Flip Stunts on Bicycle Robots using Iterative Motion Imitation

TL;DR

This paper introduces Iterative Motion Imitation (IMI), a reinforcement learning method that enables bicycle robots to perform acrobatic flips by iteratively refining infeasible reference motions into feasible, agile behaviors.

Contribution

The paper presents IMI, a novel iterative imitation approach that successfully trains bicycle robots to perform flips from infeasible references, achieving real-world robustness and first-of-its-kind acrobatic behavior.

Findings

IMI outperforms single-shot imitation with higher success rates.

Policies trained with IMI transfer robustly from simulation to real-world.

First demonstration of unassisted acrobatic flips on a bicycle robot.

Abstract

This work demonstrates a front-flip on bicycle robots via reinforcement learning, particularly by imitating reference motions that are infeasible and imperfect. To address this, we propose Iterative Motion Imitation(IMI), a method that iteratively imitates trajectories generated by prior policy rollouts. Starting from an initial reference that is kinematically or dynamically infeasible, IMI helps train policies that lead to feasible and agile behaviors. We demonstrate our method on Ultra-Mobility Vehicle (UMV), a bicycle robot that is designed to enable agile behaviors. From a self-colliding table-to-ground flip reference generated by a model-based controller, we are able to train policies that enable ground-to-ground and ground-to-table front-flips. We show that compared to a single-shot motion imitation, IMI results in policies with higher success rates and can transfer robustly to the real world. To our knowledge, this is the first unassisted acrobatic flip behavior on such a platform.