Reinforcement learning method for legged robots capable of rapid response and continued locomotion despite motor failure

Fault-Tolerant Reinforcement Learning for Legged Robots Using Knowledge Distillation and Joint Trajectory Space Learning

This technology is a reinforcement learning-based gait control method that enables robots to quickly resume adaptive walking when hardware failures, such as leg damage, occur. It achieves this by distilling knowledge from an agent trained in a normal state and utilizing it as a refined joint trajectory space through an encoder-decoder neural network.

Existing gait control technologies can adapt to terrain or environmental changes, but they face inefficiencies when hardware failures occur, often leading to a loss of control or requiring the agent to be retrained from scratch.

This technology proposes a method that uses a conditional variational autoencoder to set the joint trajectory space as the action space, narrowing the search space during failures based on knowledge learned in a normal state. It generates anchor points and paths based on conditional vectors to derive optimal joint trajectories in real time. This ensures robust autonomy, allowing robots to autonomously reconfigure their gait even when legs are damaged, making it ideal for environments where mission interruption is critical, such as disaster site exploration, defense, and industrial patrolling.

‍

‍

Key Features:

• Generating anchor points corresponding to conditional vectors in the gait learning method for n-legged robots
• Constructing a joint trajectory search space by generating paths corresponding to the generated anchor points
• Performing simulations for joint trajectories corresponding to the paths and obtaining rewards
• Updating the gait policy based on the obtained rewards and learning the gait policy

‍

This invention was developed with support from the Artificial Intelligence Graduate School Program (Korea University) funded by the Ministry of Science and ICT.