Featured project ยท CoRL 2026

GRACE: Group-Aware Socially Compliant Visual Navigation in Human Environments

Group-Aware Social Navigation for robots that move safely and naturally around people.

Lulu Huang, Wenbo Wang, Guo Cheng, Hao Cheng, Jian-Jian Jiang, Wei-Shi Zheng, and Jingke Meng

Read the paper OpenReview

Videos

Selected demonstrations of group-aware and socially compliant navigation.

Navigation demo 1

Robot navigation around people in a shared environment.

Navigation demo 2

Additional real-world rollout with group-aware behavior.

Overview

Visual navigation in a human environment requires more than finding a collision-free path. A robot can avoid a person and still behave intrusively by cutting through the shared space of a conversation or walking between members of a group.

GRACE addresses this problem with a group-aware visual navigation framework. From forward-facing RGB observations and a goal image, it learns to recognize social relationships, anticipate pedestrian motion, and generate trajectories that preserve both personal space and group interaction space.

Comparison of FlowNav and GRACE in human-shared navigation scenarios
Motivation: GRACE avoids socially intrusive paths and preserves personal and group spaces.

How GRACE works

01Social perception
Builds a pedestrian-robot graph and learns soft social relations with F-formation supervision.
02Motion forecasting
Predicts multiple short-horizon pedestrian futures conditioned on group context.
03Social trajectory generation
Uses reward-weighted flow matching to favor goal-reaching paths with fewer social intrusions.

The policy is designed for deployment with RGB input only. It does not require LiDAR, depth sensors, external tracking, or ground-truth group annotations at test time.

Overview of the GRACE group-aware visual navigation framework
Method overview: group-aware social perception conditions trajectory generation, followed by social-reward alignment.

Real-world evaluation

Experiments were conducted on a Unitree Go2 quadruped in unseen indoor and outdoor environments. The evaluation covers sparse and dense scenes with both static groups and moving pedestrians. GRACE runs at approximately 15 Hz with around 51 ms end-to-end latency on an external RTX 4060 workstation.

  • In static-sparse scenes, GRACE reaches an 87.5% success rate with zero collisions and zero group-space intrusions.
  • In dynamic-sparse scenes, it reaches an 83.3% success rate and zero group-space intrusions.
  • In dynamic-dense scenes, it reduces group-space intrusion to 8.33% and pedestrian direction interference to 6.67%.

These results show the value of combining explicit group perception, pedestrian prediction, and social reward alignment for navigation around people.