Hierarchical Safe Locomotion for Legged Robots Using Learnable Signed Distance Function and Nonlinear MPC

Hierarchical Safe Locomotion for Legged Robots Using Learnable Signed Distance Function and Nonlinear MPC

TL;DR A hierarchical planner combining a learned SDF, NMPC, and PPO policy for obstacle-aware legged locomotion with improved real-time performance.

A hierarchical motion planning framework for legged robots that integrates a learnable signed distance function (SDF) from a paper for perception, an NMPC module for trajectory optimisation, and a PPO-trained neural network policy for stable gait generation. The learnable SDF is embedded directly in the NMPC cost, allowing the optimiser to reason about obstacle proximity and produce collision-free trajectories. The method was deployed in real time in Isaac Gym.