Flip Stunts on Bicycle Robots using Iterative Motion Imitation

Jeonghwan Kim; Shamel Fahmi; Seungeun Rho; Sehoon Ha; Gabriel Nelson

cs.ROMar 30, 2026

Flip Stunts on Bicycle Robots using Iterative Motion Imitation

Authors: Jeonghwan Kim, Shamel Fahmi, Seungeun Rho, Sehoon Ha, Gabriel Nelson

Organizations: RAI Institute, Cambridge, MA, USA · Georgia Institute of Technology, Atlanta, GA, USA

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.

Figures & tables

Fig. 1: UMV executing a front-flip learned via Iterative Motion Imitation .

Fig. 2: A 2D overview of the UMV model used in this work.

Fig. 3: An overview of IMI . The first iteration starts with an initial reference

\xi_{0}

used to train policy

\pi_{1}

. Then, reference

\xi_{n-1}

generated by policy

\pi_{n-1}

becomes the new more feasible reference used to train a new policy

\pi_{n}

. This loop continues, progressively refining the motion until a high-performance, hardware-deployable policy is achieved.

Tracking Rewards
Component	Formulation	Weight	Tolerance
Base Position	$\exp(-\alpha_{\text{base}}\\|p_{\text{base}}-p_{\text{base}}^{\text{ref}}\\|^{2})$	4.0	0.4
Base Orientation	$\exp(-\alpha_{\text{ang}}\\|d_{\text{angle}}(q_{\text{base}},q_{\text{base}}^{\text{ref}})\\|^{2})$	20.0	0.8
Joint Positions	$\exp(-\alpha_{\text{joint}}\\|q_{\text{joint}}-q_{\text{joint}}^{\text{ref}}\\|^{2})$	1.0	0.1
Penalty
Action Smoothness	$\\|a_{t}-a_{t-1}\\|^{2}$	-1e-5 to -1e-3	—

TABLE I: Reward Structure with Tolerance

Fig. 4: Overlayed screenshots of UMV performing a front flip.

Fig. 5: Comparison of training from scratch ( $\xi_{0}$ ) versus using refined reference ( $\xi_{1}$ ). Each run uses 3 seeds. The solid line is the mean and the shaded area is the standard deviation.

Fig. 6: The evolution of trajectories

\xi

over two iterations. The time is aligned at the peak height (

t=0

) and the interval is

\pm 0.5

s around that peak. (a) bike-base X-Z position, (b) lower-body and upper-body joint angles, (c) bike-base vertical velocity, and (d) bike-base vertical acceleration.

Fig. 7: Three hardware flip experiments of UMV . The time is aligned at the peak height (

t=0

) and the interval is

\pm 0.5

s around that peak. (a) bike-base X-Z position, (b) lower-body and upper-body joint angles, (c) bike-base vertical velocity, and (d) bike-base vertical acceleration. The vertical dashed line indicates the time at take-off.

Fig. 8: Overlay of UMV flipping up a 70 cm box.

Fig. 9: Flipping up 70 cm (

\xi^{\prime}_{2,70}

) and 26 cm (

\xi^{\prime}_{2,26}

) boxes, each learned by imitating the reference trajectory

\xi_{1}

. (a) bike-base trajectory in the global frame, with the shaded region indicating the table position. (b) bike-base vertical trajectory aligned at peak height.

	Success Rate		Average Return
$\xi$	70cm	26cm	70cm	26cm
$\xi_{2}$	95.5%	97.2%	81.8	94.1
$\xi_{1}$	91.9%	92.4%	80.5	93.0
$\xi_{0}$	0.0%	85.1%	33.6	57.0

TABLE II: Flip-up performance from

\xi_{0}

\xi_{1}

, and

\xi_{2}

Explore similar work

Bicycle Acrobatics with Reinforcement Learning
Aug 1, 2026Shamel Fahmi, Arianna Ilvonen, Samuel Zapolsky +8Robot Skill Learning Robotic RL
1Robotics and AI Institute (RAI), Cambridge, MA, USA.
LineRides: Line-Guided Reinforcement Learning for Bicycle Robot Stunts
May 6, 2026Seungeun Rho, Shamel Fahmi, Jeonghwan Kim +3Reinforcement Learning Robot Skill Learning
1RAI Institute, Cambridge, MA, USA · Georgia Institute of Technology, Atlanta, GA, USA
System Design of the Ultra Mobility Vehicle: A Driving, Balancing, and Jumping Bicycle Robot
Feb 25, 2026Benjamin Bokser, Daniel Gonzalez, Aaron Preston +39Robot Locomotion Robot Morphology-Control Co-Design

Aug 1, 2026J/K move · Enter open · S save

Bicycle Acrobatics with Reinforcement Learning

Shamel Fahmi, Arianna Ilvonen, Samuel Zapolsky, Yu-Ming Chen+7

1Robotics and AI Institute (RAI), Cambridge, MA, USA.

Robot Skill Learning Robotic RL Robotics Robot Locomotion

Open paper AskPDF arXiv

Bicycle robots are fast and energy efficient, but their simple mechanical design and their underactuated and non-holonomic dynamics make highly agile maneuvers difficult to achieve. Here, we use Reinforcement Learning (RL) to enable a bicycle robot to learn and compose a diverse repertoire of dynamic acrobatic stunts. Using different RL formulations such as waypoint following, pose reaching, twist tracking, guided tracking, and motion imitation, the robot acquires autonomous single and multi-table forward and lateral jumps, steerable jumps, front flips, kip-ups, kip-downs, driving, wheelies, bunny hops, and three-point turns. To coordinate these behaviors, we introduce an orchestrator that transitions between policies using state-dependent triggers, enabling robust long-horizon acrobatic stunts. We validate the approach on the Ultra Mobility Vehicle (UMV), a custom bicycle robot, in simulation and hardware. The robot repeatedly traverses tables up to 1 m high, performs more than 15 consecutive autonomous jumps while following waypoints, handles previously unseen multi-table configurations, executes continuous repertoires of kipups, jumps, flips, kip-downs, over more than 20 consecutive trials, and performs more than 10 consecutive autonomous and steerable repertoires of wheelies, lateral jumps, and single-wheel jump downs. These results demonstrate that RL can endow bicycle robots with levels of agility previously associated primarily with legged platforms while preserving the speed and efficiency of wheeled locomotion, establishing a foundation for bicycle acrobatics.

Abstract

Explore similar work

Bicycle Acrobatics with Reinforcement Learning

LineRides: Line-Guided Reinforcement Learning for Bicycle Robot Stunts

System Design of the Ultra Mobility Vehicle: A Driving, Balancing, and Jumping Bicycle Robot