Flip Stunts on Bicycle Robots Using Iterative Motion Imitation
Jeonghwan Kim, Shamel Fahmi, Seungeun Rho, Sehoon Ha, Gabriel Nelson
AI summary
Problem
Standard motion imitation fails when reference trajectories are dynamically or kinematically infeasible, often causing training instability or unsafe behaviors that cannot transfer to physical hardware.
Approach
The method recursively trains a reinforcement learning policy to track a reference, then uses the policy’s rollout as a new, more feasible reference for the next training cycle, progressively refining imperfect demonstrations into safe behaviors.
Key results
- Successful ground-to-ground and ground-to-table front-flips on the UMV bicycle robot
- Higher success rates and faster convergence compared to single-shot motion imitation
- Robust sim-to-real transfer of acrobatic flip stunts on physical hardware
- First unassisted acrobatic flip demonstrated on this specific bicycle robot platform
Why it matters
Provides a scalable framework for learning extreme agile maneuvers on complex wheeled platforms by safely bridging imperfect demonstrations and real-world hardware constraints.
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.