Energy-Based Legged Robots Terrain Traversability Modeling via Deep Inverse Reinforcement Learning

This work reports ondeveloping a deep inverse reinforcement learning method for legged robots terrain traversability modeling that incorporates both exteroceptive and proprioceptive sensory data. Existing works use robot-agnostic environmental features or handcrafted kinematic features; instead, we propose to also learn robot-specific inertial from data reward approximation in single neural network. Incorporating the can improve model fidelity provide depends on robot’s state during deployment. We train network using Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) algorithm simultaneously minimizing trajectory ranking loss deal with suboptimality of robot demonstrations. The demonstrated trajectories are ranked by locomotion energy consumption, order an energy-aware function more energy-efficient policy than demonstration. evaluate our dataset collected MIT Mini-Cheetah simulator. code is publicly available. 1