Neural Network Dynamics Models for Control of Under-actuated Legged Millirobots.

Abstract

Millirobots are a promising robotic platform for many applications due to their small size and low manufacturing costs. However, controlling these millirobots is difficult due to their underactuation, power constraints, and size. While hand-engineered controllers can sometimes control these millirobots, they often have difficulties with highly dynamic maneuvers and complex terrains. We present a learning based approach in which a model of the dynamics is learned from data gathered by the millirobot, and that data is then leveraged by an MPC controller. We show that with 17 minutes of random data collected with the VelociRoACH millirobot, the VelociRoACH can accurately follow trajectories at higher speeds and on more difficult terrains than a differential drive controller. Experiment videos can be found at this https URL