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

10th Conference on Robot Learning (CoRL 2026)
Zehao Wang1, Jianpeng Yao2, Yuping Wang3, Yiping Li4, Jiaqi Li5, Zhixu Li2, Kanghoon Lee6, Jinfeng Wang2, Cong Liu2, Zhaowei Tan2, Jinkyoo Park6, Jiachen Li1*
1Georgia Institute of Technology    2University of California, Riverside
3University of Michigan    4University of Georgia
5New York University    6KAIST
*Corresponding author
TASL Lab Logo Trustworthy Autonomous Systems Laboratory (TASL)
Code BibTeX
Crossing Pedestrians
Furnished Lobby
Cluttered Space

Adapting to different layouts and pedestrian interactions

Real-robot demonstrations

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.

Motivation

Figure 1: a robot's view of an approaching pedestrian is blocked by an obstacle; another robot shares complementary observations for cooperative forecasting and navigation. The bottom row shows four real-world evaluation environments.

In human-shared spaces, walls and furniture constrain both what a robot can see and where it can maneuver. A pedestrian hidden around a corner may become visible only when little room remains to react. Another robot may already see that pedestrian. Shared observations can help the ego robot anticipate where the pedestrian may cross its path and begin slowing down or changing course before the pedestrian enters its own field of view.

Even with shared observations, future pedestrian motion remains uncertain. Pedestrians can change direction or speed, while unfamiliar layouts and denser crowds can make forecasts less reliable. Navigation must account for this uncertainty when balancing progress toward the goal with caution around potential hazards. CoopUQ connects these needs by fusing cross-robot spatiotemporal features and incorporating predicted occupancy and uncertainty into both the policy input and its training cost. This encourages caution where future occupancy is ambiguous while maintaining efficient progress through well-observed space.

Method

CoopUQ architecture. Robots encode LiDAR-derived occupancy histories and fuse aligned features across robots. Stochastic forecasting produces occupancy and uncertainty maps. A constrained multi-agent policy combines these maps with ego and neighboring robot states, using separate reward and cost critics during training.

CoopUQ links where robots need information to where they should exercise caution. Each robot encodes its LiDAR-derived occupancy history and aligns neighboring features to its own frame. Spatial disagreement between these features provides a cue for a learned fusion gate, encouraging the use of complementary observations in occluded or inconsistent regions while preserving reliable local evidence. A variational autoencoder samples possible future occupancy maps, with pixel-wise entropy used to construct uncertainty maps. Crucially, predicted occupancy and high uncertainty also contribute to an explicit cost constraint in MAPPO-Lagrangian training, allowing the policy to adapt its risk sensitivity under occlusions and distribution shifts while avoiding unnecessary conservatism in well-observed regions.

Cooperative Forecasting in Challenging Scenes

Clean Communication

Communication Delay or Localization Noise

In-Distribution

200 ms Delay

Unseen Layout

Loc. Noise

Denser Crowds

200 ms Delay

Rushing Pedestrians

Loc. Noise

Sustains Reliable Navigation as Robot Teams Grow

  • Robot
  • Goal reached
  • Goal not reached
  • Collision or timeout
  • Pedestrian
  • Social zone

Can Cooperation Improve Forecasts
despite Delays and Noisy Poses?

Occupancy forecasting; mean ± standard deviation
CooperationMethodHorizonstepsSSIM ↑Structural similarityWMSE ↓Weighted mean squared errorIoU ↑Intersection over union
CleanDelay200 msLoc. noiseCleanDelay200 msLoc. noiseCleanDelay200 msLoc. noise
No cooperationSOGMP++10.75 ± 0.00——0.16 ± 0.00——0.60 ± 0.01——
50.71 ± 0.00——0.19 ± 0.00——0.57 ± 0.01——
Early fusionCoopUQ10.86 ± 0.000.79 ± 0.010.74 ± 0.010.07 ± 0.000.13 ± 0.010.15 ± 0.010.73 ± 0.010.65 ± 0.010.58 ± 0.02
50.80 ± 0.000.74 ± 0.010.68 ± 0.020.12 ± 0.000.15 ± 0.010.20 ± 0.010.68 ± 0.010.60 ± 0.020.53 ± 0.02
Late fusionCoopUQ10.81 ± 0.010.78 ± 0.010.76 ± 0.010.12 ± 0.000.13 ± 0.010.15 ± 0.010.63 ± 0.010.61 ± 0.020.59 ± 0.02
50.76 ± 0.010.74 ± 0.010.71 ± 0.020.14 ± 0.010.15 ± 0.010.17 ± 0.010.60 ± 0.010.58 ± 0.020.56 ± 0.02
Middle fusionCoopUQ10.90 ± 0.000.88 ± 0.010.84 ± 0.010.05 ± 0.000.06 ± 0.010.08 ± 0.010.77 ± 0.010.74 ± 0.010.68 ± 0.01
50.86 ± 0.000.84 ± 0.010.80 ± 0.010.09 ± 0.000.10 ± 0.010.12 ± 0.010.72 ± 0.010.72 ± 0.020.64 ± 0.02

What Happens When Layouts
and Pedestrian Behavior Change?

  • Success rate (%)
  • Collision rate (%)
  • Timeout rate (%)

Unseen layouts

No cooperation

Late fusion

Early fusion

Middle fusion

CoopUQ Generalizes to Larger Teams without Retraining

Navigation across robot team sizes; each entry reports ID / OOD
MethodSR ↑Success rateID / OODCR ↓Collision rateTR ↓Timeout rateID / OODNT ↓Nav. time (s)ID / OODPL ↓Path length (m)ID / OODAA ↓Avg. accel.ID / OODITR ↓Intrusion (%)ID / OODSD ↑Social dist. (m)ID / OOD
ObstacleID / OODHumanID / OODRobotID / OOD
ORCA0.47 / 0.370.17 / 0.150.23 / 0.380.06 / 0.080.06 / 0.0210.04 / 10.549.30 / 11.120.10 / 0.117.63 / 11.800.36 / 0.34
SAMARL0.65 / 0.530.10 / 0.140.12 / 0.210.05 / 0.060.07 / 0.0612.98 / 13.4514.99 / 15.570.18 / 0.206.60 / 10.430.37 / 0.34
SCOPE0.67 / 0.560.10 / 0.100.11 / 0.180.03 / 0.050.09 / 0.1113.39 / 14.7915.67 / 16.920.18 / 0.215.56 / 7.560.38 / 0.36
HEIGHT0.70 / 0.600.07 / 0.080.13 / 0.190.03 / 0.050.06 / 0.0712.43 / 13.4913.80 / 14.110.14 / 0.154.08 / 8.020.37 / 0.36
CoopUQ0.90 / 0.820.03 / 0.040.03 / 0.100.00 / 0.010.03 / 0.039.80 / 10.8811.98 / 13.820.13 / 0.172.07 / 4.470.45 / 0.43

All learned policies are trained with three robots and evaluated with five and ten robots without retraining.

Does the Learned Policy Transfer to Real Robots?

  • Success rate (%)
  • Collision rate (%)
  • Timeout rate (%)

SCOPE

CoopUQ

Averaged over four indoor environments (two lobbies and two corridors), four trials per environment and pedestrian count.

BibTeX

Download .bib
@inproceedings{wang2026coopuq,
  title     = {CoopUQ: Robust Multi-Robot Social Navigation via
               Uncertainty-Aware Cooperative Forecasting},
  author    = {Wang, Zehao and Yao, Jianpeng and Wang, Yuping and
               Li, Yiping and Li, Jiaqi and Li, Zhixu and Lee, Kanghoon and
               Wang, Jinfeng and Liu, Cong and Tan, Zhaowei and
               Park, Jinkyoo and Li, Jiachen},
  booktitle = {10th Conference on Robot Learning (CoRL)},
  year      = {2026}
}