ConceptioArchivearXiv CS
arXiv CSopen access

BRAVR: An AP-Assisted Online DRL Mechanism for Interactive VR Bitrate Adaptation over Wi-Fi

Unknown · 2026 · arxiv_cs
arXiv CS · Papers · License: Open Access · 2026
Open Source ↗Direct PDF ↓
distributedsystemsprotocols
networking, internet, protocols, distributed systems

BRAVR: An AP-Assisted Online DRL Mechanism for Interactive VR Bitrate Adaptation over Wi-Fi Miguel Casasnovas, Francesc Wilhelmi, and Boris Bellalta

arXiv:2606.24389v1 [cs.NI] 23 Jun 2026

Department of Engineering Universitat Pompeu Fabra Barcelona, Spain [email protected]

Abstract—Interactive virtual reality (VR) streaming over WiFi requires stringent latency and reliability guarantees, which become increasingly difficult to achieve under dynamic channel conditions and shared medium contention. These challenges make real-time bitrate adaptation a critical yet fundamentally difficult control problem, particularly under limited visibility of the underlying network conditions. This paper formulates VR bitrate adaptation as a network-aware, online decision-making problem and proposes BRAVR, a decentralized deep reinforcement learning (DRL) mechanism designed to optimize visual quality while maintaining streaming performance and promoting airtime fairness in multi-user scenarios. BRAVR integrates application-layer observations with lightweight wireless network statistics collected at the Wi-Fi access point (AP) serving the VR client, enabling more informed bitrate adaptation decisions. We implement BRAVR in a real VR streaming system and evaluate it on a physical Wi-Fi testbed against a strong heuristic baseline and an ablated BRAVR variant without AP assistance. Experimental results show that BRAVR consistently achieves its design objectives, delivering robust quality of service (QoS) and preventing sustained airtime overutilization. It also outperforms its ablated counterpart, highlighting the benefits of incorporating network-level information into the bitrate adaptation control loop. Overall, these results demonstrate the effectiveness of AP-assisted online learning for decentralized interactive VR streaming over commodity Wi-Fi and provide practical insights into bitrate adaptation in shared wireless environments. Index Terms—Virtual Reality, Adaptive Bitrate, Wi-Fi, Deep Reinforcement Learning, Edge-assisted Networking

I. I NTRODUCTION Virtual reality (VR) streaming offloads rendering and encoding to edge or cloud servers, enabling high-fidelity immersive applications on lightweight client devices such as head-mounted displays (HMDs) while reducing on-device computation and energy consumption. To support user mobility and untethered usage, wireless connectivity is therefore essential. Among available wireless technologies, Wi-Fi is a natural candidate due to its ubiquity and access to wide unlicensed spectrum bands, enabling large channel bandwidths and consequently high achievable data rates. As a result, WiFi is, in principle, capable of meeting the stringent quality of service (QoS) requirements of VR streaming, including high frame rates, ultra-low latency, and near-lossless delivery [1]. However, Wi-Fi operates as a contention-based shared wireless medium and is therefore inherently susceptible to interference, congestion, and time-varying channel conditions—factors that

often introduce delay, jitter, and packet loss, ultimately degrading the users’ quality of experience (QoE). Under this setup, video bitrate selection becomes a critical control parameter. High bitrates enhance visual quality but increase network load, potentially causing queueing delays and loss, whereas conservative bitrates may underutilize available capacity—a trade-off typically balanced through adaptive bitrate (ABR) algorithms. Classical ABR approaches rely on hand-crafted heuristics, such as throughput estimation [2] and buffer occupancy rules [3]. While computationally efficient and widely adopted, these methods rely on simplified system assumptions and cannot fully capture the underlying dynamics or adapt to unseen operating conditions. To address this limitations, machine learning (ML), and in particular reinforcement learning (RL), has been increasingly explored for bitrate adaptation [4]–[6], enabling learning adaptive policies that can better handle non-stationary and complex conditions. In parallel to learning-based approaches, edge computing has emerged as a complementary paradigm for improving adaptive bitrate decisions. Edge-assisted adaptation mechanisms have been explored in both standardization (e.g., MPEG-DASH SAND [7]) and research [8]–[10], providing access to network-side measurements that are otherwise unavailable to the client. This additional information can support more informed adaptation decisions, including those made by ML-based approaches. For instance, Wi-DASH [10] applies deep reinforcement learning (DRL) for HTTP-based bitrate adaptation using PHY- and MAC-layer statistics collected from an OpenWrt-based Wi-Fi access point (AP). Despite these advances, most state-of-the-art ABR algorithms have been designed for conventional, non-interactive video streaming scenarios, such as video-on-demand. Their underlying assumptions—particularly regarding buffering and tolerance to latency—do not hold in emerging latency-sensitive applications, such as interactive VR streaming. This incompatibility has motivated the development of dedicated ABR solutions for this domain, such as EV E RE ST [11], [12] and N E S TVR [13], [14]. Nevertheless, existing approaches remain predominantly heuristic-based and rely solely on application-layer measurements. To date, learning-based and network-assisted bitrate adaptation remains largely unexplored in interactive VR streaming. In particular, rather than addressing bitrate adaptation, research in

this domain has focused on complementary aspects: learningbased methods have targeted radio resource allocation [15] and system parameter tuning [16], while edge-assisted approaches have addressed computation offloading [17] and scheduling [18]. In contrast, in less latency-sensitive immersive VR settings, such as panoramic (360◦ ) video streaming, which typically relies on multi-second playback buffers, learning-based bitrate adaptation has been more extensively explored [19]– [22]. This paper formulates bitrate adaptation as an online, nonepisodic learning problem and presents BRAVR, a decentralized AP-assisted BitRate Adaptation algorithm for interactive VR streaming over Wi-Fi based on DRL, aiming to improve streaming quality while discouraging airtime overutilization in multi-user scenarios. The proposed approach is integrated into the open-source Air Light VR (ALVR) [23] platform and evaluated on a physical Wi-Fi testbed against a strong nonlearning baseline, N E S T-VR, as well as an ablated variant without AP assistance. The main contributions of this paper are: • Integration of lightweight Wi-Fi network telemetry from OpenWrt-compatible APs into the bitrate adaptation control loop. • A decentralized DRL-based bitrate adaptation mechanism that leverages application- and network-level information to optimize visual quality while satisfying VR QoS constraints and ensuring airtime fairness in multi-user scenarios. • An experimental evaluation on a physical Wi-Fi testbed demonstrating the effectiveness of online learning and the advantages of incorporating AP-level information into the bitrate adaptation control loop. To the best of our knowledge, this is the first end-to-end integration of online DRL with network-assisted adaptation for interactive VR streaming over commodity Wi-Fi. II. S YSTEM A RCHITECTURE A. System Overview We consider a Wi-Fi BSS supporting one or more interactive VR streaming sessions over a shared wireless medium. Each session consists of a remote streaming server that renders and encodes immersive content and an associated client (an HMD) that receives and displays it. Sessions operate independently and evolve asynchronously, without synchronization between sessions. Communication is bidirectional: the HMD transmits tracking, control, and telemetry data in the uplink, while the server delivers encoded video frames (at the session’s target frame rate), audio streams, control information, and haptic feedback in the downlink. Both uplink and downlink transmissions are highly delay- and loss-sensitive, and impairments in either direction can degrade user experience. For instance, delayed or missing uplink tracking updates degrade pose estimation accuracy and interaction responsiveness, while delayed or incomplete downlink video frames reduce visual continuity and can lead to perceptible artifacts.

HTTP request

1

2

1

1

AP & client stats 2

ABR

ABR

2 HTTP response

Fig. 1: Overview of the VR streaming system architecture.

Among all traffic flows, video traffic is the dominant contributor to network load, and its encoding bitrate—whether fixed or adaptive—largely determines channel occupancy and airtime consumption. Consequently, bitrate decisions directly influence contention and resource availability in the shared wireless medium, coupling the performance of concurrent VR sessions. This makes bitrate selection a key control variable for managing system performance under shared medium constraints. B. Learning-based Policy for Bitrate Adaptation Bitrate adaptation is performed at each streaming server using a decentralized learning-based policy, without centralized coordination or explicit signaling between sessions. Each server executes a periodic control loop that updates the encoding bitrate at fixed decision intervals. At each interval, the server selects the next bitrate based on a set of heterogeneous inputs, including application-level information (e.g., client feedback and server-side streaming statistics) and, when available from an external source, wireless network information that reflects observed channel and medium conditions. The selected bitrate is then applied for the subsequent transmission period. This design enables each session to adapt independently to local conditions while incorporating shared medium information through network-level observations. C. AP-Assisted Network Observability To provide visibility into wireless conditions, the system leverages the Wi-Fi AP as a shared observation point across concurrent sessions. The AP, acting as the final hop in the downlink path, has access to network and wireless channel information that is not directly observable by the client device. Consequently, the AP can expose this information through a lightweight pull-based interface, allowing it to be queried on demand by streaming entities and incorporated into the decentralized bitrate adaptation process. For instance, as illustrated in Fig. 1, connected HMDs act as intermediaries that periodically retrieve AP telemetry and forward it to their associated streaming servers. The servers then incorporate this information into the decentralized bitrate adaptation process together with server-side measurements and client feedback. This design provides a simple and practical mechanism for integrating network-level observability into the control loop, while alternative push-based or publish–subscribe mechanisms could also be adopted for improved scalability.

III. BRAVR: L EARNING F ORMULATION We formulate bitrate adaptation as a sequential decisionmaking problem under uncertainty. The design of the proposed online learning mechanism, BRAVR, is detailed in this section.

observation safe actions observation reward

Environment

Agent

A. Problem Formulation We model the decision process as an infinite-horizon Markov decision process (MDP) [24], defined as M = (S, A, P, r, γ), where S denotes the state space, A the action space of size K, P the (unknown) system dynamics, r the reward function, and γ ∈ (0, 1) the discount factor. At each decision epoch t, the agent observes a state st ∈ S and selects an action at ∈ A according to its policy π. The environment then transitions to a new state st+1 and yields a scalar reward rt . The objective is to learnPa policy that ∞ maximizes the expected discounted return: E [ t=0 γ t rt ]. B. Observation Space In practice, the agent does not have access to the full system state. Instead, it observes an observation vector ot that provides partial information about the underlying system state. At each decision epoch t, the agent observes a continuous vector   hist net ot = oapp , (1) t , o t , ot hist net where oapp t , ot , and ot denote the application-level, history, and network-level feature sets, respectively. These features are computed over the preceding decision interval and normalized to improve numerical stability. The application-level features oapp t capture streaming performance and QoS objectives, and include: the signed deviations from target values of the network frame success ratio (NFR), defined as the fraction of transmitted frames successfully received, and the video frame round-trip time (VF-RTT), defined as the delay between frame transmission and reception of the corresponding receiver feedback message; their shortterm temporal variations, computed as first-order differences between consecutive observations; a latency spike measure, defined as the peak-to-mean VF-RTT deviation; the bitrate efficiency, defined as the ratio of delivered bitrate to target bitrate; and the current bitrate. The history features ohist encode recent control decisions t and include: the holding duration, defined as the number of decision steps since the last bitrate change; and the previous action. The network-level features onet t , derived from AP telemetry, capture the wireless conditions of the corresponding stream and the shared Wi-Fi medium, and include: the stream-specific downlink modulation and coding scheme (MCS), indicating the achievable PHY-layer data rate; the stream-specific packet retransmission rate; the network-wide channel utilization, defined as the fraction of time the medium is busy; the inverse client count, defined as the reciprocal of the number of active VR clients in the BSS; the stream-specific airtime fraction, corresponding to the proportion of channel occupancy attributed to the stream; and the network-wide airtime fairness

action

Fig. 2: Safe reinforcement learning via preemptive shielding. The agent–environment interaction loop.

across all active VR streams, measured using Jain’s fairness index. C. Action Space At each decision epoch t, the agent selects an action at ∈ A based on the observation ot . The action space is defined as: A = {−1, 0, +1},

(2)

corresponding to decreasing, maintaining, or increasing the current encoding bitrate. Bitrate values are selected from a predefined ladder, and each action triggers a transition to an adjacent level (for ±1) or no change (for 0). Compared to selecting absolute bitrate levels—common in ABR streaming—this relative formulation reduces the action space and enables gradual adaptation, mitigating large, abrupt bitrate changes that could degrade QoE. D. Action Shielding To ensure safe online operation, we employ preemptive action shielding [25] (Fig. 2). At each decision epoch, the action set is conditioned on the current observation: bitrate increases are forbidden under unfavorable conditions (i.e., when QoS metrics deviate beyond predefined tolerance thresholds), and invalid actions at the bitrate ladder boundaries are masked. The agent thus selects actions from a filtered set Ã(ot ) ⊆ A, ensuring that unsafe or ineffective decisions are never considered. This mechanism acts as a runtime safety layer independent of policy learning and is particularly suited to interactive, real-time applications, where unsafe actions may cause instability or disconnections. E. Reward Function At each decision epoch t, the agent receives a scalar reward X rt = wb bt − wk pkt , K={rtt,nfr,beh,air}, (3) k∈K

where bt denotes the bitrate, pkt are penalty terms (clipped for stability), and wb and wk are weighting coefficients. The bitrate bt , normalized relative to the bitrate ladder, encourages higher visual quality. A linear min–max normalization of bitrate is adopted for simplicity and interpretability. The video frame round-trip time penalty prtt and the nett work frame success ratio penalty pnfr t penalize deviations from target values under predefined tolerances.

k=0

X

π(a | ot+n )Qθ− (ot+n , a),

(4)

a∈A

with γ ∈ (0, 1) denoting the discount factor, n the return horizon, π the current policy, and θ− the parameters of a slowly updated target network, updated via Polyak averaging with coefficient β [28]. The parameters θ are optimized by minimizing the Huber loss [29] using the Adam optimizer [30] with learning rate α, and gradient norm clipping to ensure stable updates [31]. Action selection follows a policy π defined as a convex mixture of a Boltzmann (softmax) policy πB with temperature T [24] and a uniform distribution πU over admissible actions: π(a | ot ) = (1 − ε) πB (a | ot ) + ε πU (a | ot ),

(5)

where ε ∈ [0, 1] is a fixed mixing coefficient. IV. E XPERIMENTAL S ETUP AND M ETHODOLOGY

...

γ k rt+k + γ n

...

yt =

n−1 X

...

We adopt Deep n-step Expected SARSA [24], [27], a primarily on-policy temporal-difference (TD) method for online adaptation under non-stationary wireless conditions with delayed and temporally coupled effects, where the effect of an action may not be immediately observable and may influence subsequent states and rewards due to wireless network inertia. The n-step return improves temporal credit assignment by propagating delayed rewards over multiple steps, while the expected formulation reduces variance by marginalizing over the current policy. Deep function approximation is used to model the action-value function Qθ (ot , a) via a feedforward neural network (Fig. 3) for high-dimensional continuous observation spaces. Learning is performed via temporal-difference updates based on the TD error δt = yt − Qθ (ot , at ), where yt denotes the n-step expected SARSA target. The target is defined as

Hidden Layers

...

F. Learning Algorithm: Deep n-step Expected SARSA

Input

Output

...

The behavioral penalty pbeh acts as a soft regularizer, penalt izing bitrate decreases under favorable conditions and bitrate increases under unfavorable ones, discouraging unnecessary transitions. The airtime fairness penalty pair t penalizes overutilization of airtime, a finite shared resource, relative to a nominal fair share defined as the available airtime budget equally divided among active VR users. This assumes that concurrent VR streams can reach comparable demand levels, such that exceeding the nominal share indicates unfair resource usage. This linear additive structure follows common design patterns in QoE-oriented ABR formulations but differs fundamentally from classical buffer-centric models such as [2]. Reliability and latency, major contributors to VR QoE [26], are treated as soft objectives rather than hard constraints, avoiding constrained optimization formulations.

Fig. 3: Schematic of the neural network architecture. The observation ot is mapped through two fully connected layers with ReLU activations to estimate the action–value function for all actions. dedicated to VR streaming. Depending on whether the scenario is single-player or multi-player, the setup includes one or two laptops as streaming servers, a single Wi-Fi AP running OpenWrt [32], and one or two Meta Quest 2 HMDs acting as clients. Servers connect to the AP via 1 Gbps Ethernet, while HMDs connect wirelessly over a 40 MHz Wi-Fi 5 (IEEE 802.11ac) channel in the 5 GHz band with 23 dBm transmit power and no multi-user features enabled.1 Users are positioned either near the AP (Location A, approx. −48 dBm), farther away (Location B, approx. −82 dBm), or move between both in mobility scenarios. Hardware details are summarized in Table I. Streaming servers and clients run binaries from a custom ALVR v20.6.0 fork [33] integrating the full learningbased pipeline and AP telemetry retrieval. Video streaming is performed at 90 fps with a two-frame client buffer (as recommended in [12]). AP-level telemetry collection and exchange: The OpenWrtbased AP runs a lightweight monitoring service that collects wireless statistics using standard utilities (iw, iwinfo, ip). It exposes this information through a stateless HTTP interface over TCP using short-lived request–response interactions, and returns a compact JSON payload containing channel-level metrics (e.g., active and busy times), per-client wireless metrics (e.g., RSSI, downlink/uplink transmission durations, and MCS), and metadata identifying active VR clients. In the considered endpoint-driven implementation (Fig. 1), HMDs periodically issue HTTP GET requests to the AP and receive the corresponding responses. The retrieved telemetry is then forwarded to the streaming server through the system’s existing control messaging channel, without introducing a separate transport mechanism. The communication overhead of this exchange is 4–8 KB/s, which is negligible compared to video traffic. B. Experimental Methodology BRAVR (henceforth denoted as BRAVR+ ) is evaluated against BRAVR− , an ablated variant that excludes AP-level

A. Experimental Setup Experiments are conducted in a university office environment using a controlled Wi-Fi testbed with a single BSS

1 In Wi-Fi 6 mode, the AP provided sufficient capacity for stable high-bitrate streaming, obviating adaptation and motivating the use of a more constrained Wi-Fi 5 configuration.

TABLE I: Equipment details. 1x Laptop a

Model OS GPU CPU

Dell G15 5521 Windows 11 x64 NVIDIA GeForce RTX 3060 Mobile Intel Core i7-12700H

1x Laptop b

Model OS GPU CPU

Dell Pro Max 16 Windows 11 x64 NVIDIA RTX Pro 1000 Blackwell Intel Core Ultra 7 265H

1x AP

Model Firmware

ASUS TUF-AX4200 OpenWrt 23.05.5

2x HMD

Model

Meta Quest 2

TABLE II: Learning hyperparameters. Symbol

Value

Meaning

n γ α β T H εtrain εeval ∥∇∥2

3 0.8 10−3 0.01 0.25 2 × 128 0.05 0 1.0

n-step return horizon discount factor learning rate Polyak averaging coefficient softmax temperature hidden layers (units per layer) mixing coefficient (training) mixing coefficient (evaluation) gradient L2 norm clipping threshold

wb wrtt wnfr wbeh wair

1.0 1.5 1.5 0.2 10.0

bitrate reward weight latency weight reliability weight behavior weight airtime fairness weight

information, and N E S T-VR [13], [14], a heuristic VR ABR algorithm that adjusts bitrate in discrete steps using probabilistic threshold-based rules based on latency and reliability signals. N E S T-VR serves as a strong non-learning baseline, having demonstrated robust performance across single- and multi-user scenarios and consistently outperforming constantbitrate (CBR) strategies [34], [35]. A CBR configuration is additionally included in multi-user scenarios as a non-adaptive reference. All adaptive methods operate over the same bitrate ladder, ranging from 10 to 150 Mbps in 10 Mbps steps, with an initial bitrate of 10 Mbps. Bitrate decisions are made every 0.5 s, whereas AP telemetry is collected every 0.25 s. N E S T-VR, in particular, uses its Balanced profile for symmetric increase and decrease adjustments. For both learning-based methods, QoS targets are set to 22 ms (VF-RTT) and 0.99 (NFR), following recommended values for satisfactory user experience [14]. The corresponding deviation tolerances are set to 22 ms (VFRTT) and 0.05 (NFR), defining the scaling of the penalty terms and the boundary at which the penalty term reaches one (i.e., at a VF-RTT of 44 ms and an NFR of 0.94). These targets also define N E S T-VR adaptation thresholds, ensuring a consistent basis for comparison. Table II summarizes the hyperparameters used by learning-based methods. The reward design prioritizes latency and reliability over bitrate while enforcing airtime fairness in multi-user scenarios. Training and evaluation procedure. Learning-based methods

operate and learn online, without simulation-based pretraining, and retain learned model parameters across sessions. This avoids potential simulation-to-reality mismatches and reflects incremental policy refinement through continued realworld operation. In single-user scenarios, the learning-based approaches are trained over seventeen independent sessions of 240 s each, balancing sufficient training with manageable experimental overhead. In multi-user scenarios, each learningbased method is trained using a single continuous 20-minute session per configuration. All methods, including N E S T-VR and CBR, are evaluated over three independent 240 s sessions. Unless otherwise stated, all reported results correspond exclusively to these evaluation sessions. The dataset is publicly available [36]. Utility. To assess the trade-off between bitrate and QoS objectives, we define an instantaneous utility at each decision interval t: nfr ui (t) = ri (t) · srtt (6) i (t) · si (t), where ri (t) denotes the achieved bitrate of user i at time t, and nfr srtt i (t), si (t) ∈ [0, 1] quantify the degree of satisfaction of the VF-RTT and NFR targets (22 ms and 0.99), respectively. Each satisfaction factor equals one when the corresponding target is met and decreases linearly with the magnitude of the violation, reaching zero at the maximum tolerated values (44 ms and 0.95). The utility for each user, Ui , is then defined as the time average of ui (t) over the evaluation period. Airtime-aware utility. In multi-user scenarios, we also account for fair resource usage through the airtime satisfaction rate Siair ∈ [0, 1], defined as the fraction of time intervals during which user i’s airtime does not exceed its nominal fair share. The airtime-aware utility is then given by Uiair = Ui · Siair . V. R ESULTS This section presents the evaluation results in both singleuser and multi-user scenarios. A. Single-User Scenario This scenario involves single VR user moving between Location A and Location B at a walking pace (∼ 1.6 m/s) and experiencing time-varying channel conditions due to changes in distance and surrounding environment. The user trajectory is divided into four 60 s stages: stationary at A, moving to B (A→B), stationary at B, and returning to A (B→A). This enables the evaluation of adaptation under both degrading and improving channel conditions. Table III reports the average performance across stationary and mobility stages. Fig. 4 reports the stage-wise average utility, capturing the combined effect of achieved bitrate and QoS compliance. Performance. All approaches maintain QoS largely within target bounds across stages. Under favorable conditions, i.e., stationary at Location A, constraints are consistently satisfied. During mobility, moderate reliability degradation is observed for the learning-based approaches, particularly during B→A

TABLE III: Single-user performance under stationary and mobility conditions. Metrics are reported as mean ± standard deviation over evaluation sessions. Cell text colors indicate QoS compliance: black (within target), blue (tolerated deviation), red (beyond limits). Bold values indicate the best performance among methods. (a) Stationary conditions. BRAVR−

NeSt-VR

BRAVR+

A

B

A

B

A

B

Target bitrate

115.3 ± 4.3

97.5 ± 28.6

108.2 ± 2.5

69.9 ± 8.8

122.6 ± 3.5

110.1 ± 11.1

NFR VF-RTT

100.0 ± 0.0 11.7 ± 0.9

97.6 ± 3.2 19.1 ± 4.7

99.9 ± 0.1 12.8 ± 1.1

98.6 ± 1.9 21.1 ± 1.7

100.0 ± 0.0 12.4 ± 0.6

97.6 ± 1.5 21.3 ± 2.2

PLR

0.0 ± 0.0

507.3 ± 871.3

1.1 ± 1.3

31.3 ± 37.4

0.0 ± 0.0

23.2 ± 18.2

Switch rate

18.4 ± 2.3

39.1 ± 9.8

58.9 ± 5.2

63.8 ± 2.2

52.5 ± 11.8

67.9 ± 8.5

(b) Mobility conditions. BRAVR−

NeSt-VR

BRAVR+

A→B

B→A

A→B

B→A

A→B

B→A

Target bitrate

107.3 ± 4.0

106.3 ± 6.5

108.2 ± 3.6

102.4 ± 14.2

122.9 ± 2.6

115.6 ± 6.0

NFR VF-RTT

99.0 ± 0.6 18.2 ± 2.7

99.8 ± 0.1 15.3 ± 0.4

98.8 ± 0.9 21.6 ± 3.2

95.8 ± 4.6 18.9 ± 5.2

97.8 ± 1.8 19.5 ± 1.4

98.0 ± 2.2 18.1 ± 2.3

PLR

13.2 ± 11.5

6.1 ± 8.3

18.9 ± 9.7

147.3 ± 227.1

21.2 ± 18.2

22.8 ± 29.1

Switch rate

48.7 ± 8.2

45.7 ± 8.1

64.3 ± 4.9

65.2 ± 12.0

78.2 ± 8.3

70.7 ± 14.1

Target bitrate [Mbps], Network frame success ratio (NFR) [%], Video frame round-trip time (VF-RTT) [ms], Packet loss ratio (PLR) [×10−5 ], Bitrate switching rate [changes per minute].

for BRAVR− as it does not leverage AP-provided information. Under degraded conditions, i.e., B, all approaches experience reduced reliability, although performance remains within acceptable limits. These differences stem from the underlying control strategies. The learning-based approaches treat QoS targets as soft constraints, allowing temporary deviations to improve long-term performance. In contrast, N E S T-VR enforces QoS targets via threshold-based decisions, resulting in stricter adherence but more conservative bitrate selection. Accordingly, its decisions remain more closely aligned with observed conditions, whereas learning-based policies may temporarily select actions that are suboptimal for the current state in favor of maximizing cumulative return. Despite this, BRAVR+ consistently achieves the highest bitrate levels across stages while maintaining QoS within target bounds, resulting in the highest utility values across all stages. In particular, it achieves a mean relative improvement of +21.2% in achieved bitrate and +3.8% in QoS satisfaction (i.e., the fraction of samples satisfying both QoS targets) over BRAVR− , along with higher environment reward (median over all evaluation samples: 0.64 vs. 0.58). Compared to N E S T-VR, which also outperforms BRAVR− , BRAVR+ attains consistently higher bitrate, while N E S T-VR exhibits tighter QoS adherence under mobility conditions. Bitrate adaptation. Learning-based approaches exhibit more frequent target bitrate updates due to continuous state-dependent action selection and the absence of explicit switching hysteresis. This results in incremental, finegrained adjustments in response to moderate fluctuations in observed conditions, which manifest as higher switching rates.

Policy behavior. A post-hoc feature importance analysis, based on permutation importance applied to a Random Forest regressor approximating the learned policies during evaluation sessions, shows that latency-related features are the primary drivers of BRAVR+ decisions: VF-RTT deviation from target (importance 0.40), its short-term trend (0.25), and the peak-tomean VF-RTT measure (0.16). AP-derived features, including channel utilization (0.24) and stream airtime fraction (0.22), also play a significant role in the policy’s decisions. In contrast, PHY-layer indicators such as MCS (0.01) and fairness-related features, i.e., Jain’s index (0.00), have negligible influence. Takeaway: BRAVR+ treats QoS requirements as soft constraints, permitting controlled violations to improve long-term bitrate–QoS trade-offs. Under single-user mobility, it consistently maintains acceptable average QoS while achieving higher bitrate operation. Incorporating AP telemetry improves adaptation efficiency and robustness to changing wireless conditions, as medium-usage signals provide complementary information to end-to-end observations. B. Multi-User Scenario This scenario involves two users, each running an independent streaming session over a shared wireless channel with identical bitrate adaptation strategies. Unlike the singleuser case, performance becomes interdependent, as contention for the shared medium couples the users’ bitrate adaptation dynamics. Two user placement configurations are considered: (i) near–near, where both users are at Location A and experience similarly favorable channel conditions, and (ii) near–far,

TABLE IV: Multi-user performance in the near–near and near–far scenarios. Metrics are reported as mean ± standard deviation over evaluation sessions. Cell text colors indicate QoS compliance: black (within target), blue (tolerated deviation), red (beyond limits). Bold values indicate the best performance among methods. Airtime values are reported as the percentage of total channel time occupied by each user. (a) Both users (U1, U2) near the access point. CBR

BRAVR−

N E S T-VR

BRAVR+

U1

U2

U1

U2

U1

U2

U1

U2

Target bitrate

80.0 ± 0.0

80.0 ± 0.0

71.2 ± 1.6

68.3 ± 2.5

77.1 ± 5.2

65.4 ± 6.5

66.9 ± 6.8

62.5 ± 4.9

NFR VF-RTT

88.0 ± 5.8 51.0 ± 13.0

62.4 ± 29.9 97.1 ± 42.6

99.8 ± 0.1 15.7 ± 0.5

99.5 ± 0.1 16.8 ± 0.3

99.1 ± 0.5 16.9 ± 0.6

99.2 ± 0.3 16.4 ± 1.4

100.0 ± 0.1 11.2 ± 1.9

100.0 ± 0.1 11.3 ± 1.8

Airtime

46.3 ± 0.7

47.8 ± 0.6

45.1 ± 0.5

44.6 ± 1.4

45.6 ± 1.6

44.9 ± 2.5

44.5 ± 2.7

43.4 ± 2.6

PLR Switch rate

1253.9 ± 374.5 0.0 ± 0.0

3804.2 ± 1107.3 0.0 ± 0.0

3.7 ± 1.9 21.6 ± 0.7

8.5 ± 3.5 24.5 ± 0.9

13.4 ± 7.0 58.7 ± 1.3

14.2 ± 5.9 61.9 ± 1.9

0.1 ± 0.1 63.2 ± 4.9

0.6 ± 0.9 57.5 ± 3.7

(b) One user near the access point (U1) and one user far (U2). CBR

BRAVR−

N E S T-VR

U1

U2

U1

Target bitrate

70.0 ± 0.0

70.0 ± 0.0

NFR VF-RTT

99.6 ± 0.0 19.8 ± 0.2

88.5 ± 0.8 47.0 ± 0.8

Airtime

41.8 ± 0.2

PLR Switch rate

22.0 ± 31.1 0.0 ± 0.0

BRAVR+

U2

U1

U2

U1

U2

92.9 ± 17.2

28.0 ± 3.0

89.1 ± 6.9

99.2 ± 0.0 15.6 ± 0.9

99.8 ± 0.2 15.3 ± 0.4

98.0 ± 0.8 17.3 ± 0.9

27.0 ± 7.5

55.4 ± 1.7

22.2 ± 2.5

98.8 ± 0.6 26.5 ± 5.2

100.0 ± 0.0 7.9 ± 0.8

99.3 ± 0.9 13.0 ± 6.7

51.8 ± 0.1

58.4 ± 8.2

29.7 ± 5.6

57.0 ± 1.36

30.9 ± 3.4

44.1 ± 1.3

31.7 ± 0.6

1495.9 ± 268.3 0.0 ± 0.0

9.6 ± 1.3 25.1 ± 1.4

45.9 ± 32.9 20.9 ± 3.4

69.3 ± 74.8 57.5 ± 6.6

92.7 ± 79.3 52.1 ± 13.0

0.1 ± 0.2 59.5 ± 4.0

49.6 ± 64.5 51.4 ± 3.4

BRAVR+

BRAVR−

NeSt-VR

150 125.3

Utility

116.6 106.9

93.5

100

100.5

91.7 93.5

87.2

96.8

84.8

97.6

60.2

50

0

A

A→B

B

B→A

Fig. 4: Utility across mobility stages in the single-user scenario. Values are averaged over evaluation sessions. Error bars indicate standard deviation. where User 1 is at Location A while User 2 is at Location B, creating asymmetric channel conditions. 1) Both Users Near the AP (near–near): Table IVa summarizes the average per-user performance for all adaptive strategies and a CBR baseline operating at 80 Mbps per user (higher symmetric rates led to disconnections). Fig. 5 shows the corresponding airtime-aware utility, capturing the effective application-level performance under fair airtime operation. Performance. The CBR baseline leads to sustained QoS degradation for both users. At 80 Mbps per user, the network operates near saturation. This increases latency and packet/frame loss, leading to noticeable performance impairments. In contrast, all adaptive approaches maintain QoS largely within

Airtime-aware Utility

Target bitrate [Mbps], Network frame success ratio (NFR) [%], Video frame round-trip time (VF-RTT) [ms], Airtime [%], Packet loss ratio (PLR) [×10−5 ], Bitrate switching rate [changes per minute].

User 1

40 30

User 2 31.4 32.5

29.4 28.6

28.1 26.6

NeSt-VR

BRAVR−

20 10 0

BRAVR+

Fig. 5: Airtime-aware utility per user in the multi-user near–near scenario. Values are averaged over evaluation sessions. Error bars indicate standard deviation.

target bounds while achieving comparable average per-user bitrate, indicating efficient utilization of the available capacity. BRAVR+ , in particular, operates at slightly lower bitrate levels but delivers more reliable performance, achieving QoS satisfaction rates above 90%. This corresponds to relative gains of +13.2% over N E S T-VR and +19.7% over BRAVR− , at the cost of modest reductions in achieved bitrate of −7.8% and −8.2%, respectively. Moreover, it maintains average per-user airtime below the nominal fair-share level (corresponding to approximately 45% airtime per user in our setup), demonstrating long-term resource regulation. Although all approaches frequently operate near the nominal fair-share airtime limits and exhibit occasional transient overuse, BRAVR+ reduces

Airtime-aware Utility

User 1 (near)

26.6

30

24.2

19.5

20 10 0

User 2 (far)

8.0

NeSt-VR

17.7

10.0

BRAVR−

BRAVR+

Fig. 6: Airtime-aware utility per user in the multi-user near–far scenario. Values are averaged over evaluation sessions. Error bars indicate standard deviation.

the frequency of such events. As a result, it attains the highest airtime-aware utility across users. Bitrate adaptation. Learning-based approaches perform more frequent target bitrate adjustments, resulting in less stable bitrate trajectories than the non-learning baseline. In decentralized and uncoordinated settings, this increased adaptation frequency can induce oscillatory interactions, as bitrate increases by one user reduce the capacity available to the other, triggering compensatory adaptations. Policy behavior. Feature importance analysis shows that QoSrelated signals dominate BRAVR+ ’s decision-making across both users, with NFR deviation from target (0.31 and 0.48), VF-RTT short-term trend (0.24 and 0.26), and VF-RTT deviation (0.08) being the most influential features. AP- and medium-related features, such as stream airtime fraction (0.02) and channel utilization (0.02), have a minor influence on the policy. Takeaway: BRAVR+ maintains high QoS while regulating long-term resource usage across users. Static bitrate configurations, in contrast, may lead to persistent QoS degradation when configured at overly high rates, exceeding the available capacity. 2) One User Near, One User Far (near–far): Due to increased distance and propagation loss, the far user (User 2) experiences lower PHY rates as determined by the AP’s rate adaptation. As a result, it requires more airtime per delivered bit, increasing channel occupancy and reducing overall medium efficiency. This, in turn, can limit the performance of the near user (User 1) [37], motivating bitrate adaptation strategies that account for both QoS requirements and sharedmedium constraints. Table IVb reports the average per-user performance for all adaptive strategies and a CBR baseline operating at 70 Mbps per user (higher symmetric rates led to frequent disconnections). Fig. 6 shows the corresponding airtime-aware utility, while Fig. 7 depicts the per-user airtime evolution for N E S T-VR and BRAVR+ . Performance. The CBR baseline maintains acceptable QoS for the near user but leads to severe QoS degradation for the far user due to insufficient effective link capacity. In contrast, all adaptive methods maintain QoS within acceptable

bounds for the far user despite degraded link conditions while preserving stable performance for the near user. N E S T-VR and BRAVR− do not explicitly regulate airtime usage; instead they opportunistically exploit all available capacity. This yields higher target bitrates for the near user, but also persistent airtime overutilization, as illustrated in Fig. 7a for N E S TVR. While this increases instantaneous bitrate achieved, it also raises contention and drives the medium close to its capacity limit, causing transient QoS degradation. In contrast, BRAVR+ decreases the near user’s bitrate as a result of its airtime-aware regulation, preventing sustained disproportionate airtime usage, as illustrated in Fig. 7b. This reduces contention and stabilizes medium conditions, improving QoS satisfaction rates (+16.3% and +23.1% relative improvements over N E S T-VR and BRAVR− for the near user, and +6.8% and +20.5% for the far user, respectively). Overall, BRAVR+ achieves the highest aggregate utility across users, driven primarily by substantial gains for the near user resulting from controlled airtime usage and preserved QoS. For the far user, utility is comparable to other methods but slightly lower due to reduced bitrate, since the additional airtime made available does not translate into higher achieved rates. This may be because, under poor channel conditions and in the presence of contention, higher bitrate levels are less reliable due to increased sensitivity to retransmissions and short-term channel fluctuations. As a result, airtime-aware gains depend on whether users can convert available airtime into stable QoS improvements, which is ultimately constrained by channel conditions and AP-side rate adaptation. Policy behavior. Feature importance analysis indicates that, for the near user, BRAVR+ ’s decisions are mainly driven by current bitrate (0.67), as it reflects medium occupancy and guides the algorithm’s adjustments to prevent airtime overuse. Latency-related features, including VF-RTT deviation from target (0.27) and its short-term trend (0.06), as well as fairness measures (0.11), also contribute, while stream airtime fraction (0.03) and channel utilization (0.02) have minor influence. For the far user, decisions are primarily driven by QoSrelated signals, as latency and reliability constraints dominate under channel-limited operation. In particular, latency-related features, including peak-to-mean VF-RTT (0.82), its shortterm trend (0.49), and VF-RTT deviation from target (0.05), and reliability-related features, including NFR deviation from target (0.15), are the most influential factors. In contrast, bitrate- and medium-related features have comparatively lower contribution.

Takeaway: BRAVR+ maintains high QoS for both users while regulating airtime usage for the near user, preventing persistent medium overutilization and stabilizing operation. However, results indicate that, under adverse channel conditions, reducing airtime consumption on higher-quality links may not necessarily improve performance for users with weaker channel conditions.

User 1 (near)

Airtime (%)

100

User 2 (far)

80

This work was supported by the following projects: MLDR (Chist-ERA WAI 2022) PCI2023-1459582 (MCIU/AEI/10.13039), REALM (GA 101298050 European Union), TRUE Wi-Fi PID2024-155470NB-I00 (MICIU/AEI/10,13039/501100011033/FEDER,UE), ICREA Academia 2024 (00077 AGAUR), and MdM CEX2021001195-M (MICIU/AEI/10.13039/501100011033). Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union. Neither the European Union nor the granting authority can be held responsible for them. The authors would like to thank Hugo Sanz Benito for their valuable support in conducting the multi-user tests.

60

R EFERENCES

80 60 40 20 0

0

50

100

150

200

Time (s)

(a) N E S T-VR User 1 (near)

100

Airtime (%)

VII. ACKNOWLEDGEMENTS

User 2 (far)

40 20 0

0

50

100

150

200

Time (s)

(b) BRAVR+

Fig. 7: Per-user airtime evolution in the multi-user near– far scenario during representative evaluation sessions.

VI. C ONCLUSIONS This paper presents BRAVR, a decentralized, AP–assisted reinforcement learning approach for real-time virtual reality bitrate adaptation over Wi-Fi, leveraging both application- and AP-level information to jointly optimize visual quality, latency, reliability, and multi-user airtime fairness. Experimental results demonstrate that BRAVR achieves robust quality of service across single- and multi-user scenarios while preventing sustained airtime overutilization. Compared to its ablated variant, it consistently improves performance, highlighting the value of network-assisted information, and attains performance comparable to or better than a strong heuristic baseline (N E S TVR). These findings indicate that AP-level visibility enables more informed, airtime-aware adaptation and that BRAVR constitutes a practical solution for decentralized VR streaming over Wi-Fi. Future work includes evaluating the approach under higher user densities and more complex wireless environments (e.g., overlapping BSSs), where airtime- and shared-medium awareness are increasingly critical. Another direction is exploring multi-objective reinforcement learning to reduce reliance on manually tuned reward weights and enable more flexible trade-offs among competing objectives. Finally, investigating coordinated or centralized control, using the AP as a decision or information hub, represents a promising direction for improving global efficiency.

[1] M. F. Hossain, A. Jamalipour, K. Munasinghe, A Survey on Virtual Reality over Wireless Networks: Fundamentals, QoE, Enabling Technologies, Research Trends and Open Issues, Authorea Preprints (2023). [2] X. Yin, A. Jindal, V. Sekar, B. Sinopoli, A control-theoretic approach for dynamic adaptive video streaming over HTTP, in: Proceedings of the 2015 ACM Conference on Special Interest Group on Data Communication, 2015, pp. 325–338. [3] K. Spiteri, R. Urgaonkar, R. K. Sitaraman, BOLA: Near-optimal bitrate adaptation for online videos, IEEE/ACM transactions on networking 28 (4) (2020) 1698–1711. [4] H. Mao, R. Netravali, M. Alizadeh, Neural adaptive video streaming with pensieve, in: Proceedings of the conference of the ACM special interest group on data communication, 2017, pp. 197–210. [5] N. A. Hafez, M. S. Hassan, T. Landolsi, Reinforcement learning-based rate adaptation in dynamic video streaming, Telecommunication Systems 83 (4) (2023) 395–407. [6] M. Naresh, P. Saxena, M. Gupta, Ppo-abr: Proximal policy optimization based deep reinforcement learning for adaptive bitrate streaming, in: 2023 International Wireless Communications and Mobile Computing (IWCMC), IEEE, 2023, pp. 199–204. [7] Information technology — Dynamic adaptive streaming over HTTP (DASH) — Part 5: Server and network assisted DASH (SAND), https: //www.iso.org/standard/78448.html, amendment 1 to ISO/IEC 230095:2017 (2020). [8] A. Mehrabi, M. Siekkinen, A. Ylä-Jääski, Edge computing assisted adaptive mobile video streaming, IEEE Transactions on Mobile Computing 18 (4) (2018) 787–800. [9] J. W. Kleinrouweler, S. Cabrero, P. Cesar, Delivering stable high-quality video: An SDN architecture with DASH assisting network elements, in: Proceedings of the 7th International Conference on Multimedia Systems, 2016, pp. 1–10. [10] W. Wu, J. Yuan, S. Ma, M. Yang, AP-assisted adaptive video streaming in wireless networks with high-density clients, Computer Communications 219 (2024) 53–63. [11] M. Liubogoshchev, E. Korneev, E. Khorov, EVeREst: Bitrate adaptation for cloud VR, Electronics 10 (6) (2021) 678. [12] E. Korneev, M. Liubogoshchev, D. Bankov, E. Khorov, How to Model Cloud VR: An Empirical Study of Features That Matter, IEEE Open Journal of the Communications Society (2024). [13] F. Maura, M. Casasnovas, B. Bellalta, Experimenting with adaptive bitrate algorithms for virtual reality streaming over Wi-Fi, in: Proceedings of the 30th Annual International Conference on Mobile Computing and Networking, 2024, pp. 1930–1937. [14] M. Casasnovas, F. Maura, I. Vandebroeck, H. Sukmawanto, E. Joris, B. Bellalta, NeSt-VR: An Adaptive Bitrate Algorithm for Virtual Reality Streaming over Wi-Fi, arXiv preprint arXiv:2502.14947 (2025). [15] G. Kougioumtzidis, V. K. Poulkov, P. I. Lazaridis, Z. D. Zaharis, Deep Reinforcement Learning-Based Resource Allocation for QoE Enhancement in Wireless VR Communications, IEEE Access (2025). [16] Y. Sun, J. Chen, Z. Wang, M. Peng, S. Mao, Enabling mobile virtual reality with open 5g, fog computing and reinforcement learning, IEEE Network 36 (6) (2022) 142–149.

[17] R. Singh, R. Sukapuram, S. Chakraborty, Mobility-aware multi-access edge computing for multiplayer augmented and virtual reality gaming, in: 2022 IEEE 21st International Symposium on Network Computing and Applications (NCA), Vol. 21, IEEE, 2022, pp. 191–200. [18] S.-J. Lu, W.-X. Chen, Y.-S. Su, Y.-S. Chang, Y.-W. Liu, C.-Y. Li, G.-H. Tu, Practical Latency-Aware Scheduling for Low-Latency Elephant VR Flows in Wi-Fi Networks, in: 2024 IEEE International Conference on Pervasive Computing and Communications (PerCom), IEEE, 2024, pp. 57–68. [19] Z. Jiang, X. Zhang, Y. Xu, Z. Ma, J. Sun, Y. Zhang, Reinforcement learning based rate adaptation for 360-degree video streaming, IEEE Transactions on Broadcasting 67 (2) (2020) 409–423. [20] N. Kan, J. Zou, C. Li, W. Dai, H. Xiong, RAPT360: Reinforcement learning-based rate adaptation for 360-degree video streaming with adaptive prediction and tiling, IEEE Transactions on Circuits and Systems for Video Technology 32 (3) (2021) 1607–1623. [21] W. Quan, Y. Pan, B. Xiang, L. Zhang, Reinforcement learning driven adaptive vr streaming with optical flow based qoe, arXiv preprint arXiv:2003.07583 (2020). [22] Y. Li, Federated deep reinforcement learning-based caching and bitrate adaptation for VR panoramic video in clustered MEC networks, Electronics 11 (23) (2022) 3968. [23] ALVR Project, ALVR (Air Light VR), https://github.com/alvr-org/ ALVR, accessed: May 20, 2026 (2026). [24] R. S. Sutton, A. G. Barto, et al., Reinforcement learning: An introduction, Vol. 1, MIT press Cambridge, 1998. [25] M. Alshiekh, R. Bloem, R. Ehlers, B. Könighofer, S. Niekum, U. Topcu, Safe reinforcement learning via shielding, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 32, 2018. [26] H. S. Rossi, K. Mitra, C. Åhlund, I. Cotanis, QoE Models for Virtual Reality Cloud-based First Person Shooter Game over Mobile Networks, in: 2024 20th International Conference on Network and Service Management (CNSM), IEEE, 2024, pp. 1–5. [27] H. Van Seijen, H. Van Hasselt, S. Whiteson, M. Wiering, A theoretical and empirical analysis of expected sarsa, in: 2009 ieee symposium on adaptive dynamic programming and reinforcement learning, IEEE, 2009, pp. 177–184. [28] B. T. Polyak, A. B. Juditsky, Acceleration of stochastic approximation by averaging, SIAM journal on control and optimization 30 (4) (1992) 838–855. [29] P. J. Huber, Robust estimation of a location parameter, in: Breakthroughs in statistics: Methodology and distribution, Springer, 1992, pp. 492–518. [30] D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014). [31] R. Pascanu, T. Mikolov, Y. Bengio, On the difficulty of training recurrent neural networks, in: International conference on machine learning, Pmlr, 2013, pp. 1310–1318. [32] OpenWrt Project, OpenWrt, https://openwrt.org, accessed: May 20, 2026 (2026). [33] M. Casasnovas, BRAVR-DRL: AP-Assisted Deep Reinforcement Learning for VR Bitrate Adaptation over Wi-Fi, https://github.com/ miguelcUPF/BRAVR DRL (2026). [34] M. Casasnovas, M. Carrascosa-Zamacois, B. Bellalta, Can cloud-based VR streaming handle Wi-Fi OBSS contention?, in: 2025 IEEE Conference on Standards for Communications and Networking (CSCN), IEEE, 2025, pp. 1–6. [35] B. Bellalta, M. Casasnovas, F. Maura, A. Rodrı́guez, J. S. Marquerie, P. L. Garcı́a, F. Wilhelmi, J. Blat, Understanding the Wi-Fi and VR streaming interplay: A comprehensible simulation and experimental study, Journal of Network and Computer Applications (2025) 104391. [36] M. Casasnovas, F. Wilhelmi, B. Bellalta, AP-Assisted VR Streaming Dataset (BRAVR) (2026). doi:10.5281/zenodo.20072438. URL https://doi.org/10.5281/zenodo.20072438 [37] C. Michaelides, M. Casanovas, D. Núñez, B. Bellalta, Lessons learned from a large-scale virtual reality experience over Wi-Fi, IEEE Transactions on Networking (2025).

Record · ID 303169 · SHA-256 fd4faf7244a40aec
Retrieved via Conceptio — every document is proof-bundled with source, license, and retrieval metadata.