Model-free reinforcement learning for robust locomotion using demonstrations from trajectory optimization

Problem

Creating locomotion policies robust enough to deploy directly on real robots, where training from scratch on hardware is impractical.

Why it matters

Robust real-robot locomotion without extra training is a key barrier to deploying legged robots outside labs.

What the AI did

A two-stage model-free reinforcement learning approach uses a single trajectory-optimization demonstration to bootstrap exploration, then optimizes the task reward for robustness to environment uncertainties.

Results

not stated in source

Impact

Approach was demonstrated on highly dynamic hopping and bounding tasks on a quadruped robot, per the abstract.

Record

Importance rating

More in Robotics & Autonomy