Improved Friction and Dynamics Estimation in Legged Robots

Abstract

Reducing the Sim-to-Real gap between robot simulation and robot performance could lead to improved and more efficient robot design through more accurate controls and design testing in simulation and through more accurate state detection for model-based control architectures. This work built upon current research in the field of robot system dynamics by investigating the effect of using single-layer feed-forward neural nets to model non-linear friction forces and other forms of dynamics that are difficult to account for with traditional robot system identification schemes. Applying the single-layer feed-forward neural nets to system identification data from the dynamic MIT Humanoid and MIT Mini Cheetah robots significantly reduced torque prediction errors. The neural net was able to reduce torque errors by modeling both linear and non-linear effects that could not be easily fit by traditional methods. The results of this paper suggest that using the system identification methodology outlined within could lead to more accurate dynamics modeling, which would assist with closing the Sim-to-Real gap through simulated dynamics with more fidelity and a more robust representation of robot dynamics.