CoopUQ: Robust Multi-Robot Social Navigation via Uncertainty-Aware Cooperative Forecasting

Zehao Wang; Jianpeng Yao; Yuping Wang; Yiping Li; Jiaqi Li; Zhixu Li; Kanghoon Lee; Jinfeng Wang

cs.ROOct 19, 2026

CoopUQ: Robust Multi-Robot Social Navigation via Uncertainty-Aware Cooperative Forecasting

Authors: Zehao Wang, Jianpeng Yao, Yuping Wang, Yiping Li, Jiaqi Li, Zhixu Li, Kanghoon Lee, Jinfeng Wang,

Organizations: Georgia Institute of Technology · University of California, Riverside · University of Michigan, Ann Arbor · University of Georgia ·

Abstract

Reinforcement learning enables mobile robots to make fast and scalable navigation decisions in constrained human-shared environments. However, directly extending single-robot approaches to multi-robot settings often fails to ensure safety and efficiency, especially under occlusions and out-of-distribution (OOD) scenarios, where individual robots may overlook potential risks or make inaccurate predictions. Leveraging cooperation effectively for both prediction and navigation remains a major challenge. We propose CoopUQ, a multi-robot social navigation framework that bridges uncertainty-aware cooperative occupancy forecasting and robust decision making. CoopUQ predicts future occupancy grid maps through cross-robot spatiotemporal feature fusion and further leverages forecasts and their uncertainty to guide policy via constrained multi-agent reinforcement learning, where the cost penalizes navigation through regions predicted to be occupied or highly uncertain. This design enables the policy to adapt its risk sensitivity under distribution shifts. Experimental results across multiple Unity3D-based simulation environments with both in-distribution and OOD settings, including diverse layouts, pedestrian densities, and rushing behavior, demonstrate improved prediction performance and safety over representative baselines. Real-robot experiments further demonstrate its robustness across various scenes. Our code and videos are available at https://coopuq.github.io/.

Figures & tables

Figure 1: The single-agent approach relies only on the robot’s own observations and misses an incoming pedestrian behind the obstacle. In contrast, CoopUQ fuses shared information, anticipates dynamic motions under uncertainty, and plans a safe trajectory. Closed-loop experiments in Unity-3D simulation demonstrate robustness in OOD scenarios, and the learned policy transfers directly to various real-world environments, achieving safe navigation across diverse crowds and layouts.

Figure 2: The overall architecture of the CoopUQ for multi-robot social navigation.

Methods	Cooperation	Horizon	SSIM $\uparrow$			WMSE $\downarrow$			IoU $\uparrow$
Methods	Cooperation	Horizon	Clean	200ms Delay	Loc. Noise	Clean	200ms Delay	Loc. Noise	Clean	200ms Delay	Loc. Noise
SOGMP++ [ 44 ]	No	1	0.75 ±0.00	–	–	0.16 ±0.00	–	–	0.60 ±0.01	–	–
SOGMP++ [ 44 ]	No	5	0.71 ±0.00	–	–	0.19 ±0.00	–	–	0.57 ±0.01	–	–
Ours	Early Fusion	1	0.86 ±0.00	0.79 ±0.01	0.74 ±0.01	0.07 ±0.00	0.13 ±0.01	0.15 ±0.01	0.73 ±0.01	0.65 ±0.01	0.58 ±0.02
	Early Fusion	5	0.80 ±0.00	0.74 ±0.01	0.68 ±0.02	0.12 ±0.00	0.15 ±0.01	0.20 ±0.01	0.68 ±0.01	0.60 ±0.02	0.53 ±0.02
	Late Fusion	1	0.81 ±0.01	0.78 ±0.01	0.76 ±0.01	0.12 ±0.00	0.13 ±0.01	0.15 ±0.01	0.63 ±0.01	0.61 ±0.02	0.59 ±0.02

Table 1: Comparison of OGM forecasting under different cooperation settings.

Methods	Coop.	SR $\uparrow$	CR $\downarrow$			TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	AA $\downarrow$	ITR (%) $\downarrow$	SD $\uparrow$
Methods			Obstacle	Human	Robots
NavMesh	No	0.39 ±0.05	0.00 ±0.00	0.41 ±0.06	0.19 ±0.04	0.00 ±0.00	12.20 ±0.53	13.89 ±0.62	0.10 ±0.05	12.93 ±1.79	0.29 ±0.03
ORCA [ 38 ]		0.47 ±0.04	0.17 ±0.02	0.23 ±0.03	0.06 ±0.01	0.06 ±0.02	10.04 ±0.42	9.30 ±0.35	0.10 ±0.02	7.63 ±0.48	0.36 ±0.03
SAMARL [ 41 ]		0.65 ±0.03	0.10 ±0.02	0.12 ±0.02	0.05 ±0.01	0.07 ±0.02	12.98 ±0.46	14.99 ±0.38	0.18 ±0.03	6.60 ±0.42	0.37 ±0.03
SCOPE [ 45 ]		0.67 ±0.03	0.10 ±0.02	0.11 ±0.01	0.03 ±0.01	0.09 ±0.02	13.39 ±0.65	15.67 ±0.20	0.18 ±0.02	5.56 ±0.39	0.38 ±0.04
HEIGHT [ 25 ]		0.70 ±0.03	0.07 ±0.01	0.13 ±0.02	0.03 ±0.01	0.06 ±0.01	12.43 ±0.40	13.80 ±0.32	0.14 ±0.02	4.08 ±0.31	0.37 ±0.03

Table 2: In-distribution navigation performance with three robots.

Figure 3: OOD test results in (a) unseen layouts, (b) denser crowds, and (c) rushing pedestrians.

Figure 4: (a) Prediction w/o cooperation: The red robot reacts only after the human fully emerges into its view, resulting in a sharp brake and collision. (b) Prediction w/ cooperation: The red robot decelerates and changes direction earlier to avoid a collision. (c) Key frames of navigation w/o and w/ cooperation. (d) Prediction and uncertainty map from ours at step $t$ . The red ego robot proactively avoids entering high-uncertainty areas due to cost constraints, thus preventing potential collisions.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: We evaluate CoopUQ in four indoor environments with one to three pedestrians walking in the scene. In total, 48 trials are conducted (4 per environment and pedestrian count). Across all runs, CoopUQ achieves an average success rate of 0.94. The success rate with one, two, and three pedestrians is 1.0, 0.94, and 0.88, respectively.

Figure 6: The upper orange robot navigates toward its goal while detouring early to avoid colliding with a pedestrian emerging from occlusion. The demo is available in supplementary videos.

Method	Environment	Pedestrians Count
		1			2			3
		Success $\uparrow$	Collision $\downarrow$	Timeout	Success $\uparrow$	Collision $\downarrow$	Timeout	Success $\uparrow$	Collision $\downarrow$	Timeout
SCOPE [ 45 ]	Lobby-A	1.00	0.00	0.00	0.75	0.25	0.00	0.50	0.25	0.25
	Lobby-B	0.75	0.00	0.25	0.50	0.25	0.25	0.25	0.50	0.25
	Corridor-A	0.75	0.25	0.00	0.50	0.50	0.00	0.25	0.50	0.25
	Corridor-B	0.75	0.25	0.00	0.50	0.50	0.00	0.50	0.50	0.00

Appendix

Table 3: Real-world navigation performance across different environments and pedestrian densities. Results are averaged over 4 trials for each environment and pedestrian setting.

Robots	Method	SR $\uparrow$	CR $\downarrow$			TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	AA $\downarrow$	ITR (%) $\downarrow$	SD $\uparrow$
Robots	Method	SR $\uparrow$	Obstacle	Human	Robots	TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	AA $\downarrow$	ITR (%) $\downarrow$	SD $\uparrow$
Three	ORCA [ 38 ]	0.47/0.37	0.17/0.15	0.23/0.38	0.06/0.08	0.06/ 0.02	10.04/ 10.54	9.30 / 11.12	0.10/0.11	7.63/11.80	0.36/0.34
	SAMARL [ 41 ]	0.65/0.53	0.10/0.14	0.12/0.21	0.05/0.06	0.07/0.06	12.98/13.45	14.99/15.57	0.18/0.20	6.60/10.43	0.37/0.34
	SCOPE [ 45 ]	0.67/0.56	0.10/0.10	0.11/0.18	0.03/0.05	0.09/0.11	13.39/14.79	15.67/16.92	0.18/0.21	5.56/7.56	0.38/0.36
	HEIGHT [ 25 ]	0.70/0.60	0.07/0.08	0.13/0.19	0.03/0.05	0.06/0.07	12.43/13.49	13.80/14.11	0.14/0.15	4.08/8.02	0.37/0.36
	CoopUQ	0.90 / 0.82	0.03 / 0.04	0.03 / 0.10	0.00 / 0.01	0.03 /0.03	9.80 /10.88	11.98/13.82	0.13/0.17	2.07 / 4.47	0.45 / 0.43

Appendix

Table 4: Navigation performance with three, five and ten robots. Values are reported as ID/OOD. Three-robot OOD results are averaged equally over the three scenarios. All learned policies are trained with three robots and evaluated zero-shot with five and ten robots.

	1-step			5-step
Method	SSIM $\uparrow$	WMSE $\downarrow$	IoU $\uparrow$	SSIM $\uparrow$	WMSE $\downarrow$	IoU $\uparrow$
No Cooperation	$0.75\pm 0.00$	$0.16\pm 0.00$	$0.60\pm 0.01$	$0.71\pm 0.00$	$0.19\pm 0.00$	$0.57\pm 0.01$
Early Fusion	$0.86\pm 0.00$	$0.07\pm 0.00$	$0.73\pm 0.01$	$0.80\pm 0.00$	$0.12\pm 0.00$	$0.68\pm 0.01$
Late Fusion	$0.81\pm 0.01$	$0.12\pm 0.00$	$0.63\pm 0.01$	$0.76\pm 0.01$	$0.14\pm 0.01$	$0.60\pm 0.01$
Middle Fusion w/o Deviation-guided	$0.87\pm 0.01$	$0.08\pm 0.01$	$0.74\pm 0.01$	$0.83\pm 0.01$	$0.11\pm 0.01$	$0.71\pm 0.01$
Middle Fusion w/ Deviation-guided	$\mathbf{0.90\pm 0.00}$	$\mathbf{0.05\pm 0.00}$	$\mathbf{0.77\pm 0.01}$	$\mathbf{0.86\pm 0.00}$	$\mathbf{0.09\pm 0.00}$	$\mathbf{0.72\pm 0.01}$

Appendix

Table 5: Ablation study on deviation-guided cross-robot spatiotemporal fusion for OGM forecasting.

Figure 7: Training reward comparison with and without CMARL. The CMARL policy achieves faster convergence and higher final rewards, indicating that cost constraints do not merely improve safety but also facilitate efficient policy learning.

Cooperation	Policy	SR $\uparrow$	Obstacle CR $\downarrow$	Human CR $\downarrow$	Robots CR $\downarrow$	TR $\downarrow$	ITR (%) $\downarrow$	SD $\uparrow$
No Cooperation	Prediction	$0.71\pm 0.01$	$0.12\pm 0.02$	$0.13\pm 0.01$	$0.02\pm 0.00$	$\mathbf{0.01\pm 0.00}$	$5.78\pm 0.27$	$0.34\pm 0.03$
	Prediction & Uncertainty	$0.72\pm 0.01$	$0.11\pm 0.02$	$0.12\pm 0.01$	$0.02\pm 0.01$	$0.03\pm 0.00$	$5.26\pm 0.37$	$0.35\pm 0.01$
	CMARL	$0.75\pm 0.02$	$0.10\pm 0.02$	$0.10\pm 0.03$	$0.01\pm 0.00$	$0.04\pm 0.02$	$4.17\pm 0.26$	$0.39\pm 0.02$
Late Fusion	Prediction	$0.78\pm 0.03$	$0.08\pm 0.02$	$0.10\pm 0.02$	$0.01\pm 0.01$	$0.02\pm 0.02$	$3.96\pm 0.19$	$0.40\pm 0.01$
	Prediction & Uncertainty	$0.79\pm 0.03$	$0.08\pm 0.01$	$0.09\pm 0.02$	$0.01\pm 0.01$	$0.02\pm 0.03$	$3.86\pm 0.29$	$0.40\pm 0.02$
	CMARL	$0.81\pm 0.03$	$0.07\pm 0.02$	$0.09\pm 0.02$	$0.01\pm 0.01$	$\mathbf{0.01\pm 0.02}$	$3.37\pm 0.16$	$0.42\pm 0.02$

Appendix

Table 6: Ablation study on CMARL under in-distribution navigation.

Cost Limit $\epsilon$	SR $\uparrow$	Obstacle CR $\downarrow$	Human CR $\downarrow$	Robots CR $\downarrow$	TR $\downarrow$	NT $\downarrow$	PL $\downarrow$	AA $\downarrow$	ITR (%) $\downarrow$	SD $\uparrow$
$0.20$	$0.87\pm 0.03$	$0.05\pm 0.02$	$0.04\pm 0.01$	$0.01\pm 0.01$	$\mathbf{0.02\pm 0.01}$	$\mathbf{9.37\pm 0.35}$	$\mathbf{11.35\pm 0.17}$	$0.13\pm 0.03$	$2.19\pm 0.31$	$0.43\pm 0.03$
$0.15$	$0.89\pm 0.03$	$0.04\pm 0.01$	$\mathbf{0.03\pm 0.01}$	$\mathbf{0.00\pm 0.01}$	$\mathbf{0.02\pm 0.01}$	$9.46\pm 0.26$	$11.56\pm 0.19$	$0.13\pm 0.04$	$2.15\pm 0.23$	$0.43\pm 0.03$
$0.10$	$\mathbf{0.90\pm 0.03}$	$\mathbf{0.03\pm 0.01}$	$\mathbf{0.03\pm 0.01}$	$\mathbf{0.00\pm 0.00}$	$0.03\pm 0.01$	$9.80\pm 0.15$	$11.98\pm 0.29$	$0.13\pm 0.05$	$\mathbf{2.07\pm 0.19}$	$\mathbf{0.45\pm 0.03}$

Appendix

Table 7: Ablation study on the cost limit

\epsilon

in MAPPO-Lagrangian. A smaller cost limit encourages socially compliant behavior, while a higher cost limit allows more efficient goal-reaching behavior.

Social Robot Navigation

Social robot navigation studies how robots move safely and comfortably around people while respecting human behavior, preferences, and shared-space norms.

See all 81 papers

TRACER: Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning
Sep 16, 2026Lan Hu, Minghui Liwang, Wenbo Zhu +4Multi-Robot Motion Planning Multi-Robot Systems
Guohao School, Tongji University, Shanghai, China · Shanghai Research Institute for Intelligent Autonomous Systems, the State Key Laboratory of Autonomous Intelligent Unmanned Systems, Department of Control Science and Engineering, Tongji University, Shanghai, China
Optimal-Horizon Social Robot Navigation in Heterogeneous Crowds
Feb 28, 2026Jiamin Shi, Haolin Zhang, Yuchen Yan +3Social Robot Navigation Safe Robot Navigation
National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Nation Engineering Research Center for Visual Information and Applications, and Institute of Artificial Intelligence and Robotics, Xi’an Jiaotong University, Xi’an, Shaanxi 710049, P.R. China
NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation
Mar 16, 2026Tianshuai Hu, Zeying Gong, Lingdong Kong +7Social Robot Navigation World Models for Robotics
The Hong Kong University of Science and Technology · The Hong Kong University of Science and Technology (Guangzhou) · National University of Singapore +1

Sep 16, 2026J/K move · Enter open · S save

TRACER: Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning

Lan Hu, Minghui Liwang, Wenbo Zhu, Xinlei Yi+3

Guohao School, Tongji University, Shanghai, China · Shanghai Research Institute for Intelligent Autonomous Systems, the State Key Laboratory of Autonomous Intelligent Unmanned Systems, Department of Control Science and Engineering, Tongji University, Shanghai, China

Multi-Robot Motion Planning Multi-Robot Systems Multi-Agent Trajectory Prediction Safe Robot Navigation Social Navigation

Open paper AskPDF arXiv

Multi-robot navigation in human-shared spaces is inherently interactive: coordinated robot motions influence how nearby entities respond, while those responses provide valuable information for subsequent robot decisions. However, existing methods typically address action-conditioned prediction, multi-robot planning, or online adaptation separately, and therefore lack a unified mechanism for modeling joint robot-entity interactions and adapting future decisions from executed interaction outcomes. To address this gap, we propose TRACER, a bi-directional receding-horizon framework that closes the loop between prediction and adaptation. TRACER evaluates candidate (i.e., alternative feasible future motion plans for the robot team) trajectories using a per-entity probabilistic response model that separates individual-robot effects from non-additive pairwise interactions; after executing the selected trajectory prefix, it updates persistent identity-bound beliefs over latent response modes using the synchronized observed responses. These updated beliefs then guide subsequent candidate evaluation under probabilistic safety and response-aware cost criteria. Experiments show that (i) TRACER more accurately captures non-additive multi-robot interaction effects than a capacity-matched additive predictor, (ii) persistent identity-consistent evidence improves response prediction and downstream replanning, and (iii) the complete TRACER framework improves collision-free completion over an independent-robot baseline on the SocialGym2 multi-robot social-navigation benchmark.

Abstract

Social Robot Navigation

Explore similar work

TRACER: Adaptive Multi-Robot Social Navigation via Joint Human-Response Prediction and Interaction-Aware Replanning

Optimal-Horizon Social Robot Navigation in Heterogeneous Crowds

NavThinker: Action-Conditioned World Models for Coupled Prediction and Planning in Social Navigation