Paper deep dive
Less is More: Robust Zero-Communication 3D Pursuit-Evasion via Representational Parsimony
Jialin Ying, Zhihao Li, Zicheng Dong, Guohua Wu, Yihuan Liao
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%
Last extracted: 3/13/2026, 12:07:27 AM
Summary
The paper introduces a robust, zero-communication 3D pursuit-evasion framework for multi-agent reinforcement learning (MARL) in cluttered environments. By applying 'representational parsimony'—specifically reducing observation dimensionality from 83-D to 50-D by removing team-coupled channels—and implementing 'Contribution-Gated Credit Assignment' (CGCA), the authors demonstrate improved robustness against communication latency, sensing noise, and nonholonomic constraints compared to traditional centralized-critic or fully-coupled MARL methods.
Entities (5)
Relation Signals (4)
OURS-LITE → utilizes → Representational Parsimony
confidence 100% · We instantiate this principle with (i) a parsimonious actor observation interface that removes team-coupled channels
OURS-LITE → utilizes → Contribution-Gated Credit Assignment
confidence 100% · and (ii) Contribution-Gated Credit Assignment (CGCA), a locality-aware credit structure for communication-denied cooperation.
CTDE MAPPO → performsworsethan → OURS-LITE
confidence 95% · OURS-LITE reaches the high-return regime earlier and with lower oscillation amplitude than FULL OBS and CTDE MAPPO.
Representational Parsimony → improves → Robustness
confidence 90% · representational parsimony can improve communication-free coordination.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Asymmetric 3D pursuit-evasion in cluttered voxel environments is difficult under communication latency, partial observability, and nonholonomic maneuver limits. While many MARL methods rely on richer inter-agent coupling or centralized signals, these dependencies can become fragility sources when communication is delayed or noisy. Building on an inherited path-guided decentralized pursuit scaffold, we study a robustness-oriented question: can representational parsimony improve communication-free coordination? We instantiate this principle with (i) a parsimonious actor observation interface that removes team-coupled channels (83-D to 50-D), and (ii) Contribution-Gated Credit Assignment (CGCA), a locality-aware credit structure for communication-denied cooperation. In Stage-5 evaluation (4 pursuers vs. 1 evader), our configuration reaches 0.753 +/- 0.091 success and 0.223 +/- 0.066 collision, outperforming the 83-D FULL OBS counterpart (0.721 +/- 0.071, 0.253 +/- 0.089). It further shows graceful degradation under speed/yaw/noise/delay stress tests and resilient zero-shot transfer on urban-canyon maps (about 61% success at density 0.24). These results support a practical paradigm shift: explicitly severing redundant cross-agent channels can suppress compounding error cascades and improve robustness in latency-prone deployment.
Tags
Links
- Source: https://arxiv.org/abs/2603.08273v1
- Canonical: https://arxiv.org/abs/2603.08273v1
Trouble viewing inline? Open PDF directly →
Full Text
28,841 characters extracted from source content.
Expand or collapse full text
Less is More: Robust Zero-Communication 3D Pursuit-Evasion via Representational Parsimony Jialin Ying, Zhihao Li, Zicheng Dong, Guohua Wu and Yihuan Liao1 *Corresponding author: Yihuan Liao (email: 225189@csu.edu.cn) Abstract Asymmetric 3D pursuit–evasion in cluttered voxel environments is difficult under communication latency, partial observability, and nonholonomic maneuver limits. While many MARL methods rely on richer inter-agent coupling or centralized signals, these dependencies can become fragility sources when communication is delayed or noisy. Building on an inherited path-guided decentralized pursuit scaffold [1], we study a robustness-oriented question: can representational parsimony improve communication-free coordination? We instantiate this principle with (i) a parsimonious actor observation interface that removes team-coupled channels (83-D → 50-D), and (i) Contribution-Gated Credit Assignment (CGCA), a locality-aware credit structure for communication-denied cooperation. In Stage-5 evaluation (4 pursuers vs. 1 evader), our configuration reaches 0.753±0.0910.753± 0.091 success and 0.223±0.0660.223± 0.066 collision, outperforming the 83-D FULL OBS counterpart (0.721±0.0710.721± 0.071, 0.253±0.0890.253± 0.089). It further shows graceful degradation under speed/yaw/noise/delay stress tests and resilient zero-shot transfer on urban-canyon maps (about 61% success at density 0.24). These results support a practical design principle: in communication-constrained multi-robot pursuit, representational sparsity may be more robust than richer inter-agent coupling. I Introduction 3D pursuit–evasion in cluttered voxel spaces is difficult because communication latency, partial observability, and nonholonomic constraints interact with highly non-convex geometry. In the 4-vs-1 setting considered here, pursuers are maneuver-limited (vx≤8v_x\!≤\!8 m/s, |ψ˙|≤0.8| ψ|\!≤\!0.8 rad/s nominally) while the evader reaches 9 m/s in Stage 5, so stale estimates and delayed corrections can destabilize cooperative interception. A common MARL strategy is to increase inter-agent coupling via richer teammate-conditioned observations, explicit communication, or centralized critics. Although helpful in nominal simulation, these couplings may propagate stale peer beliefs under delay/noise. We therefore ask: does decentralized coordination in complex 3D pursuit necessarily require richer cross-agent coupling? Reference [1] introduced the path-guided collaborative search-and-capture scaffold (3D A* guidance + directional frontier allocation). Rather than revisiting that scaffold design, this paper studies a different question: which observation and credit structures remain robust under communication denial, delay, and sensing noise in cluttered 3D pursuit–evasion? We treat the inherited learner as a parameter-shared decentralized PPO/IPPO-style actor–critic (as implemented in [1]), distinct from centralized-critic MAPPO variants. Our core insight is a robustness-oriented design principle: representational parsimony can improve communication-free coordination. Our study highlights two counter-intuitive effects in this regime: (i) more team-coupled information can reduce rather than improve robustness; and (i) communication-free cooperation can remain strong when sparse local observations are paired with locality-aware credit assignment. Our contributions are: • Representational parsimony as a design principle for decentralized MARL in communication-constrained 3D pursuit. We identify that reducing explicit cross-agent observation coupling can improve robustness under delay and sensing noise. • Contribution-Gated Credit Assignment (CGCA). We introduce a locality-aware credit structure that sustains cooperative capture quality without explicit communication channels. • Comprehensive robustness/generalization evidence. We provide benchmark evaluation plus speed/yaw/noise/delay stress suites and zero-shot transfer on procedurally generated urban-canyon maps. The macro-topological planning scaffold itself is retained from [1]; our novelty lies in robustness-oriented information and credit design on that inherited decentralized foundation. I Related Work I-1 A. 3D Pursuit–Evasion Classical pursuit guidance is grounded in analytic controllers such as APF-style reactive potential shaping [12] and proportional navigation (PN) [13]. These methods are lightweight and interpretable, but they are fundamentally local and therefore vulnerable to topological deadlocks in voxelized 3D clutter: the controller can remain trapped in obstacle-induced local minima or oscillatory turn–overshoot cycles. Topology-aware planning (e.g., A*) alleviates some deadlocks by exposing homotopy-level structure [11], yet standalone planning remains brittle under delayed actuation and model mismatch. Recent DRL pipelines improve adaptability, but naive end-to-end formulations frequently face combinatorial explosion in high-dimensional partially observable state spaces [4]. The path-guided decentralized scaffold of [1] (named PGF-MAPPO in that paper) therefore motivates a pragmatic compromise: a planner encodes global geometry, while the policy specializes in local, uncertainty-aware residual control. I-2 B. Communication-Constrained MARL Explicit communication in MARL has been extensively studied through differentiable messaging, attention routing, and graph interaction modules, including CommNet, attentional communication, and targeted multi-agent communication [5, 6, 7, 8]. In simulation, these channels can improve coordination efficiency; however, high-frequency aerial pursuit is precisely the regime where communication quality is non-stationary. Delay, packet dropout, and asynchronous clocks convert message passing into stale belief propagation, which can trigger cross-agent error cascade rather than cooperation gains. This mismatch becomes especially acute when control horizons are short and collision margins are thin. Our zero-communication actor design is thus not an austerity choice; it is a robustness prior that explicitly severs fragile coupling pathways and forces policy learning to rely on local geometry plus shared topological guidance. I-3 C. Credit Assignment in Cooperative MARL Credit assignment remains central in cooperative MARL. Counterfactual baselines such as COMA and value-factorization methods such as VDN/QMIX reduce multi-agent non-stationarity by introducing centralized structure during training [3, 9, 10]. These methods are powerful when global state is reliable, but the dependence on centralized signals can become a practical bottleneck in realistic distributed robotics, where synchronized state aggregation is costly or unavailable. In contrast, CGCA uses only local distance and closing-kinematics signals to shape cooperative incentives. This lightweight locality makes the mechanism naturally compatible with strict decentralized execution and avoids adding another high-dimensional centralized channel on top of already difficult 3D pursuit dynamics. I Problem Formulation and Environment Setup I-A Stage-5 environment The world is discretized into a 52×52×1852× 52× 18 voxel grid (voxel size: 6.0 m) over 311.0×311.0×110.5311.0× 311.0× 110.5 m. Stage-5 settings include 60 m visibility range, 8 m capture radius, and a 70% visibility gate. The traversable occupancy ratio is 0.127, and the largest free connected component covers 0.951 of free voxels, indicating cluttered but globally navigable topology. TABLE I: Stage-5 environment and asymmetry constants. Item Value World/grid 311.0×311.0×110.5311.0× 311.0× 110.5 m, 52×52×1852× 52× 18 voxels Team setup 4 pursuers vs. 1 evader Capture radius (Stage 5) 8.0 m Visibility range (Stage 5) 60.0 m Pursuer bounds |vx|≤8|v_x|≤ 8, |vy|≤4|v_y|≤ 4, |vz|≤3|v_z|≤ 3 m/s Pursuer yaw cap |ψ˙|≤0.8| ψ|≤ 0.8 rad/s (swept down to 0.2) Evader speed schedule up to 9.0 m/s Altitude occupancy anisotropy 15.9% / 12.1% / 10.2% I-B Asymmetric game model Let pursuer i∈1,2,3,4i∈\1,2,3,4\ have position ti∈ℝ3 p_t^i ^3, yaw ψti _t^i, and body-frame control ti=[vx,ti,vy,ti,vz,ti,ψ˙ti]⊤. u_t^i=[v_x,t^i,v_y,t^i,v_z,t^i, ψ_t^i] . (1) Dynamics are t+1i=ti+Δtz(ψti)ti,ψt+1i=ψti+Δtψ˙ti, p_t+1^i= p_t^i+ t\,R_z( _t^i) v_t^i,\; _t+1^i= _t^i+ t\, ψ_t^i, (2) with constraints |vx,ti|≤8,|vy,ti|≤4,|vz,ti|≤3,|ψ˙ti|≤ω¯.|v_x,t^i|≤ 8,\;|v_y,t^i|≤ 4,\;|v_z,t^i|≤ 3,\;| ψ_t^i|≤ ω. (3) Capture is defined by mini‖ti−te‖2≤Rc,Rc=8m. _i\| p_t^i- p_t^e\|_2≤ R_c, R_c=8\,m. (4) The cooperative objective is J(θ)=πθ[∑tγtrt]J(θ)=E_ _θ[ _tγ^tr_t] with dense collision/safety penalties and capture-quality incentives. Figure 1: Training reward curves under Stage 5. CTDE MAPPO shows marked instability after the visibility-gated transition, whereas OURS-LITE converges to the highest stable return regime. IV Methodology IV-A Inherited path-guided decentralized scaffold Following [1], each agent receives a topology-aware 3D A* guidance vector instead of raw path polylines. The planner provides macro free-space structure, while the recurrent policy handles local reactive control. We retain this macro–micro decoupled pursuit scaffold as inherited infrastructure rather than a contribution of the present paper. For method-boundary clarity, we follow the inherited implementation in [1] as a parameter-shared decentralized PPO/IPPO-style actor–critic, and treat centralized-critic MAPPO variants separately. IV-B Observation-space formalization: representational parsimony via 83-D to 50-D masking Let ~it∈ℝ83 o_i^t ^83 denote the full actor-side observation for pursuer i at time t: ~it=[ℓit,Δiet,Δiet,it,it,it,ϕ(it),it,i,mt,δt,it,it], o_i^t= [ _i^t, p_ie^t, v_ie^t, v_i^t, a_i^t, ω_i^t,φ(N_i^t), g_i^t, e_i,m_t, _t, s_i^t, c_i^t ], (5) where ϕ(it)∈ℝ24φ(N_i^t) ^24 encodes top-k teammate states, it∈ℝ7 s_i^t ^7 is the tactical-slot block, and it∈ℝ2 c_i^t ^2 contains encirclement topology cues. To enforce representational parsimony under communication denial, we apply a binary masking operator ∈0,150×83M∈\0,1\^50× 83: it=~it=[~i,1:41t,~i,66:74t]∈ℝ50. o_i^t=M o_i^t= [ o_i,1:41^t, o_i,66:74^t ] ^50. (6) Equivalently, the operator explicitly nulls team-coupled channels, ϕ(it)=,it=,it=,Mφ(N_i^t)= 0, s_i^t= 0, c_i^t= 0, (7) thereby removing 24-D teammate states and 9-D slot/encirclement descriptors. This operation is not introduced as generic feature pruning; it explicitly removes team-coupled channels that are most vulnerable to stale peer estimates, reducing sensitivity to delay/noise while keeping the action space and optimization backbone unchanged. Figure 2: Inherited path-guided decentralized pipeline [1]. Our change is at the actor interface: 83-D full observation is masked to 50-D by removing teammate/slot/encirclement channels for zero-communication robustness. TABLE I: Observation-profile reduction for representational parsimony. Block 83-D 50-D LiDAR + target + self + IMU 41 41 Teammate top-k 24 0 Guidance + ID + mode + delay 9 9 Slot + encirclement 9 0 Total 83 50 IV-C Reward composition and Contribution-Gated Credit Assignment (CGCA) CGCA is designed as a locality-aware credit structure for communication-free cooperation, rather than a generic reward bonus. It rewards geometrically meaningful local participation in interception and suppresses free-rider equilibria when explicit messaging is unavailable. The cooperative objective for parameters θ is J(θ)=πθ[∑t=0Tγt1Nt∑i=1Ntrit],J(θ)=E_ _θ [ _t=0^Tγ^t 1N_t _i=1^N_tr_i^t ], (8) with per-agent reward decomposition rit= r_i^t= λdirgdir(dit)Δdit+λcapρtc^it 1capt+λqualqit _dir\,g_dir(d_i^t)\, d_i^t+ _cap\,ρ^t\, c_i^t\,1_cap^t+ _qual\,q_i^t −λcol 1col,it−λimpκit−λlazy 1lazy,it, - _col\,1_col,i^t- _imp\, _i^t- _lazy\,1_lazy,i^t, (9) where Δdit=dit−1−dit d_i^t=d_i^t-1-d_i^t, dit=‖it−et‖2d_i^t=\| p_i^t- p_e^t\|_2, and κit _i^t denotes collision-impact intensity. To stabilize cooperation under zero communication, CGCA introduces locality-aware directional gating: gdir(d)=1,d≤40m,80−d40,40<d≤80m,0,d>80m.g_dir(d)= cases1,&d≤ 40\,m,\\ 80-d40,&40<d≤ 80\,m,\\ 0,&d>80\,m. cases (10) Capture-share credit is hard-gated outside 60 m and weighted by closing behavior. Define closing speed vi,clot=−d˙itv_i,clo^t=- d_i^t. The raw contribution is c~it=[dit≤60](αpe−dit/d0+αv[vi,clot]++αr[dit≤Rc]), c_i^t=1[d_i^t≤ 60] ( _pe^-d_i^t/d_0+ _v[v_i,clo^t]_++ _r1[d_i^t≤ R_c] ), (11) and the normalized share is c^it=c~it∑j=1Ntc~jt+ϵ. c_i^t= c_i^t _j=1^N_t c_j^t+ε. (12) To suppress free-rider behavior, a participation ratio multiplies global capture reward: ηit=[dit≤60] 1[vi,clot>0.5],ρt=min(1,∑iηit0.5Nt). _i^t=1[d_i^t≤ 60]\,1[v_i,clo^t>0.5], ρ^t= \! (1, _i _i^t0.5N_t ). (13) Thus, if fewer than half of alive pursuers actively close the target, collective capture bonus is proportionally downscaled. Collision suppression is explicitly encoded as col,it=obs−hit,it+team−hit,it+shield,it,1_col,i^t=1_obs-hit,i^t+1_team-hit,i^t+1_shield,i^t, (14) which penalizes obstacle impacts, inter-agent contacts, and safety-shield triggers in a unified term. The 40/60/80 m thresholds are chosen to align with the 60 m sensing extremum in Stage 5 and the practical maneuver envelope of the platform: 40 m acts as high-confidence local-interaction core, 60 m matches reliable observability scale, and 80 m suppresses non-contributive far-field reward leakage. Figure 3: Contribution-Gated Credit Assignment. Direction reward is fully active within 40 m, decays over 40–80 m, and vanishes beyond 80 m; capture-share is hard-gated beyond 60 m. This local credit geometry suppresses free-rider equilibria in zero-communication team pursuit. V Experiments and Evaluations V-A Setup and baselines Evaluation follows the provided Stage-5 protocol (main benchmark: 3 seeds, 500 episodes/seed; robustness suites: 200 episodes/setting/seed, with delay using 4–5 evaluation seeds). Baselines include FULL OBS, EUCLIDEAN, LOCAL-NO-GATE, CTDE MAPPO, and APF+PN. Comparisons are intentionally factorized. OURS-LITE, FULL OBS, and LOCAL-NO-GATE belong to the same inherited parameter-shared decentralized PPO family; OURS-LITE vs. FULL OBS isolates information-structure effects (50-D parsimonious vs. 83-D team-coupled), and OURS-LITE vs. LOCAL-NO-GATE isolates the role of CGCA under the same 50-D setting. CTDE MAPPO is treated separately as a centralized-critic baseline to contrast centralized coupling against decentralized execution in this task. V-B Curriculum maturation and visibility gating To clarify where search capability originates, we adopt a stage-wise curriculum with explicit visibility gating. During training, target observability is conditionally released only when either the explored-space ratio exceeds a stage-specific threshold τ or a hard episode timeout is reached (>1500>1500 steps in the late-stage setting). The threshold τ is progressively increased with curriculum maturation: Stage 1 has no gating, while later blind-search stages tighten from 45% (Stage 3) to 60% (Stage 4) and 70% (Stage 5). This mechanism creates a structural bottleneck: agents must first internalize 3D exploration before reliable tracking signals become available. In other words, observability is earned through coverage rather than granted a priori, which prevents early lucky sightings from dominating learning. Critically, this gate is only active during training as an inducement for tactical emergence; it is fully removed during evaluation, where policies are tested without artificial visibility constraints. V-C Training dynamics and sample efficiency Fig. 1 shows that OURS-LITE reaches the high-return regime earlier and with lower oscillation amplitude than FULL OBS and CTDE MAPPO. A key reason is representation-level compression: removing 33 dimensions of delay-sensitive team-coupling features shrinks the actor input manifold and reduces nuisance covariance under delayed observations. In PPO, the policy-gradient estimator ^t=∇θlogπθ(ait∣it)A^it g_t= _θ _θ(a_i^t o_i^t)\, A_i^t (15) can be interpreted through a policy-input variance lens: fewer delay-sensitive channels reduce nuisance covariance, which is consistent with empirically more stable policy updates and improved sample efficiency in our curriculum. In contrast, CTDE MAPPO shows less stable value-learning dynamics in this setting: the 338-D centralized critic may increase approximation burden and potentially amplify distribution-shift sensitivity once visibility gates and clutter-induced mode switches become active. Importantly, the middle panel of Fig. 1 (Clean Win Rate) shows that CTDE MAPPO is not weak from the outset: in simpler curriculum phases (Stage 1–2), it stays above 90% for a continuous 2M-step interval before Stage 3. Instability emerges mainly at the Stage-3 transition, where severe 3D non-convex occlusions and visibility gating are introduced. These observations suggest that centralized critics may become fragile when high-dimensional value approximation and abrupt visibility-induced distribution shift occur simultaneously. Why may CTDE MAPPO become unstable here? The centralized critic consumes 338-D input, which may intensify approximation difficulty in highly non-convex 3D clutter. Meanwhile, visibility-gated phase transitions induce abrupt observation-distribution shifts, potentially amplifying critic-target variance and gradient jitter. The combined evidence is consistent with unstable value learning and weaker policy improvement despite centralized conditioning. V-D Main benchmark and ablation study TABLE I: Main Stage-5 benchmark (mean± over 3 seeds, 500 episodes/seed). OURS-LITE, FULL OBS, and LOCAL-NO-GATE are inherited parameter-shared decentralized PPO-family variants. Method Success Clean Collision Avg Steps OURS-LITE 0.753± 0.091 0.752± 0.091 0.223± 0.066 683.6± 269.4 FULL OBS 0.721± 0.071 0.721± 0.072 0.253± 0.089 655.4± 106.7 EUCLIDEAN 0.586± 0.120 0.583± 0.121 0.353± 0.092 887.9± 237.7 LOCAL-NO-GATE 0.569± 0.090 0.568± 0.089 0.389± 0.106 669.1± 223.2 CTDE MAPPO 0.006± 0.004 0.006± 0.004 0.159± 0.111 2806.8± 139.1 APF+PN 0.125± 0.007 0.125± 0.007 0.875± 0.007 124.9± 2.0 Within the decentralized parameter-shared PPO family, OURS-LITE vs. FULL OBS isolates the effect of information structure: removing explicit teammate-coupled channels improves success (0.753 vs. 0.721) and reduces collisions (0.223 vs. 0.253). This supports the hypothesis that richer coupling is not always more robust under delayed/noisy pursuit. Ablation insight (OURS-LITE vs. LOCAL-NO-GATE). Removing CGCA under the same 50-D zero-communication setting causes an 18.4-point success drop (0.753 → 0.569) and a 16.6-point collision increase (0.223 → 0.389). This isolates CGCA as a necessary mechanism rather than a cosmetic regularizer. V-E Kinematic stress tests Fig. 4 shows speed sweep (7.0–10.0 m/s), where OURS-LITE degrades gracefully from 0.907 to 0.790 success. EUCLIDEAN exhibits the classic kinematic trap: straight-line chasing delays feasible turn commitment in clutter, causing overshoot-recovery loops and inflated episode length. Figure 4: Velocity sweep. OURS-LITE maintains the strongest success–collision trade-off under increasing evader speed, while EUCLIDEAN suffers persistent high-step inefficient pursuits. Fig. 5 reports yaw-cap restriction (0.8 → 0.2 rad/s). Under severe turning limits, performance drop reflects algorithm–physics mismatch: reactive corrections arrive too late when feasible arc space collapses. Figure 5: Yaw-rate sweep. The proposed configuration preserves a robust mid-regime margin and shows graceful degradation as maneuver authority shrinks. V-F Sim-to-real robustness Noise and delay robustness are shown in Fig. 6 and Fig. 7. Under delay, FULL OBS degrades more sharply than OURS-LITE, which is consistent with state de-synchronization: densely coupled teammate channels may propagate stale estimates and potentially amplify control-error cascades. Figure 6: Observation-noise sweep (σ∈[0,0.20]σ∈[0,0.20]). OURS-LITE remains strongest across all tested noise levels. Figure 7: Action-delay sweep (0–3 steps). OURS-LITE preserves the best absolute robustness profile under latency perturbation. V-G Qualitative analysis: cross-layer vertical sweep Figure 8: 3D Z-axis tactical slicing. The team performs cross-layer vertical sweep: upper-layer pressure constrains vertical escape while lower-layer agents close horizontal corridors, yielding a communication-free topological enclosure. Figure 9: Zero-shot transfer on five unseen Urban Canyon maps (density 0.08–0.24). Left: map examples from easy to hard density. Right: seed-wise and mean success, collision, average steps, and reward without fine-tuning. OURS-LITE degrades gracefully with density, remaining near 61% success and below 40% collision at density 0.24, consistent with structural transfer rather than fixed-map memorization. Fig. 8 reveals a three-phase maneuver pattern rather than a single reactive chase. Initial Search: guided by the shared 3D A* prior, agents fan out to frontier-consistent corridors while preserving safety spacing. Altitude Stratification: once target likelihood concentrates, one pursuer climbs to the upper low-occupancy band and the remaining agents occupy mid/low bands, exploiting measured anisotropy (15.9%/12.1%/10.2%) to reduce vertical escape options. Topological Containment: the formation then uses obstacle morphology—especially U-shaped walls and corner pockets—as virtual teammates, tightening a cross-layer envelope until terminal capture. This phased behavior explains why communication-free coordination remains effective: agents do not exchange explicit messages, yet shared topology priors and CGCA-induced local incentives synchronize role allocation over time. Importantly, the enclosure trajectory is geometrically consistent with the map’s layered free-space structure, indicating true 3D tactical organization rather than incidental altitude jitter. VI Zero-Shot Generalization on Procedurally Generated Urban Canyons To test map over-fitting, we evaluate OURS-LITE zero-shot on five procedurally generated Urban Canyon maps with obstacle densities 0.08–0.24. The generator uses connected 2.5D anisotropic street-corridor layouts (rather than random voxel noise), yielding physically meaningful maneuver corridors and controlled topology shift. As shown in Fig. 9, performance degrades gracefully with density: at the hardest map (0.24), success remains around 61% while collision stays below 40%, and average steps remain in the 350–450 band. These trends are consistent with structural transfer (cross-layer containment under vertical anisotropy) rather than fixed-map memorization. VII Discussion and Limitations The empirical picture supports a robustness-oriented design principle: in communication-constrained multi-robot pursuit, less cross-agent coupling can sometimes yield more reliable coordination. In our setting, the 83-D team-coupled profile is outperformed by the 50-D parsimonious profile, and cooperation quality is maintained by locality-aware CGCA without explicit communication. This combination is consistent with the view that stale coupling channels may introduce harmful variance under delay/noise, whereas sparse local signals can remain controllable. Two limitations remain. First, yaw-restriction sweeps currently do not include FULL OBS, so low-yaw cross-method comparisons are incomplete. Second, the present noise suite is bounded at σ≤0.20σ≤ 0.20; stronger perturbation regimes are needed for firmer conclusions under extreme sensing corruption. VIII Conclusion Built on the inherited path-guided decentralized scaffold of [1], this paper studies a different question from framework design: robustness-oriented information and credit structure in communication-constrained 3D pursuit–evasion. The results support two linked findings: (i) richer team-coupled observations are not always better, and representational parsimony can improve robustness; (i) communication-free coordination can remain effective when local credit assignment is properly structured through CGCA. Across benchmark, stress, and zero-shot urban-canyon transfer evaluations, the evidence indicates a practical design principle rather than a universal rule: in communication-constrained multi-robot pursuit, reducing explicit cross-agent coupling can sometimes yield more robust coordination. References [1] J. Ying, Z. Li, Z. Dong, G. Wu, and Y. Liao, “Generalizable Collaborative Search-and-Capture in Cluttered Environments via Path-Guided MAPPO and Directional Frontier Allocation,” arXiv preprint arXiv:2512.09410, 2025. [2] R. Lowe et al., “Multi-agent actor-critic for mixed cooperative-competitive environments,” in Advances in Neural Information Processing Systems, 2017. [3] J. Foerster, G. Farquhar, T. Afouras, N. Nardelli, and S. Whiteson, “Counterfactual multi-agent policy gradients,” in AAAI Conference on Artificial Intelligence, 2018. [4] C. Yu et al., “The surprising effectiveness of PPO in cooperative multi-agent games,” in NeurIPS Datasets and Benchmarks, 2022. [5] S. Sukhbaatar, A. Szlam, and R. Fergus, “Learning multiagent communication with backpropagation,” in Advances in Neural Information Processing Systems, 2016. [6] J. Jiang and Z. Lu, “Learning attentional communication for multi-agent cooperation,” in Advances in Neural Information Processing Systems, 2018. [7] A. Das, T. Gervet, J. Romoff, D. Bouchard, L. V. Belanger, and J. Pineau, “TarMAC: Targeted multi-agent communication,” in International Conference on Machine Learning, 2019. [8] Y. Jiang, Y. Guo, and S. Li, “Graph convolutional reinforcement learning for multi-agent cooperation,” in International Conference on Learning Representations Workshop, 2020. [9] P. Sunehag et al., “Value-decomposition networks for cooperative multi-agent learning,” in International Conference on Autonomous Agents and Multiagent Systems, 2018. [10] T. Rashid et al., “QMIX: Monotonic value function factorisation for deep multi-agent reinforcement learning,” in International Conference on Machine Learning, 2018. [11] P. E. Hart, N. J. Nilsson, and B. Raphael, “A formal basis for the heuristic determination of minimum cost paths,” IEEE Transactions on Systems Science and Cybernetics, vol. 4, no. 2, p. 100–107, 1968. [12] O. Khatib, “Real-time obstacle avoidance for manipulators and mobile robots,” The International Journal of Robotics Research, vol. 5, no. 1, p. 90–98, 1986. [13] P. Zarchan, Tactical and Strategic Missile Guidance, 6th ed. Reston, VA, USA: AIAA, 2012. [14] F. Sadeghi and S. Levine, “CAD2RL: Real single-image flight without a single real image,” in Robotics: Science and Systems, 2017. [15] L. Busoniu, R. Babuska, and B. De Schutter, “A comprehensive survey of multiagent reinforcement learning,” IEEE Transactions on Systems, Man, and Cybernetics, Part C, vol. 38, no. 2, p. 156–172, 2008.