Stability Control of a Biped Robot Based on Efficient Reinforcement Learning

Reinforcement Learning (RL) allows a robot to learn from the experience of interacting with the environment. This project will aim to maintain the stability of a biped robot on a rotating platform, where a novel RL algorithm that perfectly controls the robot is proposed. The project will combine Model Free RL (MRFL) and Model-Based RL (MBRL) as a Hybrid RL structure to improve stability performance, data efficiency and reduce convergence time. A Temporal Difference Algorithm is proposed as a high-level MFRL controller, and the Hierarchical Gaussian Process is implemented as a Low-Level MBRL to estimate the transition dynamics.