Learning Adaptive Safety Margins for Visual Navigation
The paper’s core claim is that visual-navigation robots need learned, context-specific clearance, not one fixed safety buffer.
The authors introduce a safety critic that ranks diffusion-planner trajectory proposals using safety, efficiency, and clearance-matching terms. It is trained with privileged ESDF geometry in simulation, then distilled into a perception-only selector. In HM3D and MP3D PointGoal navigation tests, including cross-dataset transfer, the method reports the best SR and SPL against diffusion, optimization, and RL baselines. The authors also report transfer from simulation to a Unitree G1 humanoid in cluttered indoor scenes without task-specific tuning. ArXiv · AI/CL/LG's note
The authors introduce a safety critic that ranks diffusion-planner trajectory proposals using safety, efficiency, and clearance-matching terms. It is trained with privileged ESDF geometry in simulation, then distilled into a perception-only selector. In HM3D and MP3D PointGoal navigation tests, including cross-dataset transfer, the method reports the best SR and SPL against diffusion, optimization, and RL baselines. The authors also report transfer from simulation to a Unitree G1 humanoid in cluttered indoor scenes without task-specific tuning. ArXiv · AI/CL/LG's note
score 4