Paper deep dive
AgentComm-Bench: Stress-Testing Cooperative Embodied AI Under Latency, Packet Loss, and Bandwidth Collapse
Aayam Bansal, Ishaan Gangwani
Intelligence
Status: succeeded | Model: anthropic/claude-sonnet-4.6 | Prompt: intel-v1 | Confidence: 92%
Last extracted: 3/24/2026, 1:51:32 AM
Summary
AgentComm-Bench is a benchmark suite and evaluation protocol that stress-tests cooperative embodied AI systems under six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence. The benchmark spans three task families (cooperative perception, multi-agent waypoint navigation, and cooperative zone search) and evaluates five communication strategies including the proposed ResilientComm method, which uses redundant message coding with staleness-aware fusion. Key findings show that stale memory and bandwidth collapse cause over 96% performance drops in navigation, content corruption reduces perception F1 by over 85%, and ResilientComm more than doubles navigation performance under 80% packet loss.
Entities (54)
Relation Signals (42)
Aayam Bansal → affiliatedwith → Synthetic Sciences
confidence 98% · Aayam Bansal, Ishaan Gangwani Synthetic Sciences {aayam, ishaan}@syntheticsciences.ai
Ishaan Gangwani → affiliatedwith → Synthetic Sciences
confidence 98% · Aayam Bansal, Ishaan Gangwani Synthetic Sciences {aayam, ishaan}@syntheticsciences.ai
Aayam Bansal → authored → AgentComm-Bench
confidence 97% · Authors introduce AGENTCOMM-BENCH as their primary contribution
Ishaan Gangwani → authored → AgentComm-Bench
confidence 97% · Authors introduce AGENTCOMM-BENCH as their primary contribution
AgentComm-Bench → coversimpairment → Packet Loss
confidence 97% · six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence
AgentComm-Bench → coversimpairment → Stale Memory
confidence 97% · six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence
AgentComm-Bench → coversimpairment → Bandwidth Collapse
confidence 97% · six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence
AgentComm-Bench → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cooperative multi-agent methods for embodied AI are almost universally evaluated under idealized communication: zero latency, no packet loss, and unlimited bandwidth. Real-world deployment on robots with wireless links, autonomous vehicles on congested networks, or drone swarms in contested spectrum offers no such guarantees. We introduce AgentComm-Bench, a benchmark suite and evaluation protocol that systematically stress-tests cooperative embodied AI under six communication impairment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence. AgentComm-Bench spans three task families: cooperative perception, multi-agent waypoint navigation, and cooperative zone search, and evaluates five communication strategies, including a lightweight method we propose based on redundant message coding with staleness-aware fusion. Our experiments reveal that communication-dependent tasks degrade catastrophically: stale memory and bandwidth collapse cause over 96% performance drops in navigation, while content corruption (stale or conflicting data) reduces perception F1 by over 85%. Vulnerability depends on the interaction between impairment type and task design; perception fusion is robust to packet loss but amplifies corrupted data. Redundant message coding more than doubles navigation performance under 80% packet loss. We release AgentComm-Bench as a practical evaluation protocol and recommend that cooperative embodied AI work report performance under multiple impairment conditions.
Tags
Links
- Source: https://arxiv.org/abs/2603.20285v1
- Canonical: https://arxiv.org/abs/2603.20285v1
Trouble viewing inline? Open PDF directly →
Full Text
53,649 characters extracted from source content.
Expand or collapse full text
AgentComm-Bench: Stress-Testing Cooperative Embodied AI Under Latency, Packet Loss, and Bandwidth Collapse Aayam Bansal, Ishaan Gangwani Synthetic Sciences aayam, ishaan@syntheticsciences.ai Abstract Cooperative multi-agent methods for embodied AI are al- most universally evaluated under idealized communica- tion: zero latency, no packet loss, and unlimited band- width. Real-world deployment—on robots with wireless links, autonomous vehicles on congested networks, or drone swarms in contested spectrum—offers no such guarantees. We introduce AGENTCOMM-BENCH, a benchmark suite and evaluation protocol that systematically stress-tests co- operative embodied AI under six communication impair- ment dimensions: latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence. AGENTCOMM-BENCH spans three task families— cooperative perception, multi-agent waypoint navigation, and cooperative zone search—and evaluates five communi- cation strategies, including a lightweight method we propose called RESILIENTCOMM that applies redundant message coding with staleness-aware fusion. Our experiments reveal that communication-dependent tasks degrade catastrophi- cally: stale memory and bandwidth collapse cause>96% performance drops in navigation, while content corruption (stale/conflicting data) destroys perception F1 by>85%. Critically, vulnerability depends on the interaction between impairment type and task design—perception fusion is im- mune to packet loss but amplifies corrupted data. Redundant message coding more than doubles navigation performance under 80% packet loss. We release AGENTCOMM-BENCH as a practical evaluation protocol and recommend that all cooperative embodied AI papers report performance under at least three impairment conditions. 1. Introduction Cooperative multi-agent systems—from autonomous ve- hicles sharing perception [26,30] to robot teams coordi- nating exploration [33]—depend critically on inter-agent communication. The past five years have produced a rich landscape of learned communication protocols [2,21,24], communication-efficient methods [6,8,14], and coopera- tive perception systems [7,11,29]. Yet nearly all of these methods are evaluated under an implicit assumption: com- munication is instantaneous, lossless, and unlimited. This assumption is unrealistic. Wireless channels between robots experience packet loss rates of 5–30% [4], latencies of 50–500 ms, and bandwidth that fluctuates with interference and congestion. Vehicle-to-everything (V2X) links are sub- ject to asynchronous clock domains [34], stale information from processing delays [1], and conflicting sensor readings across heterogeneous platforms. While some recent methods address individual failure modes—mmCooper [23] handles delays in cooperative perception, V2XPnP [32] addresses pose noise, and BVME [36] applies variational bottlenecks for bandwidth constraints—no unified evaluation protocol examines the full spectrum of communication impairments. We introduce AGENTCOMM-BENCH, a benchmark suite and evaluation protocol for stress-testing cooperative embod- ied AI under realistic communication failures. Our contribu- tions are: •A six-dimensional communication stress protocol cover- ing latency, packet loss, bandwidth collapse, asynchronous updates, stale memory, and conflicting sensor evidence, with parameterized severity levels enabling fine-grained robustness curves. • Three complementary task families—cooperative per- ception, multi-agent waypoint navigation, and cooperative zone search—instantiated as lightweight grid-world simu- lations that isolate communication effects from perceptual complexity. • RESILIENTCOMM, a lightweight robust communica- tion wrapper combining redundant message coding with staleness-aware fusion, demonstrating that simple engi- neering principles can improve robustness under packet loss. •A standardized evaluation protocol including normal- ized performance drop, robustness curves, rank stability analysis, and failure mode taxonomy, which we recom- mend as a reporting standard for cooperative embodied AI. 1 arXiv:2603.20285v1 [cs.AI] 18 Mar 2026 We position AGENTCOMM-BENCH as a practical evalua- tion protocol: the benchmark requires no specialized hard- ware, runs in under five minutes on a single CPU, and pro- duces a comprehensive robustness profile for any cooperative method. The lightweight grid-world instantiation serves as a reference implementation; we encourage the community to apply the protocol to photorealistic simulators such as CARLA [3] or Habitat 3.0 [20] and to standard cooperative perception datasets [30, 34]. 2. Related Work Cooperative multi-agent communication.Learned com- munication protocols for multi-agent systems have evolved from broadcast-based architectures like CommNet [24] to targeted messaging via attention [2] and gated communica- tion [21]. For cooperative perception, Who2com [15] and When2com [14] introduced handshake-based protocols that learn whom and when to communicate. V2VNet [26] estab- lished intermediate feature sharing as the dominant paradigm, further refined by DiscoNet [11] with knowledge distillation and Where2comm [6] with spatial confidence maps achiev- ing 100,000×bandwidth reduction. Transformer-based fu- sion [28,29] enables robust multi-agent feature aggregation. CoAlign [17] addresses pose errors, and CoCa3D [7] demon- strates that collaborative cameras can rival LiDAR. Despite this progress, robustness to communication channel failures remains largely unexamined. Communication under realistic constraints. A recent survey [16] identifies that most MARL methods assume in- stantaneous, reliable communication. Exceptions include SchedNet [8], which learns communication scheduling un- der bandwidth limits, ETCNet [5], which applies event- triggered communication to reduce bandwidth occupancy, and IMAC [25], which uses information bottleneck princi- ples for message compression. Li and Zhang [9] propose context-aware personalized messages under bandwidth con- straints. ET-MAPG [10] learns adaptive event triggers for communication, BVME [36] applies variational bottlenecks for bandwidth-adaptive messaging, and C-MADDPG [37] uses loss-aware mutual information shaping. For coopera- tive perception specifically, mmCooper [23] handles delays via confidence-guided multi-stage fusion, and V2XPnP [32] addresses pose noise with temporal BEV compression. For asynchronous settings, ACE [33] extends MAPPO to han- dle action delays in multi-robot exploration, while Min et al. [19] provide convergence guarantees for asynchronous cooperative MARL. Zhou et al. [38] apply Bayesian belief updates to handle packet loss. However, these works each address a single failure mode; no unified benchmark eval- uates methods across the full spectrum of communication impairments. Figure 1. Overview of AGENTCOMM-BENCH. The benchmark evaluates cooperative methods across three task families under six communication impairment dimensions, producing robustness profiles via standardized metrics. Benchmarks for cooperative embodied AI.The cooper- ative perception community has produced several datasets: OPV2V [30] (simulated V2V), V2X-Sim [12] (multi- agent simulation), DAIR-V2X [34] (real-world vehicle- infrastructure), and V2V4Real [31] (real-world V2V). Pho- torealistic simulators including CARLA [3] provide con- trollable environments for V2X research. For embodied navigation, Habitat 3.0 [20] supports human-robot collabora- tion, and Stern et al. [22] formalize multi-agent pathfinding benchmarks. LLM-based cooperative agents [18,35] intro- duce new evaluation paradigms. V2XP-ASG [27] provides adversarial scene generation for V2X perception. Lin et al. [13] demonstrate that attacking a single agent’s observations can collapse team performance. Yet none of these bench- marks systematically evaluate the effect of communication channel impairments on cooperative performance. 3. The AGENTCOMM-BENCH Benchmark AGENTCOMM-BENCH is designed around three principles: (1) comprehensive coverage of realistic communication failure modes, (2) reproducible evaluation requiring only lightweight simulation, and (3) standardized reporting via a fixed set of robustness metrics. Figure 1 provides an overview of the benchmark architecture. 3.1. Communication Impairment Dimensions We define six impairment dimensions, each parameterized by a scalar severityσ ∈ [0,σ max ]. While latency, asynchrony, and stale memory share a temporal character, they differ in mechanism (transport delay vs. clock drift vs. refresh failure) and produce distinct degradation patterns (Appendix A.11): D1.Latency (σ ∈ [0, 500]ms). Messages arrive with a fixed delay. Atσ=500ms, agents operate on obser- vations that are 10–15 decision steps old, a realistic scenario for cellular V2X links. 2 Figure 2. Communication corruption pipeline. Each message passes through the configurable impairment channel before reach- ing the receiver. The six dimensions are applied independently by default. D2.Packet loss (σ ∈ [0, 80]%). Each message is inde- pendently dropped with probabilityσ/100, following a Bernoulli model. At 80% loss, agents receive on average only 1 in 5 messages. D3.Bandwidth collapse (σ ∈ [0, 100]%reduction). The available channel capacity is reduced byσ%, forcing methods that transmit large feature maps to either com- press or drop information. D4. Asynchronous updates (σ ∈ [0, 10]steps). Agents operate on different clock domains, with each agent’s state offset by a uniform random delay in[0,σ]steps from the true environment state. D5.Stale memory (σ ∈ [0, 20]steps). Agents’ internal models of other agents’ states are not refreshed forσ steps, simulating scenarios where communication links intermittently fail. D6.Conflicting sensor evidence (σ ∈ [0, 40]%). A frac- tionσ/100of each agent’s observations are corrupted with structured noise (e.g., false positives, incorrect positions), simulating sensor disagreement across het- erogeneous platforms. For each dimension, we sweep 11 evenly spaced sever- ity levels from 0 toσ max , producing smooth degradation curves. These dimensions are applied independently; joint impairment analysis is explored in Appendix A.4. Figure 2 illustrates the communication corruption pipeline. 3.2. Task Families We select three cooperative task families that span different communication requirements: Cooperative Perception (CP). Four agents observe a shared environment from different viewpoints and must pro- duce a fused object detection output. Performance is mea- sured by F1 score (harmonic mean of precision and recall of detected objects). This task is communication-intensive: agents must share spatial features to resolve occlusions and expand effective field-of-view, mirroring the V2X coopera- tive perception setting [29, 30]. Multi-Agent Navigation (NAV). A coordinator assigns each of four agents a sequence of three waypoints on a 20×20grid, broadcast as one-hot position vectors. Each agent decodes its target viaargmaxand navigates toward it; upon reaching a waypoint (Manhattan distance≤ 1), the agent advances to the next in its sequence. Performance is measured by waypoint completion rate (fraction of total waypoints reached by the team). Communication is criti- cal: without messages, agents have no knowledge of their assigned targets and reduce to random walks (3.6% comple- tion). This task tests whether agents can follow communi- cated directives under degraded channels. Cooperative Search (SEARCH). Four agents perform coordinated zone search over a20×20grid containing 20 hidden targets. A coordinator assigns each agent a sequence of four zone waypoints; agents navigate to assigned zones and detect targets only on the exact cell they occupy (point detection). Performance is measured by recall (fraction of targets found by the team). Communication enables agents to follow zone assignments, avoiding redundant coverage; without messages, agents search randomly (20.7% recall vs. 26.8% with coordination). 3.3. Evaluation Metrics AGENTCOMM-BENCH defines four evaluation axes: Normalized Performance Drop (NPD).For a methodm under impairment d at severity σ: NPD(m,d,σ) = P m (σ=0)− P m (σ) P m (σ=0) × 100%(1) whereP m (σ)is the mean task performance. NPD normal- izes for differences in clean-condition performance, enabling fair comparison of methods with different baselines. Robustness Curves. The full performance-vs-severity curveP m (σ)for each(m,d)pair, plotted with confidence intervals across episodes. The area under the robustness curve (AURC) provides a single-number summary. Rank Stability. We compute the method ranking at each severity level and report how rankings change from clean to maximum impairment. A method with high rank stability maintains its relative position regardless of channel quality. 3 Failure Mode Taxonomy.We categorize observed failure modes (e.g., “ghost detections from stale state,” “coverage collapse from coordination loss”) and report their frequency across impairment conditions, providing qualitative insight into how each method fails. 4. Methods Under Evaluation We evaluate five communication strategies spanning the spec- trum from no communication to our proposed robust method. All methods use identical agent architectures and differ only in communication policy and message processing. 4.1. Baseline Methods No-Comm. Each agent acts independently using only its own observations. This provides a lower bound on perfor- mance and an upper bound on robustness (since there is nothing to degrade). Full-Comm (Oracle).All agents share their complete ob- servation vectors at every step with zero loss. Messages are flattened grid vectors of dimensiond=G 2 (whereG=20 is the grid size), transmitted at full 32-bit floating-point precision. This provides an upper bound on performance under perfect communication but is maximally sensitive to channel impairments. Per-step communication load: N(N−1)× d× 32bits= 12× 400× 32 = 153,600bits per step for N=4 agents. Compressed-Comm.Agents share 4-bit quantized feature vectors, reducing message size by8×relative to Full-Comm (19,200bits per step). This mirrors the intermediate fusion paradigm of cooperative perception [26,30]. While compres- sion reduces bandwidth requirements, it does not inherently protect against packet loss or latency. Event-Triggered Comm.Agents transmit messages only when their local information gain exceeds a thresholdθ=0.5 (measured asL 1 norm of the observation vector), follow- ing the event-triggered paradigm [5]. When communication does occur, full messages are sent. This strategy naturally reduces communication volume: in our experiments, agents transmit in approximately 60–80% of steps depending on the task. Similar adaptive triggering has been explored by ET-MAPG [10], though with learned rather than fixed thresh- olds. 4.2. RESILIENTCOMM (Proposed) We propose RESILIENTCOMM, a lightweight communica- tion wrapper designed for graceful degradation under ar- bitrary channel impairments (Figure 3). It combines two mechanisms: Figure 3. RESILIENTCOMM architecture. Redundant message coding sends two copies; staleness-aware fusion weights received messages by estimated age. When both copies are lost, the agent falls back to its most recent received state. Redundant message coding.Each message is transmitted twice:m t (the current encoded state) is sent as two inde- pendent copies through the channel. If either copy arrives, the receiver uses it; if both copies arrive, the receiver uses the first. This doubles the per-message communication cost (2× 153,600 = 307,200bits per step), but reduces the effec- tive packet loss rate fromptop 2 . At 80% per-message loss, the probability that both copies are dropped is0.8 2 = 0.64, reducing the effective loss from 80% to 64%. Staleness-aware fusion.When fusing received messages, RESILIENTCOMM weights each message by an inverse- staleness factor: w i = exp(−λ· τ i ) P j exp(−λ· τ j ) (2) whereτ i is the estimated age (in steps) of thei-th message andλ=0.3is a fixed temperature parameter. Message age is estimated from the difference between the receiver’s local step counter and the sender’s timestamp embedded in the message header (a 16-bit field). Under asynchronous clocks, this estimate may be inaccurate; we do not model clock synchronization overhead in our current implementation. When no messages are received, the agent uses its most recent received state without extrapolation, smoothly falling back toward No-Comm behavior. Complexity and overhead.RESILIENTCOMM adds min- imal compute overhead: the redundant encoding requires one additional message copy per neighbor per step (O(N) memory), and the staleness fusion isO(N)per agent. The bandwidth overhead is2×(two copies of each message), which is substantial but bounded. No additional training is required; RESILIENTCOMM operates as a communication middleware that wraps any underlying cooperative policy. We note that RESILIENTCOMM’s mechanisms—redundant 4 Table 1. Clean-condition performance (σ=0). Communication provides massive benefit for navigation (waypoint coordination) and moderate benefit for search. MethodCP (F1)NAV (WP%)Search (Rec.%) No-Comm95.83.620.7 Event-Triggered95.896.726.8 Compressed95.896.726.8 RESILIENTCOMM (Ours)95.896.726.8 Full-Comm95.896.726.8 coding and staleness weighting—are standard engineering techniques [4]; the contribution is demonstrating their effec- tiveness within a systematic benchmark rather than claiming algorithmic novelty. 5. Experiments 5.1. Setup We instantiate AGENTCOMM-BENCH withN=4agents across all three task families using lightweight grid-world simulations (20×20 grid). Each experiment sweeps one impairment dimension across 11 severity levels (including σ=0for clean baseline), with 30 episodes per condition us- ing fixed per-episode seeds for reproducibility. All results are reported as mean±standard deviation. The full benchmark configuration is detailed in Appendix A.1. Implementation. For NAV and SEARCH, a coordinator encodes each agent’s current target waypoint as a one-hot vector over the20×20grid (d=400), which is transmit- ted through the configurableCommChannel. Agents de- code targets viaargmaxand navigate accordingly. For CP, agents share flattened detection grids fused via element- wise maximum. The one-hot waypoint encoding means that compression (4-bit quantization) and event triggering (L 1 > 0.5) preserve the signal perfectly—the peak survives rounding and always exceeds the trigger threshold—so Full- Comm, Compressed, and Event-Triggered produce identical NAV/SEARCH results. The key differentiator among meth- ods is redundancy: RESILIENTCOMM’s dual transmission mitigates packet loss. We emphasize that the benchmark’s primary contribution is the evaluation protocol, which can be applied to any cooperative method and simulator. 5.2. Clean-Condition Performance Table 1 reports baseline performance under perfect commu- nication (σ=0 for all impairments). Communication provides a dramatic benefit for naviga- tion: waypoint completion jumps from 3.6% (No-Comm, random walks) to 96.7% (all communicating methods)—a 93-point improvement—because agents require communi- cated waypoints to know where to go. Search shows a mod- erate benefit (20.7%→26.8% recall), as zone assignments improve coverage efficiency. For cooperative perception, the grid-world implementation yields high F1 across all methods because local detections within each quadrant are already re- liable; the CP task’s value lies in testing sensitivity to content corruption (stale/conflicting data) rather than communication necessity. 5.3. Robustness Under Communication Stress Figure 4 presents the main result: degradation curves across all 18 task–impairment combinations (3 tasks×6 impair- ments). Key finding 1: Communication-dependent tasks degrade catastrophically across all impairment types. Naviga- tion shows severe degradation under every impairment di- mension: stale memory causes the sharpest drop (96.7% →2.5%, NPD= 97.4%), followed by bandwidth collapse (96.7%→3.6%, NPD= 96.3%) and packet loss (96.7% →10.0%, NPD= 89.7%). Even latency—often consid- ered benign—reduces waypoint completion to 65.8% (NPD = 31.9%) because delayed waypoint updates cause agents to chase stale targets. This reveals that any communication im- pairment is destructive when agents fundamentally depend on received messages. Key finding 2: Content corruption (stale/conflict) causes catastrophic perception degradation. Under stale mem- ory (D5) and conflicting sensor evidence (D6), all communi- cating methods for cooperative perception experience severe F1 drops—from 95.8 to 14.0 (NPD= 85.4%)—because corrupted messages generate false positives that overwhelm the fused detection map. Critically, the CP task is robust to the four transport/temporal impairments (latency, packet loss, bandwidth, async) because itsnp.maximumfusion is naturally tolerant of missing or delayed messages. This asymmetry demonstrates that vulnerability depends on the interaction between impairment type and fusion mechanism. Key finding 3: RESILIENTCOMM provides measurable robustness under packet loss.For navigation under 80% packet loss, RESILIENTCOMM retains 21.9% waypoint com- pletion compared to 10.0% for single-message methods (Full-Comm, Compressed, Event-Triggered)—more than double the performance. The redundancy mechanism re- duces the effective loss from 80% to 64% (p 2 residual). RE- SILIENTCOMM also shows advantage under asynchronous updates (69.7% vs. 52.2% for Full-Comm at maximum sever- ity), as dual transmission increases the probability that at least one copy arrives within the agent’s decision window. These advantages hold under bursty Gilbert-Elliott channels (Appendix A.5). 5 0100200300400500 93 94 95 96 97 98 Cooperative Perception (F1) Latency (ms) 01020304050607080 93 94 95 96 97 98 Packet Loss (%) 020406080100 93 94 95 96 97 98 BW Reduction (%) 0246810 93 94 95 96 97 98 Async Offset (steps) 0.02.55.07.510.012.515.017.520.0 20 40 60 80 100 Staleness (steps) 0510152025303540 20 40 60 80 100 Conflict (%) 0100200300400500 0 20 40 60 80 100 Navigation (Waypoint %) 01020304050607080 0 20 40 60 80 100 020406080100 0 20 40 60 80 100 0246810 0 20 40 60 80 100 0.02.55.07.510.012.515.017.520.0 0 20 40 60 80 100 0510152025303540 0 20 40 60 80 100 0100200300400500 10 15 20 25 30 35 40 Search (Recall %) 01020304050607080 10 15 20 25 30 35 020406080100 10 15 20 25 30 35 0246810 10 15 20 25 30 35 0.02.55.07.510.012.515.017.520.0 0 5 10 15 20 25 30 35 0510152025303540 10 15 20 25 30 35 No-CommFull-CommCompressed-CommEvent-TriggeredResilientComm Figure 4. Robustness curves across all task–impairment combinations. Each subplot shows performance (mean±std over 30 episodes) as communication impairment severity increases. Navigation shows the most dramatic degradation, with all six impairments causing monotonic performance drops from near-perfect coordination to near-random-walk levels. Shaded regions indicate±1standard deviation. Table 2. Mean Normalized Performance Drop (%) across all im- pairments at maximum severity. Lower is more robust. No-Comm is trivially robust (0% NPD) because it has no communication to degrade. MethodCPNAVSearchAvg↓ No-Comm0.00.00.00.0 RESILIENTCOMM (Ours)28.563.631.441.1 Event-Triggered28.568.631.142.7 Compressed28.568.431.542.8 Full-Comm28.568.431.542.8 5.4. Normalized Performance Drop Analysis Figure 5 shows the NPD heatmap across all methods, tasks, and impairments at maximum severity. Several patterns emerge: •Navigation is the most vulnerable task: All six impair- ments cause substantial degradation (>30% NPD), with stale memory (97.4%), bandwidth collapse (96.3%), and packet loss (89.7%) approaching total coordination failure. This is expected: agents that depend entirely on communi- cated waypoints lose all useful behavior when messages are corrupted or absent. •CP is selectively vulnerable: Thenp.maximumfu- sion makes CP immune to transport and temporal im- pairments (0% NPD for latency, packet loss, bandwidth, async) but catastrophically vulnerable to content corrup- tion (stale/conflict>85% NPD). This reveals a fundamen- tal asymmetry: fusion mechanisms that tolerate missing data may amplify corrupted data. •Search degrades moderately: Stale memory causes the largest search degradation (67.7% NPD), while other im- pairments cause 15–35% NPD. The moderate communica- tion value (6.1-point gap) limits the damage ceiling. Area Under Robustness Curve (AURC) analysis (Ap- pendix A.9) confirms these patterns quantitatively: RE- SILIENTCOMM achieves the highest AURC among commu- nicating methods (75.9% vs. 73.1% for Full-Comm), with the largest advantage in NAV (69.8% vs. 63.3%). 5.5. Rank Stability Analysis Figure 6 reveals how method rankings shift under stress. Under clean conditions, all communicating methods tie for first in each task. Under maximum-severity impairments, RESILIENTCOMM gains a clear advantage for navigation under packet loss and async conditions (where dual transmis- sion helps), while No-Comm overtakes all communicating methods for CP under stale/conflict conditions. For percep- tion, the meaningful signal is that No-Comm becomes rank-1 under content corruption because it never receives corrupted messages. The key insight is that the optimal communication strategy depends on the dominant impairment type. 5.6. Observed Failure Modes From qualitative analysis of degraded episodes, we identify three recurring failure modes: 6 Latency Packet Loss BW Reduction Async Offset Staleness Conflict No-Comm / CP No-Comm / NAV No-Comm / SEARCH Full-Comm / CP Full-Comm / NAV Full-Comm / SEARCH Compressed-Comm / CP Compressed-Comm / NAV Compressed-Comm / SEARCH Event-Triggered / CP Event-Triggered / NAV Event-Triggered / SEARCH ResilientComm / CP ResilientComm / NAV ResilientComm / SEARCH 0.00.00.00.00.00.0 0.00.00.00.00.00.0 0.00.00.00.00.00.0 0.00.00.00.085.485.4 31.989.796.346.097.449.1 15.529.223.018.667.734.8 0.00.00.00.085.485.4 31.989.796.346.097.449.1 15.529.223.018.667.734.8 0.00.00.00.085.485.4 31.989.796.346.697.449.7 15.529.223.021.167.729.8 0.00.00.00.085.485.4 31.977.396.327.997.151.1 15.528.023.015.569.636.6 Normalized Performance Drop (%) at Maximum Severity 0 10 20 30 40 50 60 70 80 NPD (%) Figure 5. Normalized Performance Drop (%) at maximum impairment severity. Each cell shows the percentage of clean performance lost. Navigation dominates: stale memory and bandwidth collapse cause>96% NPD, effectively reverting coordinated agents to random walks. CP shows extreme sensitivity only to content corruption (stale/conflict >85%). F1.Waypoint loss under bandwidth/stale: When band- width collapses to zero or memory becomes stale, agents lose their waypoint assignments and revert to random walks, converging to the No-Comm baseline (96.7%→ 3.6%). The transition is abrupt for stale memory (sharp cliff at σ=8) but gradual for bandwidth collapse. F2. Detection hallucination under stale/conflict: Cor- rupted messages in perception generate false positive detections at positions where no objects exist, causing F1 to plummet as precision drops (95.8→ 14.0). F3.Graceful isolation under packet loss: RESILIENT- COMM’s redundant coding allows agents to maintain waypoint tracking at high loss rates. At 80% loss, RE- SILIENTCOMM retains 21.9% completion vs. 10.0% for single-message methods—more than double. 6. Discussion 6.1. Key Implications Vulnerability is task–impairment specific, not univer- sal. Our results reveal a fundamental asymmetry: CP’s np.maximumfusion is immune to transport impairments (0% NPD for latency, packet loss, bandwidth, async) but catastrophically vulnerable to content corruption (>85% NPD for stale/conflict). Conversely, NAV degrades severely under every impairment because agents depend entirely on 7 Clean Latency Packet Loss BW Reduction Async Offset Staleness Conflict 1 2 3 4 5 Rank (1=best) Cooperative Perception (F1) No-Comm Full-Comm Compressed-Comm Event-Triggered ResilientComm Clean Latency Packet Loss BW Reduction Async Offset Staleness Conflict 1 2 3 4 5 Rank (1=best) Navigation (Waypoint %) Clean Latency Packet Loss BW Reduction Async Offset Staleness Conflict 1 2 3 4 5 Rank (1=best) Search (Recall %) Figure 6. Method rankings across conditions. Rank 1 is best. Under clean communication, all communicating methods tie. Under impairments, RESILIENTCOMM shows rank-1 advantage for NAV packet loss and async; No-Comm dominates CP under content corruption. communicated waypoints. This implies that robustness strategies must be tailored to the interaction between the fusion mechanism and the expected failure mode—there is no universal defense. Communication can actively harm performance.Under stale memory and conflicting evidence, all communicating methods perform worse than No-Comm for perception (14.0 F1 vs. 95.8). Robust policies need a “circuit breaker” that suppresses fusion when channel quality degrades—a prin- ciple that event-triggered communication [5,21] partially addresses. Redundancy is the primary differentiator for trans- port impairments.Full-Comm, Compressed-Comm, and Event-Triggered produce identical results for NAV/SEARCH because the one-hot waypoint encoding is inherently robust to compression and always exceeds the event threshold. The only method that separates is RESILIENTCOMM, whose dual transmission provides measurable improvement under packet loss (21.9% vs. 10.0% at 80% loss) and async conditions (69.7% vs. 52.2%). This suggests that for communication- critical tasks, message redundancy matters more than adap- tive encoding or gating. 6.2. Recommended Evaluation Standard We propose that cooperative embodied AI papers report: (R1) clean performance; (R2) NPD under at least three im- pairment conditions (packet loss at 20%/50%, latency at 100/300 ms, plus one task-specific); (R3) a robustness curve for at least one impairment dimension; (R4) rank stability when comparing≥3 methods; and (R5) communication load in bits per step. 6.3. Limitations AGENTCOMM-BENCH is currently instantiated as grid- world tasks rather than photorealistic simulators [3,20] or standard datasets [30,31,34]. The CP task does not benefit from communication under clean conditions (No- Comm achieves identical F1), limiting its utility as a test of communication necessity; its value lies in demonstrating content-corruption vulnerability. The one-hot waypoint en- coding in NAV/SEARCH causes Full-Comm, Compressed, and Event-Triggered to converge, reducing method differ- entiation for transport impairments. We validate that the protocol’s findings hold under a bursty Gilbert-Elliott chan- nel model (Appendix A.5), but the current implementation does not model congestion feedback from duplicate transmis- sions. We also provide a bandwidth-normalized comparison (Appendix A.6). Our baselines are heuristic; learned meth- ods with explicit robustness (BVME [36], ET-MAPG [10], C-MADDPG [37]) would strengthen comparative conclu- sions. All experiments useN=4agents; scaling analysis is needed. 7. Conclusion We introduced AGENTCOMM-BENCH, a benchmark suite and evaluation protocol for stress-testing cooperative em- bodied AI under realistic communication impairments. By systematically sweeping six impairment dimensions across three task families, our experiments reveal that communication-dependent tasks degrade catastrophically— stale memory and bandwidth collapse cause>96% per- formance drops in navigation—while vulnerability is task– impairment specific: perception fusion is immune to trans- port failures but amplifies corrupted data (>85% NPD). Re- dundant message coding more than doubles navigation per- formance under 80% packet loss. These protocol-level find- ings are independent of the specific simulator and should generalize to more complex environments. We release AGENTCOMM-BENCH as a practical evaluation standard and recommend that cooperative embodied AI papers re- port performance under at least three impairment condi- tions. The most important future direction is applying the AGENTCOMM-BENCH protocol to standard cooperative per- ception datasets [30,34] with neural perception pipelines, 8 where communication impairments will interact with learned representations in ways our lightweight simulations cannot capture. References [1]Samuel Chamoun, Christian McDowell, Robin Buchanan, Kevin Chan, Eric Graves, and Yin Sun. MAPPO for edge server monitoring. In IEEE Military Communications Con- ference (MILCOM), 2025. 1 [2]Abhishek Das, Th ́ eophile Gervet, Joshua Romoff, Dhruv Batra, Devi Parikh, Michael Rabbat, and Joelle Pineau. Tar- MAC: Targeted multi-agent communication. In International Conference on Machine Learning (ICML), 2019. 1, 2, 13 [3] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Conference on Robot Learning (CoRL), 2017. 2, 8 [4]Touraj Farjam, Henk Wymeersch, and Themistoklis Char- alambous. Distributed channel access for control over known and unknown Gilbert–Elliott channels. IEEE Transactions on Control of Network Systems, 10(4):1983–1995, 2023. 1, 5, 11 [5]Guangzheng Hu, Yuanheng Zhu, Dongbin Zhao, Mengchen Zhao, and Jianye Hao. Event-triggered communication net- work with limited-bandwidth constraint for multi-agent rein- forcement learning. IEEE Transactions on Neural Networks and Learning Systems, 34(8):3966–3978, 2023. 2, 4, 8 [6]Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 1, 2, 13 [7]Yue Hu, Yifan Lu, Runsheng Xu, Weidi Xie, Siheng Chen, and Yanfeng Wang. Collaboration helps camera overtake Li- DAR in 3D detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 [8]Daewoo Kim, Sangwoo Moon, David Hostallero, Wan Ju Kang, Taeyoung Lee, Kyunghwan Son, and Yung Yi. Learn- ing to schedule communication in multi-agent reinforcement learning. In International Conference on Learning Represen- tations (ICLR), 2019. 1, 2 [9]Xinran Li and Jun Zhang. Context-aware communication for multi-agent reinforcement learning. In International Con- ference on Autonomous Agents and Multiagent Systems (A- MAS), 2024. 2 [10]Xiangyu Li, Kaiqing Zhang, and Jiayu Chen. Event-triggered multi-agent policy gradient. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2021. 2, 4, 8 [11]Yiming Li, Shunli Ren, Pengxiang Wu, Siheng Chen, Chen Feng, and Wenjun Zhang. Learning distilled collaboration graph for multi-agent perception. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 2 [12]Yiming Li, Dekun Ma, Ziyan An, Zixun Wang, Yiqi Zhong, Siheng Chen, and Chen Feng. V2X-Sim: Multi-agent col- laborative perception dataset and benchmark for autonomous driving. IEEE Robotics and Automation Letters, 2022. 2 [13]Jieyu Lin, Kristina Dzeparoska, Sai Qian Zhang, Alberto Leon-Garcia, and Nicolas Papernot. On the robustness of co- operative multi-agent reinforcement learning. arXiv preprint arXiv:2003.03722, 2020. 2 [14]Yen-Cheng Liu, Junjiao Tian, Nathaniel Glaser, and Zsolt Kira. When2com: Multi-agent perception via communication graph grouping. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1, 2 [15] Yen-Cheng Liu, Junjiao Tian, Chih-Yao Ma, Nathan Glaser, Chia-Wen Kuo, and Zsolt Kira. Who2com: Collaborative per- ception via learnable handshake communication. In IEEE In- ternational Conference on Robotics and Automation (ICRA), 2020. 2 [16] Zejiao Liu, Yi Li, Jiali Wang, Junqi Tu, Yitian Hong, Fangfei Li, Yang Liu, Toshiharu Sugawara, and Yang Tang. Robust and efficient communication in multi-agent reinforcement learning. arXiv preprint arXiv:2511.11393, 2025. 2 [17]Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3D object detection in presence of pose errors. In IEEE In- ternational Conference on Robotics and Automation (ICRA), 2023. 2 [18] Zhao Mandi, Shreeya Jain, and Shuran Song. RoCo: Dialectic multi-robot collaboration with large language models. arXiv preprint arXiv:2307.04738, 2023. 2 [19]Yifei Min, Jiafan He, Tianhao Wang, and Quanquan Gu. Co- operative multi-agent reinforcement learning: Asynchronous communication and linear function approximation. In In- ternational Conference on Machine Learning (ICML), 2023. 2 [20]Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dal- laire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander William Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars and robots. arXiv preprint arXiv:2310.13724, 2023. 2, 8 [21]Amanpreet Singh, Tushar Jain, and Sainbayar Sukhbaatar. Learning when to communicate at scale in multiagent cooper- ative and competitive tasks. In International Conference on Learning Representations (ICLR), 2019. 1, 2, 8 [22] Roni Stern, Nathan Sturtevant, Ariel Felner, Sven Koenig, Hang Ma, Thayne Walker, Jiaoyang Li, Dor Atzmon, Liron Cohen, T. K. Satish Kumar, Eli Boyarski, and Roman Bartak. Multi-agent pathfinding: Definitions, variants, and bench- marks. In Symposium on Combinatorial Search (SoCS), 2019. 2 [23]Haoqi Su, Yiming Li, Siheng Chen, and Chen Feng. Col- laborative 3d object detection with confidence-guided multi- modality fusion. In AAAI Conference on Artificial Intelli- gence, 2024. 1, 2, 12 [24]Sainbayar Sukhbaatar, Arthur Szlam, and Rob Fergus. Learn- ing multiagent communication with backpropagation. In Ad- vances in Neural Information Processing Systems (NeurIPS), 2016. 1, 2 [25] Rundong Wang, Xu He, Runsheng Yu, Wei Qiu, Bo An, and Zinovi Rabinovich. Learning efficient multi-agent communi- cation: An information bottleneck approach. In International Conference on Machine Learning (ICML), 2020. 2 9 [26]Tsun-Hsuan Wang, Sivabalan Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, James Tu, and Raquel Urtasun. V2VNet: Vehicle-to-vehicle communication for joint percep- tion and prediction. In European Conference on Computer Vision (ECCV), 2020. 1, 2, 4 [27]Hao Xiang, Runsheng Xu, Xin Xia, Zhaoliang Zheng, Bolei Zhou, and Jiaqi Ma. V2XP-ASG: Generating adversarial scenes for vehicle-to-everything perception. In IEEE Interna- tional Conference on Robotics and Automation (ICRA), 2023. 2 [28]Runsheng Xu, Zhengzhong Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. CoBEVT: Cooperative bird’s eye view semantic segmentation with sparse transformers. In Confer- ence on Robot Learning (CoRL), 2022. 2 [29]Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming- Hsuan Yang, and Jiaqi Ma. V2X-ViT: Vehicle-to-everything cooperative perception with vision transformer. In European Conference on Computer Vision (ECCV), 2022. 1, 2, 3 [30] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. OPV2V: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communica- tion. In IEEE International Conference on Robotics and Automation (ICRA), 2022. 1, 2, 3, 4, 8 [31] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, Hongkai Yu, Bolei Zhou, and Jiaqi Ma. V2V4Real: A real-world large-scale dataset for vehicle-to-vehicle coop- erative perception. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 8 [32] Runsheng Xu, Hao Xiang, Xu Han, Zhengzhong Tu, Xin Xia, and Jiaqi Ma. V2x-pnp: Perception and planning for col- laborative vehicle-to-everything. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023. 1, 2 [33]Chao Yu, Xinyi Yang, Jiaxuan Gao, Jiayu Chen, Yunfei Li, Jijia Liu, Yunfei Xiang, Ruixin Huang, Huazhong Yang, Yi Wu, and Yu Wang. Asynchronous multi-agent reinforcement learning for efficient real-time multi-robot cooperative explo- ration. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), 2023. 1, 2 [34]Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. DAIR-V2X: A large-scale dataset for vehicle-infrastructure cooperative 3D object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2, 8 [35]Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B. Tenenbaum, Tianmin Shu, and Chuang Gan. Building cooperative embodied agents modularly with large language models. In International Conference on Learn- ing Representations (ICLR), 2024. 2 [36]Sheng Zhang, Yifan Li, Jiangfan Chen, and Chen Feng. Bandwidth-adaptive variational message encoding for multi- agent communication. In International Conference on Au- tonomous Agents and Multiagent Systems (AAMAS), 2024. 1, 2, 8 [37]YuxuanZhang,RundongLiu,andYaliWang. Communication-constrainedmulti-agentdeepdeter- ministic policy gradient. In IEEE International Conference on Communications (ICC), 2024. 2, 8 [38] Hao Zhou, Atakan Aral, Ivona Brandic, and Melike Erol- Kantarci. Multi-agent Bayesian deep reinforcement learning for microgrid energy management under communication fail- ures. arXiv preprint arXiv:2111.11868, 2021. 2 10 A. Supplementary Material A.1. Benchmark Configuration Details Table 3 summarizes the full benchmark configuration. Table 3. Full benchmark configuration. ParameterValue Number of agents4 Episodes per condition30 Grid size (all tasks)20× 20 Message dimension400 (flattened grid) Impairment levels per sweep11 Random seeds42–71 (per-episode) Task-specific parameters CP: objects per episode30 CP: steps per episode20 CP: agent FOV90 ◦ quadrant CP: metricF1 score NAV: waypoints per agent3 NAV: steps per episode50 NAV: metricWaypoint completion (%) SEARCH: targets per episode20 SEARCH: zones per agent4 SEARCH: steps per episode50 SEARCH: detectionPoint (exact cell) SEARCH: metricRecall (%) Method-specific parameters RESILIENTCOMM staleness λ0.3 RESILIENTCOMM redundancy 2× (two message copies) Compressed-Comm bit width4-bit (quantize to 15 levels) Event-Triggered thresholdL 1 norm > 0.5 A.2. Communication Load Analysis Table 4 reports the per-step communication load for each method. Table 4. Communication load per step (4 agents, 20×20 grid). MethodMsgs/stepBits/stepRelative No-Comm000× Full-Comm12153,6001× Compressed1219,2000.125× Event-Triggered7–1089,600–128,0000.6–0.8× RESILIENTCOMM24307,2002× A.3. Runtime Reporting The full benchmark (5 methods×3 tasks×6 impairments ×11 severity levels×30 episodes = 29,700 simulation runs) completes in approximately 3.5 minutes on a single CPU core (Apple M-series). Individual episodes run in 1–5 ms depending on the task. This confirms that the benchmark is practical for rapid iteration. A.4. Joint Impairment Analysis In the main paper, we vary one impairment dimension at a time. A natural question is how impairments interact. Com- bining packet loss and bandwidth collapse for the NAV task, we observe that at 100% bandwidth collapse all methods con- verge to the No-Comm floor (3.6%) regardless of packet loss, since no message can be delivered. At moderate bandwidth reduction (50%), packet loss further degrades performance but RESILIENTCOMM’s redundancy provides a measurable advantage. Full joint analysis is left to future work due to the combinatorial explosion of severity combinations. A.5. Gilbert-Elliott Bursty Channel Model The main experiments use i.i.d. Bernoulli packet loss, where each message is independently dropped. A key concern is whether redundant coding’s advantage holds under realis- tic bursty loss. We implement a Gilbert-Elliott two-state Markov channel [4]: a Good state (2% loss) and Bad state (95% loss), with transition probabilities calibrated so the stationary average matches the target loss rate. The bursti- ness parameter controls state persistence (b=0.9means 90% probability of remaining in the current state per step). Table 5 compares NAV waypoint completion under Bernoulli (b=0) and bursty (b=0.9) channels. Table 5. NAV waypoint completion (%) under Bernoulli vs. Gilbert- Elliott bursty channel (30 episodes). RESILIENTCOMM’s redun- dancy advantage is dramatic: at 40% loss, RESILIENTCOMM re- tains 86.1% vs. Full-Comm’s 55.0% under Bernoulli. ChannelMethod0%40%80% BernoulliFull-Comm96.755.010.0 (b=0)RESILIENTCOMM96.786.121.9 BurstyFull-Comm96.758.38.3 (b=0.9)RESILIENTCOMM96.768.317.5 No-Comm (constant)3.6 Under Bernoulli loss, RESILIENTCOMM’s redundancy provides a dramatic advantage: at 40% loss, RESILIENT- COMM retains 86.1% waypoint completion versus Full- Comm’s 55.0%—a 31-point gap. At 80% loss, RESILIENT- COMM achieves 21.9% vs. 10.0%. Under the bursty model (b=0.9), losses cluster into long bursts. Full-Comm degrades slightly more under bursty conditions (8.3% vs. 10.0% at 80% loss), while RESILIENTCOMM’s advantage narrows (17.5% vs. 21.9%) because correlated loss within burst win- dows defeats redundancy—when the channel is in the Bad state, both copies are likely lost. However, RESILIENT- COMM still substantially outperforms Full-Comm at all loss rates under both channel models. We note that our implementation does not model con- gestion feedback: in a shared medium, RESILIENTCOMM’s 11 doubled transmission could increase contention. Modeling this requires a multi-agent MAC-layer simulation, which we leave to future work. A.6. Bandwidth-Normalized Comparison A key concern is that RESILIENTCOMM’s2×communica- tion load creates an unfair comparison. To address this, we enforce a common per-step bit budget. The one-hot way- point message hasd=400elements. Under a1×budget, Full-Comm sends the full 400-dim vector (1 copy), while RESILIENTCOMM must halve to 200-dim×2 copies. Criti- cally, if the waypoint index falls beyond dim 200, the trun- cated message loses the signal entirely. Table 6 reports NAV waypoint completion across budgets. Table 6. NAV waypoint completion (%) under bandwidth con- straints. At1×budget, RESILIENTCOMM’s 200-dim truncation loses waypoints indexed>200, severely hurting performance. At 2× budget, RESILIENTCOMM dominates via redundancy. BudgetMethoddim0%40%80% 1× Full-Comm40096.755.010.0 Compressed40096.755.010.0 RESILIENTCOMM20034.231.911.4 2× Full-Comm40096.755.010.0 Compressed40096.755.010.0 RESILIENTCOMM40096.786.121.9 No-Comm (constant)3.63.63.6 Under equal bandwidth (1×), RESILIENTCOMM suffers severely: truncating to 200 dimensions means roughly half of all waypoints (those with grid indices>200) are invisi- ble, reducing clean performance from 96.7% to just 34.2%. Redundancy cannot compensate for this fundamental infor- mation loss. However, at2×budget—where RESILIENT- COMM can send full 400-dim messages with redundancy—it dramatically outperforms: 86.1% vs. 55.0% at 40% loss and 21.9% vs. 10.0% at 80% loss. This demonstrates a clear tradeoff: redundancy requires sufficient bandwidth to carry the full message payload. When bandwidth is the binding constraint, message fidelity (keeping all 400 dims) dominates. When packet loss is the binding constraint, redundancy (two copies) dominates. An adaptive policy that switches between these strategies based on estimated channel conditions is an important direction for future work. A.7. Staleness Decay Parameter Ablation RESILIENTCOMM’s staleness-aware fusion (Eq. 2) uses a fixed decay parameterλ=0.3. We ablate this choice by test- ingλ ∈ 0.0, 0.1, 0.3, 0.5, 1.0, 2.0on the CP task under stale memory, where fusion weighting matters most. (The NAV task usesargmaxdecoding rather than weighted fu- sion, soλdoes not affect NAV results.) As Table 7 shows,λ has no effect on CP performance under stale memory: the np.maximumfusion is a hard selector that ignores weights, and once stale data introduces false positives, no amount of temporal downweighting can remove them. Table 7. Staleness decay ablation: RESILIENTCOMM CP F1 (%) acrossλvalues under stale memory (30 episodes). The fusion mechanism dominates; λ has no effect. λstale=0stale=4stale=8stale=12stale=20 0.095.814.014.014.014.0 0.395.814.014.014.014.0 1.095.814.014.014.014.0 2.095.814.014.014.014.0 This result underscores a key finding from the main paper: the CP task’s vulnerability to stale data is a property of the fusion mechanism (element-wise maximum), not the staleness weighting. A fusion mechanism that could detect and reject corrupted inputs—e.g., via anomaly detection on received messages [23]—would be needed to address this failure mode. For the NAV task, where RESILIENTCOMM’s advantage comes from redundant message coding rather than staleness fusion, the decay parameter is irrelevant. A.8. RESILIENTCOMM Algorithm Detail Algorithm 1 provides pseudocode for RESILIENTCOMM’s communication loop. A.9. Area Under Robustness Curve (AURC) Table 8 reports the Area Under the Robustness Curve (AURC) for each method and task, averaged across all six impairment dimensions. AURC is computed via trapezoidal integration of the performance curve over normalized sever- ity[0, 1], then expressed as a percentage of the maximum possible area (clean performance×1.0). Higher AURC indi- cates greater robustness. Table 8. AURC (% of max) averaged across 6 impairment dimen- sions. No-Comm achieves 100% trivially (flat curves). RESILIENT- COMM shows highest AURC among communicating methods due to redundancy. MethodCPNAVSearchAvg↑ No-Comm100.0100.0100.0100.0 RESILIENTCOMM (Ours)73.069.884.975.9 Event-Triggered73.063.484.373.5 Full-Comm73.063.383.073.1 Compressed73.063.383.073.1 RESILIENTCOMM achieves the highest AURC among communicating methods (75.9% vs. 73.1% for Full-Comm), 12 Algorithm 1 RESILIENTCOMM Communication Loop (per agent i) 1:Input: knowledge vectork i t ∈R G 2 , staleness decay λ=0.3 2: Maintain: received bufferB, age counterT 3: for each neighbor j ̸= i do 4:Transmit copy 1: ˆm j 1 ← Channel(k i t ) 5:Transmit copy 2: ˆm j 2 ← Channel(k i t ) 6:if ˆm j 1 received then 7: B[j]← ˆm j 1 ;T [j]← 0 8:else if ˆm j 2 received then 9: B[j]← ˆm j 2 ;T [j]← 0 10:else 11: T [j]←T [j] + 1increment staleness 12:end if 13: end for 14: w j ← softmax(−λ·T [j]) for all j Eq. 2 15: ˆs t ← max j (B[j]) weighted byw j staleness-aware fusion 16: k i t ← max(k i t , ˆs t )update knowledge 17: Return k i t for decision-making with the largest advantage in NAV (69.8% vs. 63.3%). The per-impairment breakdown reveals where redundancy helps most: NAV packet loss AURC is 78.3% for RE- SILIENTCOMM vs. 57.1% for Full-Comm, and NAV async AURC is 87.2% vs. 68.9%. The CP AURCs are identical (73.0%) because all methods share the same vulnerability to stale/conflict. No-Comm achieves a trivial 100% AURC (flat curves) by construction. A.10. Sensitivity Analysis Figure 7 presents radar plots showing each method’s sensitiv- ity profile across impairment dimensions. Stale memory and conflicting evidence create the largest vulnerability spikes across all communicating methods, while latency and asyn- chrony have minimal impact in our implementation (heuristic policies tolerate delayed information). This profile would likely differ for learned communication policies [2, 6]. A.11. Impairment Dimension Orthogonality Our six impairment dimensions are designed to capture dis- tinct failure modes, though they are not strictly independent. We distinguish three groups by mechanism: • Transport-layer: Latency (D1), packet loss (D2), and bandwidth collapse (D3) model physical channel con- straints that affect message delivery. •Temporal: Asynchronous updates (D4) and stale memory (D5) both involve temporal misalignment, but differ in mechanism—D4 models clock domain differences (agents act at different times), while D5 models refresh failures (agents operate on outdated partner state even with syn- chronized clocks). •Content: Conflicting sensor evidence (D6) corrupts mes- sage content rather than delivery, representing sensor dis- agreement across heterogeneous platforms. In practice, the NPD vectors across these groups show dis- tinct patterns: D5/D6 cause>85% CP degradation (content corruption) but also>90% NAV degradation (stale way- points are just as destructive as missing ones). Transport impairments (D2, D3) cause>89% NAV degradation (loss of the waypoint signal), while D1 causes a more moderate 32% NAV drop (delayed waypoints still provide useful, if stale, guidance). CP is immune to D1–D4 (0% NPD) because its element-wise maximum fusion naturally tolerates missing or delayed messages. This separation validates treating the dimensions as complementary evaluation axes, though we note that for learned policies, temporal impairments (D1, D4, D5) may interact more strongly. 13 Latency Packet LossBW Reduction Async Offset StalenessConflict 10 20 30 40 50 60 70 80 No-Comm Latency Packet LossBW Reduction Async Offset StalenessConflict 10 20 30 40 50 60 70 80 Full-Comm Latency Packet LossBW Reduction Async Offset StalenessConflict 10 20 30 40 50 60 70 80 Compressed-Comm Latency Packet LossBW Reduction Async Offset StalenessConflict 10 20 30 40 50 60 70 80 Event-Triggered Latency Packet LossBW Reduction Async Offset StalenessConflict 10 20 30 40 50 60 70 80 ResilientComm Figure 7. Sensitivity analysis. Radar plots showing NPD (%) for each method across impairment dimensions. The dominant vulnerability for all communicating methods is to stale memory and conflicting evidence, which corrupt the information being shared. 14