Session: MR-5-2: Motion Planning, Dynamics, and Control of Robots

Paper Number: 116959

116959 - Gaussian Process Regression for Sim-to-Real Transfer of Hopping Gaits 

To accelerate the development of robotic applications in the real world, it is important to use a simulation-based methodology to develop controllers. Indeed, simulations are computationally cheap (e.g., take a fraction of the time to set up and run), can be fully automated (e.g., running test cases without supervision), and inexpensive (e.g., robots don't break in simulation) as compared to hardware. However, simulation-based controllers 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

Presenting Author: Jeremy Krause University of Illinois at Chicago

Presenting Author Biography: Later, after paper acceptance, maybe.

Authors:

Jeremy Krause University of Illinois at Chicago
Adel Alaeddini The University of Texas at San Antonio
Pranav Bhounsule University of Illinois at Chicago