Learning Adaptive Safety Margins for Visual Navigation

In visual navigation, robots often struggle to balance safety and efficiency. Fixed safety margins—the minimum distance a robot maintains from obstacles—frequently cause robots to either take unnecessary detours in open spaces or attempt risky shortcuts in tight, cluttered areas. This paper addresses the challenge of selecting the best trajectory from a set of diverse candidates generated by a diffusion model, specifically by replacing static, hand-tuned safety rules with a learned, context-aware safety critic.

The authors introduce a context-conditioned safety critic that evaluates candidate trajectories based on three learned components: a safety term (using a clearance-budget penalty and control-barrier-function residuals), an efficiency term (using a smoothness penalty and a safety-gated detour-ratio penalty), and a distance-constraint matching term to prevent margin collapse.

The training process follows a two-stage teacher-student pipeline. First, a teacher model is trained in simulation using privileged geometric information (ESDFs) to learn the optimal safety-efficiency trade-off. Second, this knowledge is distilled into a student selector that operates solely on egocentric RGB-D observations and trajectory tokens, allowing the robot to perform real-time navigation without requiring explicit map building or privileged geometry during deployment.

The proposed method consistently outperforms strong baselines, including diffusion-based planners (NavDP), optimization-based planners (iPlanner), and reinforcement learning agents (PointNav), across both HM3D and MP3D datasets. It achieves the highest success rates and success weighted by path length (SPL), demonstrating superior adaptability to varying levels of scene clutter. Furthermore, the model demonstrates successful zero-shot sim-to-real transfer on a Unitree G1 humanoid, navigating complex indoor environments without requiring task-specific tuning.