GAUSSIAN PROCESS REGRESSION FOR SIM-TO-REAL TRANSFER OF HOPPING GAITS

Simulation-based controllers are relatively easy to build and evaluate, but rarely transfer seamlessly to hardware. This is because of the reality-gap which is the discrepancy between simulations and hardware. Narrowing the reality-gap can speed up the deployment of simulated controllers to hardware, also known as sim-to-real. This paper presents sim-to-real transfer of controllers on a single leg hopping robot, a system that cycles between under-actuation during the stance phase to no-actuation during flight phase. Using simulations, we design a controller to achieve speed and height regulation once-per-step, but the controller cannot achieve accurate control on hardware. Using data from hardware, we model the mis-match between simulation and hardware using Gaussian Process Regression (GPR), recompute the controller, and redeploy it in hardware. It takes about 4 iterations to achieve accurate tracking. The results show that when GPR is used to model the step-to-step level model inaccuracy, it can lead to high accuracy sim-to-real transfer while maintaining sample efficiency. A video is here: tiny.cc/ idetc2023