Efficient and Robust Learning on Elaborated Gaits with Curriculum Learning

Developing efficient walking gaits for biomechanical robots is a difficult task that requires optimizing parameters in a continuous, multidimensional space. In this paper we present a new framework for learning complex gaits with musculoskeletal models. We use Deep Deterministic Policy Gradient which is driven by the external control command, and apply curriculum learning to acquire a reasonable starting policy. We accelerate the learning process with large-scale distributed training and bootstrapped deep exploration paradigm. As a result, our approach won the NeurIPS 2018: AI for Prosthetics competition, scoring more than 30 points than the second placed solution.