RaFM-Nav demonstration
Risk-aware navigation in cluttered environments.
Risk-aware navigation · IROS 2026
RaFM-Nav: A Risk-Aware Flow Matching Framework for Collision-Free Visual Navigation
A visual navigation framework that makes flow-matching policies more aware of obstacles, risk, and safe open-space traversal.
Liang Li, Wenbo Wang, Jianjian Jiang, Jingke Meng, and Wei-Shi Zheng
Videos
Real-world demonstrations of risk-aware visual navigation.
Outdoor navigation demo
Risk-aware navigation in an outdoor environment.
Cluttered-scene demo
Collision-free navigation through a challenging scene.
Indoor experiment
Indoor real-world experiment with obstacle-aware motion.
Overview
Visual navigation policies learned from human demonstrations can reach goals, but they may still generate unsafe trajectories in cluttered environments. Human demonstrations contain safety cues, yet imitation learning does not always separate those cues from individual preferences and scene-specific conditions.
RaFM-Nav addresses this gap with a risk-aware flow matching framework. It combines explicit obstacle occupancy perception with offline trajectory alignment so that the policy can prefer trajectories with larger clearance and safer open-space routes.
How RaFM-Nav works
01Obstacle Perception Grid
Back-projects estimated monocular depth into a top-down grid with occupancy and height cues.
Back-projects estimated monocular depth into a top-down grid with occupancy and height cues.
02Risk-aware trajectory alignment
Ranks sampled trajectories with collision penalties and open-space rewards during training.
Ranks sampled trajectories with collision penalties and open-space rewards during training.
03Flow matching policy
Generates multimodal waypoint trajectories conditioned on RGB, depth, goal, and grid features.
Generates multimodal waypoint trajectories conditioned on RGB, depth, goal, and grid features.
04Efficient deployment
Keeps trajectory alignment offline, so the safety preference adds no extra inference-time computation.
Keeps trajectory alignment offline, so the safety preference adds no extra inference-time computation.
The safety-oriented reward combines an obstacle collision penalty with an open-space incentive. Group-relative weighting then increases the contribution of safer candidates while retaining the original behavior-cloning objective.
Real-world evaluation
RaFM-Nav was evaluated on a Unitree Go2 quadruped across static indoor clutter, unstructured outdoor terrain, and dynamic indoor scenes with sudden pedestrian interference.
- Goal-conditioned navigation success improves from 30% with FlowNav to 70% with RaFM-Nav.
- Average collisions per run decrease from 2.5 to 1.1 in goal-conditioned navigation.
- Goal-free exploration reaches 1.0 collision per run, compared with 3.5 for FlowNav and 3.6 for NoMaD.
- The policy runs at approximately 10 Hz with an average forward-pass latency of 65 ms.
These results show that explicit spatial risk perception and offline preference alignment can improve safety while preserving efficient visual navigation.