上海科技大学知识管理系统

The recent rapid development of deep-learning-based control strategies has made the reality gap a critical issue at the forefront of robotics, especially for legged robots. We propose a novel system identification framework, Progressive Bayesian Optimisation (ProBO), to bridge the reality gap by tuning simulation parameters. Since dynamic locomotion trajectories are usually harder to narrow the reality gap than their static counterpart, we train a Gaussian process model with the easier trajectory data set and make it a prior to start the learning process of a harder one. We implement ProBO on a quadruped robot to narrow the reality gaps of a set of bounding gaits at different speeds. Results show that our methods can outperform all other alternatives after training the initial gait.