exactly (mean IoU 0.9997), while the other two matched 24/24. This is a sample check rather than full-bank parity. A frame-paced check ran fixed-local Mask R-CNN on the device over six recordings per processor, with non-inference stages emulated and one constant-urgency, favorable-link, ample-battery setting. The resulting capture sets matched replay in 6/6 GPU and

Intent-driven adaptation across mission requirements

To examine how mission requirements can guide perception decisions, we augment the dynamic-study policies with a simple stopping rule and compare them with their original versions. For each mission phase, the stopping rule counts distinct labels in completed runtime reports whose evidence lies within that phase. Once their count reaches the number of people requested by the contract, it skips later perception in that phase. These runtime labels need not identify distinct people; mission success is evaluated separately using target identities and the remaining contract requirements. We compare four contracts (confirmation, strict, revisit, and timecritical), six base policies, three cost profiles, and 72 conditions with the stopping rule enabled and disabled. The 10,368 replays reuse six recordings and stored predictions; remote service and network costs remain modeled. Across 5,184 matched pairs, no success became a failure and three failures became successes. On the Jetson GPU, fixed-local schedules selected by replay with and without the stopping rule were executed over six recordings for each of the confirmation and strict contracts. Inference was framepaced, non-inference stages were emulated, and the stopping rule was not executed online on the device. Mean gross module energy during inference fell from 109.1 to 74.4 J and from 107.9 to 84.6 J, respectively; whole-pass module energy fell from 323.7 to 298.9 J and from 323.8 to 305.3 J. Both include idle draw and exclude flight energy. Mission outcomes were scored offline using stored predictions.

16

Related documents

• QAOA Inside ADMM: Subproblem Wins Do Not Become Decomposition Wins at Matched Budgets#1605226
• Reproducible LoRA Inference Through Deterministic Split-K Reduction#1605228
• Continual Instruction Tuning for LLM-ased Management of Continuum Resources#1605229
• FirstSwap: Pre-Consensus Trade Approvals for Decentralized Market Design#1605230
• PACE: Keeping the Right Tokens in the KV Cache Is Not Enough for Long-Horizon LLM Agents#1605231