Bipedal Locomotion Primitive Learning, Control and Prediction from Human Data

Abstract

At the current stage bipedal robot locomotion is quite different from human walking. Imitation learning framework from human demonstrations is an efficient approach to lead towards human-like behaviors. This paper addresses a framework for real-time whole-body human motion imitation by a humanoid robot. The framework is a structured mixture of whole body motion control, learning and prediction. Human movements are mapped to robot's kinematics in combination with a balancing algorithm in order to ensure the dynamic constraints during different stance phases. Once locomotion primitives are learned from human demonstrations using hidden Markov models, the robot can recognize human's current locomotion state and predict future trajectories using Gaussian regression. The proposed concepts are implemented and evaluated with a small humanoid robot NAO.