TARMM: Scaling Delay-Critical Edge AI Offloading in 5G O-RAN via Temporal Graph Mobility Management Peihao Yan
Yun Chen
Jie Lu
Qijun Wang
Huacheng Zeng
Michigan State University
arXiv:2604.24501v1 [cs.NI] 27 Apr 2026
Abstract
As a user equipment (UE) moves across cells, handover introduces transient disruptions in connectivity, resource reallocation, and scheduling, often resulting in delay spikes and increased jitter. While handover decisions are made at the level of individual UEs, they are inherently influenced by global network conditions, such as target cell load and interference level. For example, selecting a target cell with strong signal strength but high congestion may increase queuing delay, whereas choosing a lightly loaded but distant cell may degrade radio link quality. The coupling between local mobility decisions and global network dynamics makes delay-aware handover management a challenging problem. While handover management has been extensively studied under various objectives (e.g., maximizing throughput [27], improving load balancing [11, 33], enhancing fairness [35], and reducing signaling overhead [38]), most existing work relies on simplified network models or simulated datasets that are incapable of capturing the complexity and dynamics of real-world networks. Moreover, most existing approaches were evaluated through numerical analysis, simulation, or network emulation, which often do not reflect the operational constraints and variability of deployed systems. The design and evaluation of handover management strategies in realistic network environments remain underexplored, particularly for delay-critical edge AI applications. In this paper, we present TARMM, a 5G Open Radio Access Network (O-RAN) system with optimized UE mobility management for delay-critical edge AI offloading. TARMM focuses on the handover optimization for cell-edge UEs, with the aim of jointly minimizing uplink and downlink communication delays while respecting UE’s demands and cell load constraints. Our key insight is that effective handover decisions must jointly consider temporal dynamics and spatial dependencies of the entire network. To this end, we design a handover framework that combines temporal graph representation learning, multi-agent reinforcement learning (MARL), and rule-based control mechanisms to determine when, which target cell, and how to perform handover of individual UEs in dense small cell networks. One core component of TARMM is a temporal graph model. To effectively capture the complex interactions in multi-cell networks, we model the RAN as a temporal graph where both UEs and RUs are represented as nodes, and their timevarying interactions form dynamic edges. Each UE aggregates
Emerging delay-critical edge AI applications, such as VR perception and real-time video analytics, impose stringent latency and reliability requirements on 5G networks. However, existing mobility management mechanisms are largely reactive and fail to adapt to dynamic network conditions, resulting in suboptimal handover decisions and degraded performance. In this paper, we present TARMM, a 5G Open Radio Access Network (O-RAN) system that optimizes user mobility management for delay-critical edge AI offloading. The core of TARMM is a temporal graph model that captures the spatiotemporal dynamics of the RAN across users and cells, enabling near real-time handover decisions. Building on this representation, we design a multi-agent reinforcement learning (MARL) framework with rule-based action masking and proactive resource preparation to ensure safe, stable, and efficient handovers. We implement TARMM on a multicell indoor 5G O-RAN testbed and evaluate it using diverse VR workloads. Extensive experiments show that TARMM reduces tail latency by up to 44% and packet loss by up to 56% compared to state-of-the-art approaches.
1
Introduction
Mobile devices such as smartphones, AR/VR headsets, and mobile robots are fundamentally constrained by limited battery capacity and on-device computational resources. As a result, advanced AI applications increasingly rely on edge AI offloading, where sensor data are transmitted to nearby edge GPU servers for inference, and results are returned in real time to support decision-making, rendering, and control. For delaycritical applications, communication latency is not merely a performance metric but a fundamental functional requirement. Applications such as immersive AR/VR and real-time robotics operate in closed-loop systems, where sensing, inference, and actuation must occur within strict latency bounds. Even short-lived latency spikes or jitter can disrupt user experience, degrade control stability, or lead to system-level failures. Therefore, achieving low and stable communication delays is essential for edge AI offloading. The communication delay in cellular networks is determined by many factors such as spectrum availability, radio link quality, cell load, and scheduling policies. Among these, handover plays a particularly critical role for mobile users. 1
cross-layer information from neighboring cells, including radio measurements (SINR, RSRP, CQI), traffic statistics, and cell load conditions. A Temporal Graph Network (TGN) is employed to learn embeddings that encode both spatial dependencies (e.g., inter-cell interference and topology) and temporal dynamics (e.g., mobility patterns and traffic evolution). This representation enables the system to move beyond instantaneous measurements and instead reason over historical and contextual information, forming the foundation for predictive and proactive network control. Based on the graph-based representations, we adopt a MARL model for UE handover decision making, where each UE is modeled as an autonomous agent that makes decisions based on local observations derived from temporal graph embeddings. While MARL provides strong adaptability, it learns through interaction with the real network, where exploration may lead to unsafe or suboptimal handover actions. For instance, an RL agent may temporarily hand over a UE to a distant cell for exploration, causing service disruption. To address this challenge, we integrate rule-based constraints into the learning process. Domain knowledge (e.g., signal strength thresholds, connectivity requirements, and load constraints) is encoded as differentiable action masks that filter out invalid or unsafe actions before handover execution. This integration ensures that all handover actions satisfy practical network constraints while allowing the learning model to optimize performance within the feasible action space. We implement TARMM on an indoor multi-cell 5G ORAN testbed that comprises five cells and a number of smartphones, and evaluate its performance from both network and application perspectives. Experimental results show that TARMM outperforms existing rule-based and state-of-the-art learning-based handover methods in terms of both latency and path loss rate. From the network perspective, TARMM reduces the 95th-percentile latency by up to 44% and the path loss rate by up to 56% compared to the state of the art. From the VR application perspective, TARMM reduces the frame drop rate by up to 46% and the frame freeze ratio by up to 61% compared to the state of the art. The contributions of this work are summarized as follows: • We propose a temporal graph-based representation of multicell 5G networks that captures both spatial interactions and temporal dynamics, enabling delay-aware decision making beyond instantaneous measurements.
works. It highlights the importance of integrating learningbased decision-making with domain knowledge and systemlevel design in next-generation RANs.
• We design a unified framework that integrates reinforcement learning with rule-based action masking, achieving both adaptability and safety in handover control.
To understand the communication latency in incumbent networks, we conducted a measurement campaign in a university building equipped with 5G NR and Wi-Fi networks, as shown in Fig. 1. Our indoor multi-cell 5G NR O-RAN testbed is built on the srsRAN project [41], which is fully compliant with 3GPP and O-RAN specifications. Multi-Cell Indoor 5G O-RAN. The 5G O-RAN comprises five commercial radio units (RUs) that provide the coverage of the entire building floor, five distributed units (DUs), and one central unit (CU), Near-RT RIC, and Non-RT RIC. The
2 Handover and Delays in 5G NR 2.1 5G Handover Protocols in 3GPP In 5G NR, handover is a network-controlled procedure specified in 3GPP [1–3]. The procedure consists of three phases. • Phase-1: Measurement and decision. Each UE continuously reports its radio link conditions to its home gNB, and the network determines whether to trigger handover. In 5G NR, handover decisions are typically made based on the Event A3 criterion. Under this rule, handover is triggered if the signal power of a neighboring cell exceeds that of the serving cell by a hysteresis offset ∆A3 for a time duration of TTTT . This requirement is to filter out short-term fluctuations and reduce premature handovers. • Phase-2: Handover Preparation. If handover is triggered, the serving gNB sends a Handover Request to the target gNB via the Xn or NG interface to transfer the UE context and QoS requirements and initiate resource preparation. The target gNB then performs admission control and reserves the necessary resources for the incoming UE. If admitted, it replies with a Handover Request Acknowledge, which includes a dedicated Random Access Channel (RACH) preamble and a Radio Resource Control (RRC) container with the target cell configuration for subsequent access. • Phase-3: Handover Execution. The UE performs RRC reconfiguration to detach from the serving cell and access the target cell. When triggered under weak RSRP conditions, unreliable synchronization and repeated random access, including multiple RACH attempts and retransmissions, can significantly prolong this phase, resulting in a transient service interruption where data transmission is suspended. During this interruption, uplink and downlink traffic continue to arrive but cannot be served, accumulating at the UE and DU buffers. After the UE reconnects to the target cell, the buffered data is released in bursts, causing temporary congestion at the MAC/RLC layers.
2.2
• We implemented and evaluated the proposed system on a realistic 5G O-RAN testbed, demonstrating substantial improvements in communication latency and stability for delay-critical edge AI applications. To the best of our knowledge, this is the first work on system-level handover optimization for realistic cellular net2
Delay Measurements
GPS
O-DU 1
O-DU 2
O-DU 5 E2 5G Core
Split 7.2 Front-Haul
5G FH Switch
Midhaul Near-RT RIC
O-CU
O-RU 1
O-RU 2
O-RU 3
O-RU 4
O-RU 5 NVIDIA RTX A6000 GPU Server
1.0 0.8 0.6 0.4 0.2 0.0
Wi-Fi w/o Mobility Wi-Fi w/ Mobility 5G w/o Mobility 5G w/ Mobility
0
100
200
RTT (ms)
300
(a) CDF of measured RTT.
RTT (ms)
CDF
Figure 1: An indoor 5G NR O-RAN testbed deployed in a university building. 500 400 353 379 300 223 200 100 40 0 P99
Wi-Fi w/o Mobility Wi-Fi w/ Mobility 5G w/o Mobility 5G w/ Mobility Handover Event
101106 7
70
Mdev
(b) 99-percentile and std dev.
Figure 3: Impact of handover events on queuing delay in 5G.
Figure 2: RTT measurement results of Wi-Fi and 5G.
with and without mobility. It confirms that mobility significantly increases both metrics. The 99th-percentile RTT rises from 40 ms to 124 ms. Queueing Delay During Handover. To examine handover behavior, we carry a Samsung smartphone generating 10 Mbps persistent iperf downlink traffic and walk across two neighboring cells. The device performs automatic handover following the 3GPP Event A3 criterion. Fig. 3 shows the corresponding RSRP evolution together with the measured queueing delay. We observe that the queueing delay consistently exhibits sharp spikes aligned with handover events. Under steady connectivity, the delay remains below 20 ms, whereas during handover it can surge to over 200 ms. These delay spikes are short-lived but recurrent. Notably, the experiment is conducted under light cell load, indicating that the observed delay inflation is not due to resource saturation, but rather transient disruptions in buffering and scheduling during the handover procedure. Packet Loss During Handover. Fig. 4 illustrates the impact of handover on latency and reliability. We observe that packet loss occurs within a short window immediately after the handover event (highlighted region around packets 40-45), indicating a transient service disruption during the transition. This period is also accompanied by a pronounced RTT spike exceeding 1.5 s, in sharp contrast to the baseline RTT, which remains consistently below 50 ms before and after handover. Outside the handover interval, RTT quickly returns to a stable low-latency regime with no packet loss. These results suggest that while handover is generally efficient, it can introduce brief but severe latency spikes and packet loss due to temporary interruption and buffering effects.
commercial RUs are from different vendors including Benetel and Pegatron. The DU and CU use the srsRAN protocol stacks [41]. Near-RT and Non-RT RICs were deployed in a local NVIDIA A6000 GPU server. The system operates on the n78 band under an FCC experimental license. The center frequency is 3350.01 MHz, and the subcarrier spacing is 30 kHz. The bandwidth is 100 MHz and the MIMO configuration is 4x4. This network provides Internet access to 20+ smartphones, which are distributed within the building. The building is also equipped with a Wi-Fi network comprising dense Aruba Wi-Fi 6 routers [16]. The Wi-Fi network operates on the 5 GHz band with 40 MHz bandwidth. The routers use the same Extended Service Set Identifier (ESSID), providing building-wide coverage and enabling automatic roaming between routers. Measurement Results. We measure the round-trip time (RTT) of smartphones ping an edge GPU server under two cases: (i) the smartphones are stationary with no handover, and (ii) a smartphone moves across the entire floor. The experiments are conducted under very light traffic conditions for both 5G and Wi-Fi. Fig. 2a reports the results. Wi-Fi exhibits high latency and significant variability, with a large fraction of samples exceeding 200 ms, primarily due to channel contention and backoff. In contrast, 5G achieves substantially lower and more concentrated RTT distributions, with the majority of data packets falling within 15–40 ms even under mobility, owing to centralized scheduling. A key observation for 5G is that UE mobility significantly impacts RTT performance, as shown in Fig. 2a. Fig. 2b further compares the 99th-percentile latency and standard deviation 3
Handover
RTT (s)
1.5
Part 1: Rule-Based Module
RTT (s) Packet Loss Handover
1.0
Connectivity: Neighbor
0.0
0
10
20
30
40
50
Packet Index
60
70
80
RSRP (dBm)
Message
Memory and Updater
Actor
HO Decision
𝑖 𝑥𝑎𝑔𝑔
80
Batch
100 Serving cell Neighbour Cell
120 0
2
4
6
8
10
12
14
16
Temporal Graph Attention Layer
Embeddings
Centralized Critic
Loss
Figure 6: The architecture of TARMM.
18
Distance (m) Figure 5: Measured RSRP profile on a smartphone when it moves from one cell to another in our indoor 5G NR network.
while the rule-based masking filters out unsafe actions using domain knowledge. We describe each component in detail below.
Why Learning-Based Handover Needed?
3.1
In current 5G systems, handover decisions are predominantly based on rule-based criteria such as Event A3, which triggers a handover when a neighboring cell’s signal strength exceeds that of the serving cell by a fixed threshold. While effective in macro-cellular networks, this approach becomes inadequate in micro- and small-cell deployments, particularly in urban and indoor environments. For example, indoor settings are characterized by rich multipath propagation, frequent occlusions, and complex layouts, leading to highly irregular and rapidly fluctuating signal strength even over short distances (see Fig. 5). As a result, instantaneous signal measurements can be misleading, causing premature or unnecessary handovers, or delayed decisions that degrade link quality. This limitation is critical because handover events introduce transient service interruptions, which can lead to significant RTT spikes and packet loss even under light load conditions. Moreover, RSRP-based handover strategies overlook other important factors such as cell load, queuing delay, and application-specific latency requirements, potentially resulting in suboptimal target cell selection and degraded performance. Since each network scenario exhibits unique spatial signal patterns and traffic dynamics, fixed threshold-based policies lack the adaptability required for reliable operation. Therefore, a learning-based handover solution is needed to capture complex spatiotemporal network characteristics, leverage historical and contextual information, and dynamically optimize handover decisions based on both radio conditions and overall network state.
3
Load: Aware
Part 2: Learning-based Module
Figure 4: Impact of handover on packet latency and reliability.
2.3
Action Masking
Signal: Strength
Decision Logic
0.5
Temporal Graph-Based Representation
Temporal Graph Modeling. The RAN environment under mobility exhibits both strong temporal and spatial dependencies. From temporal perspective, user mobility and network conditions evolve continuously over time. The future state of a UE, including its serving cell, channel quality, and experienced delay, is highly correlated with its historical trajectory. Therefore, relying solely on instantaneous observations leads to myopic decisions that fail to anticipate upcoming handovers or delay variations. Capturing such temporal dependencies requires maintaining historical information of UE mobility and network dynamics. From a spatial perspective, the performance of each UE is inherently influenced by its surrounding environment, including neighboring cells and nearby UEs. In particular, neighboring cells determine candidate handover targets and their load conditions, while neighboring UEs compete for shared radio resources, affecting scheduling and delay performance. As a result, effective decision making requires modeling interactions among nodes in the network. To jointly capture temporal evolution and spatial interactions, we model the RAN as a time-evolving graph, where nodes represent individual UEs and cells, and edges represent serving and neighboring links, as illustrated in Fig. 7. Both node states and edge relationships are continuously updated to reflect mobility dynamics and inter-cell dependencies. Specifically, we adopt a Temporal Graph Network (TGN) to learn dynamic node representations by integrating historical information with neighborhood context. This formulation enables the model to capture both the temporal progression of UE states and the influence of neighboring cells and UEs, providing a unified framework for modeling spatiotemporal dependencies in the network. Message Passing and Feature Learning. We model the RAN system as a continuous-time dynamic graph represented by a sequence of time-stamped events, as shown in Fig. 8.
TARMM: Design
Fig. 6 presents the high-level architecture of TARMM, which consists of two key components: (i) temporal graph embedding and (ii) a graph-based actor–critic MARL framework with rule-based action masking. The learning-based module adapts handover decisions to dynamic network conditions, 4
UE-Cell Interaction
Memory 𝒉𝑖 (𝑡 − )
𝒉𝑖 (𝑡)
GRU
Q
K=V
𝑈𝐸 𝑖
𝑁𝑒𝑖𝑔ℎ𝑏𝑜𝑟 𝑗
Multi-Head Attention MLP
𝒙𝑖𝑗 (𝑡) Service Edge Neighbor Edge Node
gNB UE
Graphical Representation
Message
t0
t1
t2
t3
t4
Aggregation
Embedding
Figure 8: Temporal graph modeling and message passing in TARMM.
Temporal GNN
Figure 7: Graph representation of an RAN.
Table 1: KPM data used for message representation in graph. Source
Let G (t) = (N , E (t)) denote the graph, where N is the set of nodes corresponding to individual UE and base station (BS), and E (t) is the set of edges corresponding to each UE and its serving and neighboring BS. Each node maintains a time-evolving memory state hi (t), encoding its historical observations. The state of each node updates through message passing, triggered by the periodically KPM reportings from O-RAN’s E2 interface. For the graph, we define two types of events for message passing: (i) edge event and (ii) node event. • An edge event captures the interaction between a UE node i and a BS node j, defined as xi j (t) = (i, j,t, ei j (t)), where ei j (t) is constructed from the KPM in Table 1. This event triggers message generation for both UE node i and BS node j. The generated messages are: h i mi (t) = hi (t − ), h j (t − ), ∆ti , xi j (t) , (1) h i m j (t) = h j (t − ), hi (t − ), ∆tv , xi j (t) , (2)
Feature
Serving RSRP / RSRQ / SINR UE-BS Edge Neighbor RSRP/ RSRQ/ SINR RAN / AMF / UE / DU IDs
UL/DL Description DL DL –
Measurements from serving cell. Measurements from neighboring cells. Cross-layer identifiers.
UE-Node
Throughput Queuing Delay Air Interface Delay Packet Drop Rate PRB Usage Traffic Volume CQI / RSRP / SINR UE Identifier
DL/UL Per-UE achieved data rate. DL/UL Packet delay at the RLC layer. UL Transmission delay over the air. DL Fraction of dropped packets. DL/UL Number of PRBs allocated to the UE. DL/UL Transmitted data volume. UL Uplink radio quality indicators. – UE identifiers for cross-layer.
Cell-Node
Cell Throughput PRB Usage PRB Utilization
DL/UL Aggregated throughput at the cell level. DL/UL Total PRBs used in the cell. DL/UL Percentage of PRB utilization.
Attention-Based Embedding. To mitigate memory staleness and incorporate neighborhood dynamics, we compute node i’s embedding zi (t) using a temporal multi-head attention mechanism over node i’s recent temporal neighborhood: zi (t) = MLP hi (t) ∥ h̃i (t) , (5) where ∥ is the operator of vector concatenation and h̃i (t) = MultiHeadAttn qi (t), Ki (t), Vi (t) . In this attention, the query is defined as the current node memory, i.e., qi (t) = hi (t), and the keys and values are constructed from its temporal neighbors: h i Ki (t) = Vi (t) = h j (t − , (6) j ) ∥ xi j (t j ) ∥ φ(t − t j )
where hi (t − ) denotes the memory state of node i immediately before time t, and ∆ti is the elapsed time since the last event of node i. • A node event is designed for node-specific state updates, defined as xi (t) = (i,t, ni (t)), where ni (t) is constructed based on the KPM in Table 1 that are associated with this UE or BS node. The generated message can be written as follows: h i mi (t) = hi (t − ), ∆ti , xi (t) . (3)
j∈Ni (t)
where Ni (t) denotes the set of neighbors of node i, and t j is the timestamp of the most recent interaction between nodes i and j prior to time t. h j (t − j ) is the memory of node j immediately before time t j . φ(·) is a time encoding function capturing recency effects. This attention mechanism enables the model to selectively focus on neighboring cells whose recent interactions are most relevant to the UE’s current radio and QoE conditions.
For a UE or BS node in the graph, it may receive multiple messages within a training batch. In this case, messages are aggregated using a most-recent aggregator, which selects the message with the latest timestamp to form the aggregated message m̄i (t). When a node receives a message, its memory is updated via a gated recurrent unit (GRU): hi (t) = GRU m̄i (t), hi (t − ) . (4)
3.2
Graph-Based Actor-Critic MARL
With the node embedding zi (t), we formulate the handover problem as a multi-agent decision-making process, where each UE i ∈ U (t) is treated as an agent. Each agent makes handover decisions based on its local observation, while sharing a common policy across all agents. The observation is defined as oi (t) = zi (t). The policy is parameterized as
The GRU enables each UE and cell to capture long-term temporal dependencies, such as persistent congestion or gradual signal degradation, which are difficult to infer from instantaneous KPM. 5
UE Embedding 𝑜𝑖 (𝑡) Actor MLP
Centralized Critic
and feeds it into an MLP to produce a policy πi (ai | oi ). A rule-based action masking module is applied before sampling to filter out infeasible actions, and the policy is renormalized over the valid action space. The final action is multi-discrete, consisting of handover triggering and target cell selection. The actor outputs a multi-discrete action with two categorical heads for handover triggering and target cell selection. The centralized critic estimates the state value using both local and global context. The centralized critic leverages both local and global context to estimate the state value. Specifically, it takes as input the concatenation of the UE embedding oi (t) and a graph embedding g(t), where g(t) is obtained via mean pooling over all node embeddings, and outputs the value function Vt (oi (t), g(t)).
Graph & UE Embedding
Critic MLP
Rule-Based Action Masking 𝑉𝑡 (𝑜𝑖 (𝑡), 𝑔(𝑡)) Valid Actions 𝜋𝑖 (𝑎𝑖 |𝑜𝑖 ) Agent11 Agent 2 Agent 3 Agent 4 Agent 5 Agent
Handover
Target Cell
RAN
Figure 9: Diagram of proposed actor-critic MARL framework. πθ ai (t) | oi (t) , where ai (t) denotes the action of agent i at time t, and θ represents the shared policy parameters. Action and Observation Space. At each time step t, each UE agent i ∈ U (t) observes a local observation oi (t) derived from the temporal graph embedding gi (t). This observation has already encoded: (i) UE channel quality, (ii) neighboring cell conditions, and (iii) temporal dynamics reflected through past interactions. Each UE agent selects a handover action defined as target ai (t) = pHO (t) , (7) i (t), ci
3.3
Rule-Based Action Masking
In MARL for handover decision-making in O-RAN, the learned policy is inherently stochastic, requiring a balance between exploration and exploitation. While exploration is essential for discovering optimal strategies, it can also lead to unsafe or impractical actions during training and deployment. For instance, a UE may be driven to connect to a far-away RU despite having a strong nearby serving cell, or be redirected from an underutilized RU to a heavily congested one purely for exploration. Such actions can cause severe performance degradation, including increased latency and service interruption. Therefore, unconstrained exploration is undesirable in this practical networks. To ensure safe and stable handover decisions, we incorporate rule-based action masking into the MARL policy. Specifically, we define a binary mask Mi (t) ∈ {0, 1}|Ai | over the action space Ai , where each entry indicates whether a candidate action is valid. The masked policy is given by:
target
where pHO (t) speci (t) ∈ {0, 1} is handover decision and ci ifies the target cell index. Learning Objective. A handover event triggers networkside RRC reconfiguration for the UE, introducing a transient service interruption that manifests as elevated queuing delay at the O-RAN DU. For each UE agent i ∈ U (t), let Di denote its delay requirement, and let τi (t) denote the experienced packet delay at time t. We define τi (t) as the sum of uplink DL and downlink queuing delays, i.e., τi (t) = τUL i (t) + τi (t). If τi (t) ≤ Di , the delay requirement is satisfied and no penalty is incurred. Otherwise, we quantify the violation as a normalized delay deficit. The delay regret of UE i at time t is defined as: τi (t) − Di Ji (t) = max ,0 . (8) Di
π̃θ (ai (t) | oi (t)) =
πθ (ai (t) | oi (t)) · Mi (ai (t),t) , ∑a′ ∈Ai πθ (a′ | oi (t)) · Mi (a′ ,t)
(9)
which renormalizes the probability distribution over valid actions only. Domain Knowledge–Based Masking. To ensure that the learned policy produces valid and deployable handover decisions, we introduce three action-masking constraints derived from domain knowledge of 5G network operations. • Connectivity masking. A target cell must be observable in the UE’s measurement reports. Let Ni ′ (t) denote the set of neighboring cells reported by UE i. Then, the connectivity mask is defined as: (1) Mi, j (t) = I j ∈ Ni ′ (t) , (10)
This regret captures the end-to-end latency of a UE. In particular, during handover, packets are temporarily buffered due to link interruption, scheduling delay, and resource reallocation, leading to queue buildup at the RLC/MAC layers. The queuing delay reflects both the instantaneous channel condition and the stability of the connection. Therefore, minimizing Ji (t) enables the agent to take the actions that can reduce handover-induced latency spikes and maintain delaycritical service requirements. We define the reward of agent i as the negative of its delay regret: Ri (t) = −Ji (t). Actor-Critic Network. We adopt a parameter-sharing multi-agent actor–critic architecture with decentralized actors and a centralized critic, as illustrated in Fig. 9. Each UE acts as an agent and makes decisions based on its local observation oi (t). The actor takes the UE embedding oi (t) as input
where I(·) return 1 if condition is true and 0 otherwise. • Signal strength masking. Although the MARL policy incorporates the target cell’s signal strength alongside other KPMs, it may be underweighted relative to other KPM inputs, potentially leading to handover decisions toward weak 6
(2)
(3)
0
0
1
0
0
0
0
0
0
0
1
0
0
0.8 0.6
Cell 1 Cell 2 Cell 3
Masked Cell
Cell 4
Handover
t1
t2
t3
t4
t5
t6
t7
t8
t9
0.4 0.2
Cell 5
3
t10 t11 t12 t13 t14 t15 t16
Cell 2
CU
0.0
2
UE
RB
DL
1
UE_INFO
Allocate for the incoming UE
Cell 3’s DU Frequency RB
Handover RIC Decision
Cell 3
UL
Time Slots
Handover
Figure 11: Illustration of proactive resource reservation at the target cell. In particular, LTGN in Eq. (15) is adopted to stabilize representation graph learning. It is defined as a self-supervised link prediction loss:
(13)
LTGN = −
h i 1 yi j log σ(ŷi j ) + (1 − yi j ) log 1 − σ(ŷi j ) , ∑ N (i, j)∈D (16)
where D is the set of sampled UE-RU pairs and N = |D |. yi j ∈ {0, 1} indicates whether the UE-cell interaction is observed, ŷi j = z⊤ i z j is the predicted link score computed from node embeddings and σ(·) is the sigmoid function. To validate the quality of learned representations, we further analyze the temporal dynamics and consistency of TGN embeddings in Appendix B, where we show that the embeddings exhibit smooth temporal evolution and strong locality in representation space.
∇θ log π̃θ = ∇θ log πθ − ∇θ log ∑ πθ (a′ | oi (t)) · Mi (a′ ,t), a′ ∈Ai
(14) which corresponds to a reweighted version of the original policy gradient over the valid action space.
3.5
Online Training Process
Proactive Resource Preparation
Fig. 10 illustrates the sequential handover decisions of the RL agent over time. The top row shows the binary handover decisions, while the heatmap below visualizes action preferences over neighboring cells. Hatched entries indicate masked cells removed from the feasible action space by rule-based constraints, restricting the agent to valid candidates only. The results show that the policy consistently concentrates on a specific feasible target cell with high preference, indicating that the RL agent learns to jointly decide when and where to hand over while respecting network constraints. Proactive Resource Preparation. While optimized handover decisions can reduce connectivity disruption, they do not address transient delay spikes immediately after handover. During this phase, the target DU treats the UE as a new arrival, resulting in cold-start scheduling, temporary queue buildup, and increased latency.
TARMM includes three networks: graph encoder network parameterized by ψ, MARL actor network parameterized by θ, and MARL critic network parameterized by φ. The actorcritic networks are updated using standard PPO loss function, and the graph encoder network is updated by both actor loss and self-supervised link prediction loss. Specifically, they are updated as follows: (θ, ψ) ← (θ, ψ) − απ ∇θ,ψ Lactor , (φ, ψ) ← (φ, ψ) − αv ∇φ,ψ Lcritic ,
0
1
which ensures that only feasible, high-quality, and noncongested target cells are considered during action selection. Model Differentiability. Differentiability is critical for efficient learning. Since the mask Mi (a,t) is deterministic and independent of the policy parameters θ, gradients propagate only through the original policy πθ and do not alter the underlying learning dynamics. Taking the gradient of the masked log-policy yields:
3.4
0
Figure 10: Illustration of handover decision-making process under action masking.
which filters out candidate cells whose load exceeds the allowable threshold. Combining these three cases, the overall mask applied to the MARL is: (1)
0
Time Step
• Load-aware masking. In practice, a UE should not be handed over to a highly congested cell. Thus, we restrict the action space based on cell load. We define the load of cell j as ρ j (t) = B1j ∑i∈U j (t) bi, j (t), where U j (t) is the set of UEs associated with cell j. Let η j denote the load threshold for cell j. The load-aware mask is then defined as: (3) Mi, j (t) = I ρ j (t) ≤ η j , (12)
Mi, j (t) = Mi, j (t) · Mi, j (t) · Mi, j (t),
HO
Attention Weight
Neighboring Cells
cells. Insights from our experiments indicate that a candidate cell must provide sufficient signal strength to ensure j reliable handover execution. Let Pi (t) denote the RSRP of UE i from candidate cell j, and let γ denote the minimum RSRP threshold for handover. Then, the signal-based mask is defined as: (2) j Mi, j (t) = I Pi (t) ≥ γ . (11)
(15)
ψ ← ψ − αTGN ∇ψ LTGN , where απ , αv , and αTGN are the learning rates of actor, critic, and graph updates. Lactor and Lcritic are actor-critic PPO losses (see Appendix A). 7
4
Delay (ms) Used RBs
HO: DU2 DU5
DL UL
100 0 60 40 20
10
20
30
40
Time (s) (a) w/ resource reservation.
50
HO: DU2 DU5
DL UL
100
DL UL 0
200
Delay (ms) Used RBs
200
To address this issue, we introduce a simple yet effective proactive resource reservation mechanism at the target DU. As shown in Fig. 11, once the RIC makes a handover decision, it sends the UE’s state information (e.g., recent resource usage and QoS requirements) to the target cell before the handover is executed. The target DU then reserves resources in advance, targ with the reserved amount defined as bi = κ · bserv i , where targ serv bi and bi denote the PRBs at the target and serving cells, respectively, and κ ∈ (0, 1] is a scaling factor (set to 1 when the target cell is not overloaded). The UE is also assigned the highest scheduling priority upon arrival to ensure immediate service. This lightweight mechanism significantly reduces posthandover queuing and scheduling delay. By enabling preallocation and informed scheduling, the UE can be served promptly after attachment, mitigating cold-start effects and stabilizing latency. More importantly, the approach is fully compatible with 3GPP and O-RAN, requiring no changes to standard handover procedures or scheduler design. Effectiveness of Proactive Handover. We evaluate this approach using a mobile UE with persistent 10 Mbps uplink and 30 Mbps downlink traffic. Fig. 12 compares performance with and without proactive resource reservation. With proactive preparation, the UE experiences a smooth handover transition, with gradual resource adaptation and consistently low delay. In contrast, without reservation, the target cell reacts only after handover completion, leading to a sudden surge in uplink demand and a sharp delay spike (up to 60 ms) due to transient buffering and scheduling contention.
0 60 40 20
DL UL 0
10
20
30
40
Time (s) (b) w/o resource reservation.
50
Figure 12: Comparison of network resource utilization and UE’s delay between two cases during handover events. ing node embeddings from 16 to 32 dimensions. Relative time encodings are used as edge attributes, and node-type embeddings distinguish between UE and DU nodes. The encoder outputs 32-dimensional UE embeddings zi (t) and a graph-level embedding g(t) via mean pooling. On top of the graph encoder, we adopt a parameter-sharing actor-critic architecture. The actor takes zi (t) as input and feeds it into an MLP with hidden size (64, 64) (ReLU + LayerNorm) to generate multi-discrete actions. The critic takes [zi (t), g(t)] as input and uses an MLP (64, 64) with a linear head to estimate Vφ zi (t), g(t) . Source Code and Demo. GitHub link to source code and demo videos is provided to the program chairs as instructed.
5
Experimental Evaluation
We would like to answer the following questions through the evaluation. • Q1: How does TARMM impact RTT and packet loss for smartphones under varying network loads? (§5.2) • Q2: How does TARMM compare with state-of-the-art mobility management solutions? (§5.3) • Q3: What is the impact of key design components in TARMM, including Temporal Graph modeling and rulebased action masking? (§5.4) • Q4: How does TARMM scale with increasing numbers of users and cells? (§5.5) • Q5: How does TARMM improve user experience for VR applications under diverse AI workloads? (§5.6)
Implementation
To emulate realistic network conditions, a set of smartphones remains stationary to generate background traffic and maintain cell load, while another set moves along predefined routes that traverse multiple cells to trigger handovers. These routes cover the entire network area. The measurement data is collected while continuously traversing the central regions of cells, focusing on mobility periods where handovers occur, with handover events observed approximately every 5-10 seconds. Near-RT RIC and KPM Details. All learning-based components are implemented in the O-RAN Software Community (OSC) Near-RT RIC [32], The RIC connects to one CU and five DUs via the E2 interface. UEs report their radio measurements (e.g., signal quality indicators) to the network every 120 ms, while fine-grained network statistics are collected from the CU and DUs at a 10 ms interval through E2 subscriptions. Graph Embedding and MARL Parameters. We use a temporal graph to model dynamic UE-RU interactions. Each node maintains a 16-dimensional memory, updated via incoming messages and an aggregation function. For each event, we perform one round of temporal message passing followed by a TransformerConv layer (2 heads, dropout 0.1), project-
5.1
Performance Metrics and Baselines
In our experiments, we use the following metrics to quantify the performance of TARMM: uplink queuing delay, downlink queuing delay, packet loss rate, and the defined reward function. We also compare TARMM against the following SOTA learning-based baselines. • 3GPP Standard [4]. The standard Event A3-based handover mechanism defined in 3GPP specifications, where a handover is triggered when the RSRP of a neighboring cell exceeds that of the serving cell by a predefined offset for a time-to-trigger duration. • MR-PHO [15]. MR-PHO stands for measurement-reportbased predictive handover. It is a learning based handover 8
Packet Loss (%)
approach, leveraging historical measurement reports to predict future signal conditions for proactively triggering handovers. • Adaptive Parameter Tuning (APT) [7]. It employs a deep reinforcement learning for mobility load management, where each cell dynamically adjusts handover bias parameters to influence UE association decisions.
1.00
CDF
0.75 Static 3GPP No-MM TARMM
0.50 0.25
Better
0.00 101
102
1.55%
1
0
0.00%
0.08%
0.02%
Static
3GPP
No-MM TARMM
RTT (ms)
(a) All cells have light traffic loads.
Main Results
Packet Loss (%)
1.00 0.75
CDF
5.2
2
With TARMM deployed in our live 5G O-RAN testbed, we first conduct a comprehensive evaluation from a UE’s perspective. Our data collection campaigm involves four Samsung smartphones, along with many other smartphones to generate background traffic for the control of cell traffic load. Of those four smartphones, one is stationary, placed close to an RU. This smartphone does not have mobility. It is used as a reference serving as the upper bound for mobility management. The other three smartphones are carried by the same person, walking around the entire network coverage area. The mobility of these three smartphones are managed by three different handover control algorithms: one uses in-house 3GPP handover algorithm; one uses TARMM; and one does not have explicit mobility management (No-MM). In the case of No-MM, handover is disabled at the CU for this smartphone, forcing it to reconnect only after link failure when crossing cell boundaries. During the experiments, the RTT and packet loss rate of those four smartphones are continuously recorded for data analysis. Apparently, the performance of a UE depends on the traffic load of the network. Therefore, we define two network traffic load cases using other smartphones: (i) light traffic load (10% to 20% of cell capacity), and (ii) heavy traffic load (about 80% of cell capacity). Apparently, the performance is also dependent on the traffic type, packet size of the UE under test. Therefore, we design two experiments: (i) ping small-sized bursty data traffic and (ii) iperf persistent data traffic. Ping Data Traffic on UE. The ping command on smartphones generates very small data packets, representing the highly bursty traffic from the end users. In total, we collected 5 hours of ping data traces from those four smartphones over 3 days. Fig. 13 reports our experimental results. We have the following observations. First, static UE has the best performance in terms of both RTT and PER. When the cell has light traffic load, most of its RTTs remain with 20 ms, and its PLR stay on zero. This is as expected, thanks to its strong signal and the network’s centralized scheduling mechanism. Second, the mobility significantly worsens the RTT and PLR performance of UEs, regardless of the handover management algorithms. This can be attributed to the dynamic link quality and the imperfection of handover timing. Third, cell load has a significant impact on the performance of UE. Compared Fig. 13b against Fig. 13a, we see that both RTT and PLR increase dramatically when the cell load becomes heavy.
0.50 0.25 0.00 101
Better 10
2
Static 3GPP No-MM TARMM
10
3
7.5
6.41%
5.0
3.60%
2.5 0.0
0.98%
0.16%
Static
3GPP
No-MM TARMM
RTT (ms)
(b) All cells have heavy traffic loads.
Figure 13: Comprehensive measurement results of RTT and packet loss rate when UEs ping an edge GPU server. Table 2: Measured RTT (ms) of a smartphone under different handover policies. ‘–’ indicates that the RTT exceeds the timeout threshold. M: Method; P: Percentile; L: Light load; H: Heavy load. M 3GPP [4] No-MM APT [7] MR-PHO [15] TARMM (ours) P 95th 90th 95th 90th 95th 90th 95th 90th 95th 90th L 139 57 353 137 121 68 97 54 57 38 Ping H 607 283 – 718 352 201 264 156 185 118 L 1132 400 601 396 356 248 379 303 262 170 Iperf H – 482 – 451 638 376 470 282 318 182
Fourth and most importantly, TARMM outperforms its peer handover strategies in both cases. The gain is more significant in a network with heavy traffic load, where TARMM has a performance close to the static UE. This indicates that the main contribution of RTT and PLR is the resource competition rather than the handover events. Iperf Data Traffic on UE. The iperf command is used on the UEs to generate persistent data traffic with large data packets. Each of the four UEs generates 1–3 Mbps persistent uplink traffic and 10–30 Mbps persistent downlink traffic, emulating practical edge AI offloading applications. For this test, we collected 5 hours of data traces in the network over 3 days. Fig. 14 presents our experimental results. We have the following observations. First, similar to the previous ping case, TARMM demonstrates a significant gain compared to 3GPP and No-MM handover control. The gain stems from the fact that TARMM considers the global network state encoded by the graph network, rather than using local link information, to make the handover decision for UEs. Second, iperf traffic experiences much larger RTT and higher PLR compared to the ping traffic. This is expected, because iperf generates more data packets and thus intensifies the resource competition. Third, advanced handover management is effective mainly in the networks with low traffic load and becomes ineffective in heavy-loaded network, where resource is the main determining factor other than handover timing. 9
CDF
Static 3GPP No-MM TARMM
0.50 0.25 0.00 101
10
2
1.00 4
3.33%
2
0.25
No-MM TARMM
0.00
0.00%
0
Static
3GPP
0.50
1.03%
RTT (ms)
101
0.25 0.00 101
10
2
10
3
10.75%
10
1.87%
0.00%
Static
3GPP
No-MM TARMM
RTT (ms)
(b) Cells have heavy traffic loads.
Figure 14: Comprehensive measurement results of RTT and packet loss rate when UEs iperf an edge GPU server. 1.00
0.75
CDF
CDF
0.00
101
102
RTT (ms)
103
(b) TARMM’s action masking.
300
RL with TGN-sum
RL with Current Snapshot
250 RL with GNN
200 RL with TGN-attn
150
20
30
40
50
Better
0.75 0.50 APT MR-PHO TARMM
0.25 0.00 101
w/ action masking w/o action masking
0.25
Inference Time (ms)
1.00
Better
103
0.50
Figure 16: Experimental results of TARMM’s ablation studies.
7.20%
5 0
102
RTT (ms)
(a) TARMM’s TGN.
15
95th Percentile RTT (ms)
CDF
Static 3GPP No-MM TARMM
0.50
Packet Loss (%)
Better
0.75
Better
0.75 TGN GNN Current Snapshot
(a) Cells have light traffic loads. 1.00
1.00
Better
0.75
4.06%
CDF
Better
0.75
CDF
Packet Loss (%)
1.00
102
RTT (ms)
Figure 17: Performance of TARMM with different KPM embedding methods.
0.50 APT MR-PHO TARMM
0.25 0.00 101
102
5.4
103
RTT (ms)
Ablation Studies
We conduct ablation study to evaluate the effectiveness of individual components of TARMM. TGN Embedding Module. TGN is a key component of TARMM, with alternatives including a static GNN and a Current Snapshot approach. While GNN captures spatial relationships through graph aggregation, it ignores temporal dynamics. In contrast, the Current Snapshot approach relies solely on instantaneous local observations for KPM embedding, without incorporating either temporal history or rich spatial context. Fig. 16a compares the performance of TARMM when TGN is replaced by these alternatives. TGN significantly shifts the RTT CDF to the left, reducing the median RTT from approximately 40–50 ms (GNN) and 60–80 ms (Current Snapshot) to around 20–30 ms. At the tail, TGN keeps the 90th-percentile RTT below 100 ms, whereas GNN and Current Snapshot exceed 200 ms and 400 ms, respectively. Fig. 17 evaluates TARMM with different embedding modules in terms of inference time and RTT. Compared to the 3GPP baseline, all RL-based approaches significantly reduce tail latency, though their effectiveness depends on the quality of state representation. RL with a Current Snapshot performs the worst due to the lack of temporal and structural context, while incorporating spatial information via GNN improves performance. TGN-based models further reduce RTT by capturing temporal dynamics. In particular, RL with TGN-attn achieves the best tradeoff, lowering the 95th-percentile RTT to around 200 ms with minimal inference overhead, highlighting the importance of jointly optimizing model efficiency and network-aware decision-making. Rule-Based Action Masking. We compare the performance of TARMM with and without the rule-based action
(a) Cells have light traffic load. (b) Cells have heavy traffic load.
Figure 15: Comparison of TARMM against SOTA learningbased handover methods.
5.3 Comparison with SOTA Learning Methods While the previous subsection compared TARMM against the existing rule-based handover policies and demonstrated its gain, this subsection focuses on the comparison between TARMM and the SOTA learning-based handover methods: APT [7] and MR-PHO [15]. Similar to the previous experiments, we implement APT and MR-PHO in our live network for two smartphones, and the same person carries these three UEs walking around the entire network area. The three smartphones generate 3 Mbps uplink and 30 Mbps downlink persistent traffic using iperf command. In total, 5 hours of data traces have been collected for this comparison. Fig. 15 reports our experimental results in both light- and heavy-traffic networks. TARMM consistently outperforms MR-PHO and APT by shifting the RTT distribution left, reducing median latency by about 30–40% and keeping tail latency (e.g., 90th percentile) significantly lower. It also improves reliability, achieving a lower packet loss rate. We attribute the gain of TARMM to the temporal graph embedding, making it possible to make handover decision for individual UEs based on global network information rather than local individual link quality. Table 2 summarizes the comparison results, demonstrating that TARMM outperforms the baselines in mobile scenarios. 10
Object Detection
Segmentation
Pose Estimation
Boundary
100 0
1
4
7 UE number
9
1.0 0.8
50
0.6 0.4
Object Detection Pose Estimation Segmentation
0.2
Object Detection
Pose Segmentation Estimation
(a) Latency Breakdown.
0.0
0
50
100
150
Time (ms)
200
250
(b) CDF of breakdown latency.
Figure 20: Average and CDF of delay breakdown components under different VR offloading tasks. [Yellow: upload time; blue: inference time; orange: download time.]
masking module. Fig. 16b shows that masking significantly improves performance, reducing the median RTT from around 80–100 ms to 20–30 ms and lowering the 95th-percentile RTT from over 500 ms to below 100 ms. It also eliminates extreme latency events above 1 s. These gains stem from constraining unsafe exploration in MARL. Without masking, the agent tends to make aggressive handover decisions for exploration, often selecting transiently strong but unstable links or overloaded cells. Such actions lead to severe performance degradation, including large RTT spikes and increased packet loss.
End-to-End Delay Profile. The three applications exhibit distinct data transmission characteristics. Object detection and pose estimation generate approximately 22 KB uplink and 2 KB downlink per frame, while Semantic Segmentation produces significantly larger traffic (110 KB uplink and 6 KB downlink). Fig. 20 shows the breakdown of end-to-end delay across workloads. As illustrated in Fig. 20a, object detection achieves the lowest overall latency, pose estimation is dominated by inference delay, and semantic segmentation incurs the highest upload latency due to larger input size. This highlights heterogeneous bottlenecks across workloads. Fig. 20b further presents the latency distributions: object detection exhibits consistently low latency with a tight distribution, pose estimation shows moderate tail latency, while semantic segmentation has significantly heavier tails, with the 95th-percentile exceeding 100 ms. These results indicate that compute- and bandwidth-intensive workloads are more sensitive to network dynamics and resource allocation. Comparison with SOTA Methods. We compare TARMM with rule-based and learning-based handover methods in terms of end-to-end delay, including upload, inference, and download time consumptions. Fig. 21 shows that TARMM consistently outperforms all mobility management baselines across the three applications, with larger gains for latencysensitive workloads. Specifically, TARMM reduces median latency by 30–40% and 90th-percentile latency by about 35% compared to the 3GPP baseline, with the most significant improvement observed for segmentation. Compared to learningbased baselines (APT and MR-PHO), TARMM achieves an additional 15–25% reduction in both median and tail latency. These results demonstrate that TARMM effectively improves both typical and worst-case delay by enabling timely and loadaware handover decisions across diverse edge AI workloads. Quantitative User Experience. Since VR user experience
Impacts of Network Size
To evaluate the scalability of TARMM, we vary the number of active RUs in our 5G O-RAN testbed from 3 to 5, and adjust the number of active UEs in the target cell from 1 to 12. Fig. 19 shows the 95th-percentile RTT measured at the UEs, capturing latency during handover events. The RTT remains relatively stable as the network scales, increasing only modestly from about 180 ms to 220 ms. This result demonstrates that TARMM is robust to both user density and network size, maintaining stable latency performance under increasing system complexity.
5.6
Upload AI Inference Download
100 0
12
Figure 19: The 95th-percentile RTT when the network has different cells and different number of UEs.
5.5
150
CDF
200
3 Cells 4 Cells 5 Cells
200
Time (ms)
95th Percentile RTT (ms)
Figure 18: Illustration of three VR perception offloading tasks for edge AI computing.
VR Offloading Applications
We now extend the evaluation of TARMM from a network perspective to application-level performance. We consider three YOLO-based VR perception tasks on Meta Quest 3: Object Detection, Pose Estimation, and Semantic Segmentation, as shown in Fig. 18. For each task, the VR headset captures images, compresses them, and offloads them to an edge GPU server over the 5G network. The GPU server performs inference and returns the results for real-time rendering. The image frame rate is set to 20 FPS. Since Quest 3 does not support 5G connectivity, a smartphone hotspot is used to bridge the connection to the 5G network. 11
1.0
0.8
0.8
0.6
Static
0.4
3GPP
0.2
MR-PHO
APT
0.6
Static
0.4
3GPP
0.2
MR-PHO
APT
TARMM
0.0
0
250
500
CDF
1.0
0.8
CDF
CDF
1.0
0.6
Static 3GPP
0.4
APT MR-PHO
0.2
TARMM
0.0
750
0
Latency (ms)
200
400
600
TARMM
0.0
Latency (ms)
(a) Object detection. (b) Pose estimation.
that jointly consider radio conditions, traffic dynamics, and queueing behavior to minimize latency [39, 40, 43, 51]. Handover Management. Handover management has been extensively studied with various objectives, including maximizing throughput [27], improving load balancing [11, 33], enhancing fairness [35], and reducing signaling overhead [38, 43, 45]. A wide range of techniques have been proposed, such as rule-based heuristics [44], optimization-based formulations [28], and learning-based approaches such as reinforcement learning [7, 17] and predictive models [22, 49]. These methods aim to improve network efficiency by adapting handover thresholds, predicting mobility patterns, or optimizing cell selection policies. However, most prior work relies on simplified mathematical models or simulated datasets that fail to capture the complexity and dynamics of real-world networks. Moreover, most studies were evaluated primarily through numerical analysis, simulation, or network emulation, which often do not reflect the operational constraints and variability of deployed systems. TARMM advances this research line by developing mobility management from an O-RAN system perspective and evaluating it on a realistic O-RAN testbed. Learning-based Network Control. Learning-based approaches have been widely adopted as a general framework for O-RAN testing [42], network control [18, 24], improving QoE [23, 53, 55],enabling adaptive decision-making [9] in complex and dynamic environments. A wide range of techniques have been proposed, including supervised learning for traffic and performance prediction [34], optimization-guided learning frameworks [25], and reinforcement learning for sequential decision-making [54]. More recently, graph-based models have been introduced to capture spatial dependencies among network entities [52], while multi-agent learning frameworks enable decentralized control across distributed network components [26]. These approaches aim to learn adaptive control policies directly from data and interactions with the environment. TARMM differs from them in both objective and approach.
250
500
750
1000
Latency (ms)
(c) Segmentation.
Figure 21: End-to-end delay comparison of different mobility management schemes for different edge AI tasks. TARMM
3 2 1 0 ObjDet PoseEst Segment
MP-PHO
Freeze Ratio (%)
Drop Rate (%)
Static
APT
3GPP
7.5 5.0 2.5 0.0 ObjDet PoseEst Segment
Figure 22: VR’s user experience comparison of different mobility management schemes. is inherently subjective, we quantify it using two metrics measured on VR headsets. (i) Frame Drop Rate, which captures VR headset’s rendering continuity under network dynamics, and (ii) Frame Freeze Ratio, which measures the fraction of time the display stalls due to delayed or missing frames, directly reflecting perceived smoothness. These two metrics are directly reflective of user experience [46–48]. Fig. 22 presents the measurement results for three representative VR offloading tasks under different handover schemes. TARMM consistently achieves lower frame drop rates and smaller frame freeze ratios than all baselines, with the largest gains observed in throughput-demanding workloads such as PoseEst and Segment, where mobility-induced latency variations are more disruptive. In contrast, the rule-based 3GPP baseline exhibits the worst performance due to reactive, signal-only decisions that incur higher packet loss and queuing delay. Learning-based methods (MP-PHO and APT) offer moderate improvements over 3GPP but still lag behind TARMM, as they do not effectively capture temporal dynamics or network conditions. These results demonstrate that reducing handover-induced disruptions directly translates into improved application-level QoE, as well as the effectiveness of TARMM from an application perspective.
6
7
Conclusion
In this paper, we presented TARMM, a 5G O-RAN system that optimizes user mobility management for delay-critical edge AI offloading. By leveraging temporal graph embedding, TARMM captures the spatiotemporal dynamics of user mobility and network conditions, enabling more informed and proactive handover decisions. To ensure safe and deployable operation, we incorporate rule-based action masking, which constrains MARL exploration using domain knowledge and prevents harmful or infeasible actions. We implement TARMM on a real-world testbed and demonstrate its effectiveness under both network-level and application-level scenarios. Overall, this work highlights the importance of integrating learning-based decision-making with domain knowledge and system-level design in future RANs.
Related Work
Latency Optimization in 5G. Reducing end-to-end latency has been a central objective in 5G networks, particularly for Ultra-Reliable Low-Latency Communications (URLLC) and delay-sensitive applications. Prior work has explored latency optimization from multiple perspectives, including physical-layer enhancements [37], scheduling and resource allocation [20, 50], and transport/network-layer optimizations such as edge deployment and protocol tuning [5, 6, 21]. In addition, several studies have proposed cross-layer designs 12
References
[11] Quan Chen, Xiaoqin Song, Tiecheng Song, and Yang Yang. Vehicular edge computing networks optimization via drl-based communication resource allocation and load balancing. IEEE Transactions on Mobile Computing, 2025.
[1] 3GPP. Nr; nr and ng-ran overall description; stage-2. Technical Report TS 38.300, 3GPP, 2020. [2] 3GPP. Nr; radio resource control (rrc); protocol specification. Technical Report TS 38.331, 3GPP, 2022.
[12] Christopher Downie. Latency standards in trading systems. https://www.luxalgo.com/blog/ latency-standards-in-trading-systems/, 2025. LuxAlgo Blog, Accessed: 2026-03-13.
[3] 3GPP. Ng-ran; architecture description. Technical Report TS 38.401, 3GPP, 2024. [4] 3rd Generation Partnership Project (3GPP). 3GPP TSGSA Meeting #3 Report (SP-99100). Technical report, 3GPP, 1999.
[13] Mohammed S. Elbamby, Cristina Perfecto, Mehdi Bennis, and Klaus Doppler. Toward low-latency and ultrareliable virtual reality. IEEE Network, 32(2):78–84, 2018.
[5] Forough Shirin Abkenar, Parisa Ramezani, Saeid Iranmanesh, Sarumathi Murali, Donpiti Chulerttiyawong, Xinyu Wan, Abbas Jamalipour, and Raad Raad. A survey on mobility of edge computing networks in iot: State-of-the-art, architectures, and challenges. IEEE Communications Surveys & Tutorials, 24(4):2329–2365, 2022.
[14] Håkon Medhus Fornes, Elias Hoel Birketvedt, Carsten Griwodz, Magnus Skjegstad, Michael Welzl, and Özgü Alay. Acceptable latency in predictable first-person vr cloud games. In Proceedings of the 17th International Workshop on IMmersive Mixed and Virtual Environment Systems, pages 58–64, 2025.
[6] Mukhtiar Ahmad, Faaiq Bilal, Mutahar Ali, Syed Muhammad Ali Nawazish, Amir Salman, Shazer Ali, Fawad Ahmad, and Zafar Ayyub Qazi. Warping the edge: Enabling instant mobility for stateful applications over 5g and beyond. In Proceedings of the Tenth ACM/IEEE Symposium on Edge Computing, pages 1–18, 2025.
[15] Ahmad Hassan, Arvind Narayanan, Anlan Zhang, Wei Ye, Ruiyang Zhu, Shuowei Jin, Jason Carpenter, Z Morley Mao, Feng Qian, and Zhi-Li Zhang. Vivisecting mobility management in 5g cellular networks. In Proceedings of the ACM SIGCOMM 2022 Conference, pages 86–100, 2022. [16] Hewlett Packard Enterprise. Aruba wi-fi 6 networks deployment guide. https://higherlogicdownload. s3.amazonaws.com/HPE/UploadedImages/ 2e31f761-75c1-4680-8d55-009a31e45948/ Aruba_WiFi6_Networks_Deployment_Guide.pdf, 2020. Accessed: 2026.
[7] Ghada Alsuhli, Karim Banawan, Kareem Attiah, Ayman Elezabi, Karim G Seddik, Ayman Gaber, Mohamed Zaki, and Yasser Gadallah. Mobility load management in cellular networks: A deep reinforcement learning approach. IEEE Transactions on Mobile Computing, 22(3):1581– 1598, 2021.
[17] Shengyi Huang and Santiago Ontañón. A closer look at invalid action masking in policy gradient algorithms. arXiv preprint arXiv:2006.14171, 2020.
[8] AssemblyAI. The 300ms rule: Why latency makes or breaks voice ai. https://www.assemblyai. com/blog/low-latency-voice-ai, 2025. Accessed: 2026-03-13.
[18] Johanna Andrea Hurtado Sánchez, Katherine Casilimas, and Oscar Mauricio Caicedo Rendon. Deep reinforcement learning for resource management on network slicing: A survey. Sensors, 22(8):3031, 2022.
[9] Raouf Boutaba, Mohammad A Salahuddin, Noura Limam, Sara Ayoubi, Nashid Shahriar, Felipe EstradaSolano, and Oscar M Caicedo. A comprehensive survey on machine learning for networking: evolution, applications and research opportunities. Journal of Internet Services and Applications, 9(1):1–99, 2018.
[19] Dahyun Kang, Changjoo Nam, and Sonya S Kwak. Robot feedback design for response delay. International Journal of Social Robotics, 16(2):341–361, 2024.
[10] Dries Cardinaels, Tom Veuskens, Raf Ramakers, and Kris Luyten. Challenges and opportunities for delayinvariant telerobotic interactions. In Companion Proceedings of the 30th International Conference on Intelligent User Interfaces, pages 26–30, 2025.
[20] Minwoo Kim, Jonggyu Jang, Youngchol Choi, and Hyun Jong Yang. Distributed task offloading and resource allocation for latency minimization in mobile edge computing networks. IEEE Transactions on Mobile Computing, 23(12):15149–15166, 2024. 13
[21] Adam Langley, Alistair Riddoch, Alyssa Wilk, Antonio Vicente, Charles Krasic, Dan Zhang, Fan Yang, Fedor Kouranov, Ian Swett, Janardhan Iyengar, et al. The quic transport protocol: Design and internet-scale deployment. In Proceedings of the conference of the ACM special interest group on data communication, pages 183–196, 2017.
[30] Tan Nguyen. Low-latency mixed reality headset. Lowlatency VR/AR Headset project from Conix Research Center, Computing On Network Infrastructure for Pervasive Perception, Cognition and Action, 2:9, 2020. [31] Nokia. Tackling the network challenges for xr: Understanding, developing and delivering an e2e xr connectivity solution. White paper, Nokia, 2022.
[22] Yuanjie Li, Qianru Li, Zhehui Zhang, Ghufran Baig, Lili Qiu, and Songwu Lu. Beyond 5g: Reliable extreme mobility management. In Proceedings of the Annual conference of the ACM Special Interest Group on Data Communication on the applications, technologies, architectures, and protocols for computer communication, pages 344–358, 2020.
[32] O-RAN Software Community. O-RAN Software Community: Near-Real-Time RAN Intelligent Controller (Near-RT RIC). https://docs. o-ran-sc.org/en/latest/projects.html# near-realtime-ran-intelligent-controller-ric. Accessed: 2026-03-26. [33] Emmanouil Pateromichelakis, Mehrdad Shariat, Atta Ul Quddus, and Rahim Tafazolli. Graph-based multicell scheduling in ofdma-based small cell networks. IEEE Access, 2:897–908, 2014.
[23] Jie Lu, Peihao Yan, and Huacheng Zeng. Eexapp: Gnnbased reinforcement learning for radio unit energy optimization in 5g o-ran. arXiv preprint arXiv:2602.09206, 2026.
[34] Yarin Perry, Felipe Vieira Frujeri, Chaim Hoch, Srikanth Kandula, Ishai Menache, Michael Schapira, and Aviv Tamar. {DOTE}: Rethinking (predictive){WAN} traffic engineering. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), pages 1557–1581, 2023.
[24] Erfan Farhangi Maleki, Weibin Ma, Lena Mashayekhy, and Humberto J La Roche. Qos-aware content delivery in 5g-enabled edge computing: Learning-based approaches. IEEE Transactions on Mobile Computing, 23(10):9324–9336, 2024. [25] Hongzi Mao, Ravi Netravali, and Mohammad Alizadeh. Neural adaptive video streaming with pensieve. In Proceedings of the conference of the ACM special interest group on data communication, pages 197–210, 2017.
[35] Anna Prado, Franziska Stöckeler, Fidan Mehmeti, Patrick Krämer, and Wolfgang Kellerer. Enabling proportionally-fair mobility management with reinforcement learning in 5g networks. IEEE Journal on Selected Areas in Communications, 41(6):1845–1858, 2023.
[26] Hongzi Mao, Malte Schwarzkopf, Shaileshh Bojja Venkatakrishnan, Zili Meng, and Mohammad Alizadeh. Learning scheduling algorithms for data processing clusters. In Proceedings of the ACM special interest group on data communication, pages 270–288. 2019.
[36] R Raffik, VJ Balamurugan, and N Gopinath Pandian. Edge computing for real-time decision-making in industrial automation systems-a comprehensive review. In 2025 3rd International Conference on Advancements in Electrical, Electronics, Communication, Computing and Automation (ICAECA), pages 1–6. IEEE, 2025.
[27] Nazanin Mehregan and E Robson. Gcn-based throughput-oriented handover management in dense 5g vehicular networks. In 2025 21st International Conference on Distributed Computing in Smart Systems and the Internet of Things (DCOSS-IoT), pages 895–902. IEEE, 2025.
[37] Andrey Samuylov, Dmitri Moltchanov, Anna Gaydamaka, Eike Lyczkowski, Andreas Frotzscher, Frank von Schoettler, Juho Pirskanen, Jussi Numminen, Juha Salokannel, Endika Llaguno, et al. Empowering nearurllc iot with 5g dect-2020 nr: Current state and the road ahead. IEEE Communications Magazine, 63(4):130– 136, 2025.
[28] Michael S Mollel, Attai Ibrahim Abubakar, Metin Ozturk, Shubi Felix Kaijage, Michael Kisangiri, Sajjad Hussain, Muhammad Ali Imran, and Qammer H Abbasi. A survey of machine learning applications to handover management in 5g and beyond. IEEE Access, 9:45770– 45802, 2021.
[38] Yang Shen, Shuqi Chai, Bing Li, Xiaodong Luo, Qingjiang Shi, and Rongqing Zhang. Decentralized handover parameter optimization with marl for load balancing in 5g networks. arXiv preprint arXiv:2504.13424, 2025.
[29] Zainab Yusufali Motiwala, Arusha Desai, Rahul Bisht, Saloni Lathkar, Sidharth Misra, and Danny Darlington Carbin. Telesurgery: current status and strategies for latency reduction. Journal of Robotic Surgery, 19(1):153, 2025.
[39] Yibin Shen and Zili Meng. Law: Towards consistent low latency in 802.11 home networks. In Proc. USENIX NSDI, 2026. 14
[40] Yixin Shen, Ruihua Chen, Bo Wang, Jing Chen, Haochen Zhang, Minhu Wang, Yan Liu, Mingwei Xu, and Zili Meng. Mortise: Auto-tuning congestion control to optimize qoe via network-aware parameter optimization.
[50] Zhiying Wang, Gang Sun, Hanyue Su, Hongfang Yu, Bo Lei, and Mohsen Guizani. Low-latency scheduling approach for dependent tasks in mec-enabled 5g vehicular networks. IEEE Internet of Things Journal, 11(4):6278–6289, 2023.
[41] Software Radio Systems Ltd. srsRAN Project. https: //github.com/srsran/srsRAN_Project. Accessed: 2026-03-26.
[51] Jiali Xing, Akis Giannoukos, Paul Loh, Shuyue Wang, Justin Qiu, Henri Maxime Demoulin, Konstantinos Kallas, and Benjamin C Lee. Rajomon: Decentralized and coordinated overload control for {LatencySensitive} microservices. In 22nd USENIX Symposium on Networked Systems Design and Implementation (NSDI 25), pages 21–36, 2025.
[42] Sixu Tan, Zeyu Li, Zhutian Liu, Harsh Patel, and Zhaowei Tan. Automated model-based fuzzing for 5g o-ran. In Proceedings of the 31st Annual International Conference on Mobile Computing and Networking, pages 201–215, 2025.
[52] Zhiying Xu, Francis Y Yan, Rachee Singh, Justin T Chiu, Alexander M Rush, and Minlan Yu. Teal: Learningaccelerated optimization of wan traffic engineering. In Proceedings of the ACM SIGCOMM 2023 Conference, pages 378–393, 2023.
[43] Zhaowei Tan, Jinghao Zhao, Yuanjie Li, Yifei Xu, and Songwu Lu. {Device-Based}{LTE} latency reduction at the application layer. In 18th USENIX Symposium on Networked Systems Design and Implementation (NSDI 21), pages 471–486, 2021.
[53] Peihao Yan, Jie Lu, Huacheng Zeng, and Y Thomas Hou. Near-real-time resource slicing for qos optimization in 5g o-ran using deep reinforcement learning. IEEE Transactions on Networking, 34:1596–1611, 2025.
[44] Muhammad Tayyab, Xavier Gelabert, and Riku Jäntti. A survey on handover management: From lte to nr. IEEE Access, 7:118907–118930, 2019.
[54] Peihao Yan, Huacheng Zeng, and Y Thomas Hou. xdiff: Online diffusion model for collaborative inter-cell interference management in 5g o-ran. IEEE Transactions on Networking, 2025.
[45] Muhammad Tayyab, George P Koudouridis, Xavier Gelabert, and Riku Jäntti. Signaling overhead and power consumption during handover in lte. In 2019 IEEE wireless communications and networking conference (WCNC), pages 1–6. IEEE, 2019.
[55] Wei Ye, Xinyue Hu, Steven Sleder, Anlan Zhang, Udhaya Kumar Dayalan, Ahmad Hassan, Rostand AK Fezeu, Akshay Jajoo, Myungjin Lee, Eman Ramadan, et al. Dissecting carrier aggregation in 5g networks: Measurement, qoe implications and prediction. In Proceedings of the ACM SIGCOMM 2024 Conference, pages 340–357, 2024.
[46] Rahul Dev Tripathi, Minzhao Lyu, and Vijay Sivaraman. Assessing the impact of network quality-of-service on metaverse virtual reality user experience. In 2024 IEEE International Conference on Metaverse Computing, Networking, and Applications (MetaCom), pages 206–213. IEEE, 2024. [47] Yuk Hang Tsui, Ze Wu, Ahmad Alhilal, Matti Siekkinen, and Pan Hui. Fovrl: Joint foveation and quality control for immersive vr streaming using reinforcement learning. In Proceedings of the ACM Web Conference 2026, pages 5648–5656, 2026. [48] Anouk Van Kasteren, Kjell Brunnström, John Hedlund, and Chris Snijders. Quality of experience of 360 video– subjective and eye-tracking assessment of encoding and freezing distortions. Multimedia tools and applications, 81(7):9771–9802, 2022. [49] Donglin Wang, Anjie Qiu, Sanket Partani, Qiuheng Zhou, and Hans D Schotten. Mitigating unnecessary handovers in ultra-dense networks through machine learning-based mobility prediction. In 2023 IEEE 97th Vehicular Technology Conference (VTC2023-Spring), pages 1–7. IEEE, 2023. 15
Appendix A
Actor-Critic Loss
The actor and critic are optimized using PPO with a clipped surrogate objective to ensure stable policy updates. h
Lactor = −Et min rt (θ)Ât , r̃t (θ)Ât
i
− cH H (πθ ),
r̃t (θ) = clip(rt (θ), 1 − ε, 1 + ε) h i Lcritic = Et max (Vφ (ot ) − Rt )2 , (Ṽt − Rt )2 ,
(17)
(18)
Ṽt = clip(Vφ (ot ),Vφold − ε, Vφold + ε) where rt (θ) = ππθ (a(at |ot |ot )t ) denotes the importance sampling
(a) PCA Visualization of UE Embedding.
θold
ratio, and Ât is the advantage estimate, computed using generalized advantage estimation (GAE). The clipping operation restricts policy updates within a trust region to prevent excessively large policy shifts. cH is the entropy coefficient that encourages exploration. For the critic, Vφ (ot ) is the value function and Rt is the empirical return. The clipped value loss stabilizes training by limiting large deviations from the previous value estimate Vφold (ot ). The policy πθ (ai | oi ) takes as input the TGN embedding oi = zi (t), and the value function Vφ (oi ) takes as input the concatenation si = [zi (t), g(t)], both parameterized by ψ.
B
Embedding Interpretability
To better understand the learned temporal representations, we visualize the UE embeddings generated by TGN. Fig. 23a shows a PCA projection of UE embeddings over time for multiple UEs. Each trajectory corresponds to a single UE, where consecutive points are connected in temporal order. We observe that embeddings evolve smoothly over time while remaining distinguishable across UEs, indicating that the TGN encoder captures both temporal continuity and UE-specific characteristics. Fig. 23b presents the pairwise cosine similarity matrix of time-ordered embeddings. The strong diagonal structure reflects high similarity between temporally adjacent samples, confirming temporal consistency in the learned representations. Meanwhile, the absence of uniformly high similarity across the matrix suggests that the embeddings preserve sufficient diversity to distinguish different network states.
C
(b) Pairwise Cosine Similarity of Time-Ordered UE Embeddings.
Figure 23: Temporal Dynamics and Consistency of TGNGenerated UE Embeddings Table 3: End-to-end latency requirements for some edge AI offloading applications. Category Immersive AI Haptic AI Industrial Edge AI General AI
Edge AI Offloading Tasks
Latency Requirement
VR/AR Motion-to-Photon [13, 14]
AR<10ms VR<20 ms
Cloud XR rendering [30]
<50 ms
Remote Robotic Surgery [29]
Vision<150 ms, Touch<10 ms
Industrial tele-robotics [10]
<10 ms
Collaborative Robots [19]
<20 ms
AI Vision Inspection [36]
<50 ms
High-Frequency Trading [12]
<100 ms
inference often operates in closed-loop pipelines in which sensor data must be transmitted to edge or cloud servers for processing, and the resulting decisions [36] must be returned within strict latency constraints. Table 3 shows the end-to-end delay requirements of edge AI offloading for some emerging applications, which includes both (uplink and downlink) communications delays as well as edge computing delays. The string delay requirements pose a grand challenge in the design and optimization of wireless networks.
Delay Requirements
Delay-critical AI applications are increasingly deployed in domains such as real-time video analytics, AR/VR, autonomous robotics (e.g., remote surgery), and industrial automation [8, 31]. In these applications, timely data delivery is critical to ensure responsive and reliable system behavior. AI 16