Adaptive motor patterns and reflexes for bipedal locomotion on rough terrain

The Bio-inspired Behavior-Based Bipedal Locomotion Control (B4LC) system consists of control units encapsulating feed-forward and feedback mechanisms, namely motor patterns and reflexes. To optimize the performance of motor patterns and reflexes in terms of stable locomotion on both even and uneven terrains, we present a learning scheme embedded in the B4LC system. By combining the Particle Swarm Optimization (PSO) method and the Expectation-maximization based Reinforcement Learning (EM-RL) method, a learning unit is comprised of an optimization module and a learning module embedded in the hierarchical control structure. The optimization module optimizes the motor patterns at hip and ankle joints with respect to energy consumption, stability and velocity control. The learning module generates compensating torques against disturbances at the ankle joints by combining the basis function derived from state information and the policy parameters. The optimization and learning procedures are conducted on a simulated robot with 21 DoFs. The simulation results show that the robot with optimized motor patterns and learned reflexes performs a more robust and stable locomotion on even and uneven terrains.