Reproduce “Generating Robust Supervision for Learning-Based Visual Navigation Using Hamilton-Jacobi Reachability” on OpenResearch

About this paper

Anjian Li, Somil Bansal, Georgios Giovanis, Varun Tolani, Claire Tomlin, Mo Chen

A framework integrates Hamilton-Jacobi reachability analysis to generate robust training supervision for learning-based visual navigation, explicitly accounting for neural network prediction errors as disturbances. This approach enables a CNN-based system to achieve significantly higher success rates in challenging cluttered environments and demonstrates reliable sim-to-real transfer on a physical robot.

Key insights

• Explicitly modeling neural network prediction errors as disturbances within a formal control framework during training data generation enhances the robustness of learned navigation policies.
• Hamilton-Jacobi reachability analysis provides a theoretically sound method to generate inherently conservative and robust expert supervision, leading to safer navigation behaviors in cluttered spaces.
• A principled integration of learning-based perception with model-based control, informed by formal methods, yields better generalizability and reliability compared to purely end-to-end learning or heuristic-based approaches.

Summary by alphaXiv.