Trajectory-model-based reinforcement learning: Application to bimanual humanoid motor learning with a closed-chain constraint

摘要

We propose a reinforcement learning (RL) framework to improve policies for a high-dimensional system through fewer interactions with real environments than standard RL methods. In our learning framework, we first use off-line simulations to improve the controller parameters with an approximated environment model to generate samples along locally optimized trajectories. We then use the approximated dynamics to improve the performance of a tool manipulation task in a path integral RL framework, which updates a policy from the sampled trajectories of the state and action vectors and the cost. In this study, we apply our proposed method to a bimanual humanoid motor learning task in which we need to explicitly consider a closed-chain constraint. We show that a 51-DOF real humanoid robot can learn to manipulate a rod to hit via-points using both arms within 36 interactions in a real environment.