From Simulation to Reality

Bridging the Learning Gap on a Compliant Quadruped Robot using Domain Randomnization.

Research on adaptive locomotion has been conducted for many years, especially through neurophysiological and biomechanical studies generally carried independently. However, those complex motor behaviours originate from interactions between the neural system, the musculoskeletal system and the environment which makes exhaustive research in vivo hard to realize in practice. Embodied experiments on real robots constitutes an promising solution to test the models in closed-loop but the latter generally require a long learning phase that can hardly be conducted on the real robot. A large community of researcher try to pre-train the models in simulation but faces the reality gap when transferring results on real platforms.

In his last paper, Alexander Vandesompele investigates the role of body parameters randomnization as a promising method of regularization during the transfer of a locomotion controller from simulation to reality. He used the Tigrillo platform, a compliant quadruped robot whose improvements in electronics and software had been presented in a previous post. I had the opportunity to help with the design of some experiments on what seems a good direction for improvements on robot learning.

Going Further

To better understand the challenges behind sim-to-real gap, I recommend the following keynote lecture of Pieter Abbeel, specialist in advanced Reinforcement Learning for Robotics.

The BAIR Lab, led by Pieter Abbeel has become a reference on bridging the sim-to-real gap.