💻 computer science

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

Inspired by trials cyclists, this paper presents the design and control of the Ultra Mobility Vehicle (UMV), a 23.5 kg robotic bicycle that utilizes simulation-driven optimization and constrained reinforcement learning to achieve zero-shot transfer of diverse athletic behaviors, including high-speed driving, single-wheel balancing, and jumping over obstacles up to 130% of its height.

Original authors: Benjamin Bokser, Daniel Gonzalez, Aaron Preston, Alex Bahner, Annika Wollschläger, Arianna Ilvonen, Asa Eckert-Erdheim, Ashwin Khadke, Bilal Hammoud, Dean Molinaro, Fabian Jenelten, Henry Mayne, Howie

Published 2026-03-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Benjamin Bokser, Daniel Gonzalez, Aaron Preston, Alex Bahner, Annika Wollschläger, Arianna Ilvonen, Asa Eckert-Erdheim, Ashwin Khadke, Bilal Hammoud, Dean Molinaro, Fabian Jenelten, Henry Mayne, Howie Choset, Igor Bogoslavskyi, Itic Tinman, James Tigue, Jan Preisig, Kaiyu Zheng, Kenny Sharma, Kim Ang, Laura Lee, Liana Margolese, Nicole Lin, Oscar Frias, Paul Drews, Ravi Boggavarapu, Rick Burnham, Samuel Zapolsky, Sangbae Kim, Scott Biddlestone, Sean Mayorga, Shamel Fahmi, Surya Singh, Tyler McCollum, Velin Dimitrov, William Moyne, Yu-Ming Chen, Farbod Farshidian, Marco Hutter, David Perry, Al Rizzi, Gabe Nelson

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a robot that doesn't just walk or roll; it rides.

Meet the Ultra Mobility Vehicle (UMV). It looks like a bicycle, but it's actually a high-tech acrobat. While most robots are either slow and steady (like a walking dog) or fast but clumsy (like a car that can't climb stairs), the UMV is designed to do both. It can zip down a smooth road at 18 mph, but it can also hop over a 3-foot wall, balance on one wheel like a circus performer, and even do a front flip.

Here is the story of how they built it, explained simply.

1. The Big Idea: The "Human Cyclist" Trick

Think about a professional trials cyclist (the kind who jumps over rocks and balances on tiny ledges). They don't need a third wheel or a stabilizing bar. They stay upright by constantly shifting their body weight. If they lean left, they steer right. If they want to jump, they throw their body up and pull the bike up with them.

The engineers wanted a robot that could do this, but without the complexity of a robot with four legs (which has too many moving parts) or a robot with a giant spinning wheel inside it (which is heavy and wastes energy).

The Solution: They built a robot that is 80% bicycle and 20% "reaction mass."

The Bicycle: The bottom part is a lightweight bike frame with two wheels.
The "Head": The top part is a heavy, powerful box containing the brain, batteries, and motors.
The Secret Sauce: The "Head" is connected to the bike by a special mechanical arm. The motors in the Head can swing the heavy top part around wildly. By throwing its "head" one way, the robot forces its "body" (the bike) to go the other way. It's like a cat twisting its body in mid-air to land on its feet.

2. How It Learned to Move: The "Video Game" Coach

You can't program a robot to do a front flip by writing a list of rules like "move motor A, then motor B." It's too complicated. Instead, the team used Reinforcement Learning (RL).

Think of this like training a dog, but for a robot:

The Simulator: They created a perfect digital twin of the robot in a video game.
The Trial and Error: They let the robot try to move in the game millions of times.
- If it fell over? Game Over. (No points).
- If it stayed balanced? Good job! (Points).
- If it jumped high? Awesome! (More points).
The Result: The robot "learned" the physics of balancing and jumping on its own. It discovered tricks that humans might not have thought of, like rolling its back wheel backward slightly before jumping to get more lift.

Once the robot became a master in the video game, they turned it on in the real world. Because the training was so robust, the robot didn't need to re-learn anything. It just started doing the same tricks in real life. This is called "Zero-Shot Transfer," which is a fancy way of saying, "It learned in the game and performed perfectly in real life on the first try."

3. What Can It Actually Do?

The paper highlights some incredible stunts:

The "Shimmy-Turn": Imagine a unicycle rider spinning in place without moving forward. The UMV can do this by lifting its front wheel and using its heavy "head" to twist the bike around the back wheel.
The "Track Stand": It can balance perfectly still on a hill without moving a muscle, just like a pro cyclist.
The "Table Jump": It can see a table, run up, launch itself 3 feet into the air, land on the table, drive across, and jump off the other side.
The Front Flip: It can jump, tuck its body in tight (to spin faster), do a full front flip, and land on its wheels ready to drive again.

4. Why Is This a Big Deal?

Most robots are built with a specific job in mind. A delivery robot is built to carry boxes; a rescue robot is built to climb stairs. The UMV is a hybrid.

Wheels are efficient: They are great for moving fast on flat ground (like a car).
Legs are agile: They are great for jumping over obstacles (like a goat).

The UMV combines the best of both worlds. It's as efficient as a bike but as agile as a parkour athlete. It weighs about as much as a large dog (52 lbs), but it can jump 130% of its own height.

The Bottom Line

This robot proves that you don't need a complex, expensive, multi-legged machine to navigate rough terrain. By combining a simple bicycle frame with a smart, heavy "head" and teaching it through video-game-style trial and error, the engineers created a machine that is fast, energy-efficient, and incredibly acrobatic.

It's the difference between a robot that walks over a puddle and a robot that jumps over it, lands, and keeps going without breaking a sweat.

1. Problem Statement

Robots navigating complex terrains face a fundamental trade-off between agility and efficiency.

Legged robots (bipeds/quadrupeds) excel at overcoming obstacles and uneven terrain but suffer from high mechanical complexity, high mass, and lower energy efficiency due to the large number of actuated degrees of freedom (DoFs) required (typically 12+ DoFs).
Wheeled robots are highly efficient and fast on smooth terrain but struggle with vertical discontinuities and obstacles.
Existing Hybrids: Current "wheel-legged" hybrids often use multi-track wheel configurations (wheels side-by-side) or add reaction wheels/gyroscopes for stability. These solutions increase mass, footprint, and mechanical complexity, often failing to achieve the agility of elite human cyclists who can balance on one wheel, jump, and perform flips with minimal actuation.

The core challenge is to design a robot that combines the energy efficiency and speed of a bicycle with the dynamic agility and obstacle negotiation of a legged system, using a minimal number of actuated DoFs and without relying on passive stability mechanisms (like kickstands) or heavy auxiliary stabilizers (like reaction wheels).

2. Methodology

The authors propose the Ultra Mobility Vehicle (UMV), a dynamically stable, single-track bicycle robot controlled via a simulation-driven design and a constrained Reinforcement Learning (RL) framework.

A. System Design & Morphology

Structure: The robot consists of three main links: a Bike (rear wheel drive and steering), a Neck, and a Head.
Actuation: It utilizes only 5 actuated DoFs:
1. Steering ( $q_s$ )
2. Rear-wheel drive ( $q_w$ )
3. Three DoFs in the Head ( $q_h, q_l, q_r$ ) controlling a spatial linkage (Neck + tie rods) that moves the Head relative to the Bike.
Key Design Choices:
- Mass Remotization: The heavy components (battery, compute, jumping actuators) are concentrated in the Head. This maximizes the center of mass (CoM) height, which is critical for maximizing vertical jump height.
- Spatial Linkage: The Head is connected via a 3-DoF spatial linkage. This allows the robot to "throw" its mass to generate angular momentum for balancing and jumping.
- Out-of-Plane Motion: A specific DoF ( $\phi$ ) allows lateral movement of the Head, enabling single-wheel balancing (like a unicycle) without reaction wheels.
Hardware: The robot weighs 23.5 kg and is 0.8 m tall (crouched). It uses high-power density motors (CubeMars AK10-9) and a custom battery pack capable of 7.3 kW peak discharge to support explosive jumps.

B. Control Framework: Constrained Reinforcement Learning

The control policy is learned via Constrained Markov Decision Processes (CMDP) using Proximal Policy Optimization (PPO).

Zero-Shot Transfer: Policies are trained entirely in simulation (Isaac Lab) with extensive Domain Randomization (randomizing mass, friction, delays, noise) and transferred directly to the real robot without fine-tuning.
Constrained RL: Instead of soft penalties, safety constraints (e.g., max motor current, joint limits, landing velocity) are enforced as hard constraints. Violating a constraint immediately terminates the training episode. This forces the agent to learn safe behaviors naturally.
Unified Pipeline: A single policy architecture is used for diverse behaviors (driving, wheelies, hopping, flipping) by modifying the reward function and observation inputs (e.g., adding gait mode commands or target locations).
State Estimation: In the lab, a motion capture system is fused with onboard IMUs via a factor graph (GTSAM) for high-frequency state estimation. The system is also capable of self-contained state estimation using LiDAR-inertial odometry.

C. Design Optimization

The morphology was optimized via simulation to maximize vertical jump height and single-wheel balancing controllability.

Jump Optimization: The design prioritized moving mass to the top link (Head) and optimizing gear ratios to maximize the ratio of extended height to crouched height.
Balance Optimization: Controllability analysis confirmed that adding an out-of-plane joint ( $\phi$ ) significantly reduced the control effort required to balance on a single rear wheel, outperforming designs relying solely on steering.

3. Key Contributions

Novel Morphology: Introduction of a low-DoF (5 DoFs), single-track bicycle robot that achieves high-agility acrobatics without reaction wheels or complex leg mechanisms.
Principled Design Process: A simulation-driven optimization workflow that balances jump height, mass distribution, and mechanical complexity, resulting in a specific spatial linkage topology.
Unified Learning Framework: A constrained RL pipeline that synthesizes policies for a diverse set of athletic behaviors (driving, balancing, wheelies, hopping, flipping) using a single architecture and minimal reward modifications.
Zero-Shot Sim-to-Real: Successful deployment of complex dynamic maneuvers on hardware without explicit datasets or demonstrations, relying purely on "real-to-sim" modeling and domain randomization.

4. Results

The UMV robot demonstrated state-of-the-art performance in both speed and agility:

Speed: Achieved a maximum speed of 8 m/s, outperforming standard quadrupeds and rivaling wheel-legged robots like ANYmal on Wheels.
Obstacle Clearance: Successfully jumped onto and over 1-meter tall obstacles (approx. 130% of the robot's height).
Dynamic Behaviors:
- Track-Standing: Balancing in place using only wheel and steering actuation (with Head actuation improving stability).
- Shimmy-Turn: Turning in place by laterally stepping the front wheel while balancing on the rear wheel.
- Rear-Wheel Hopping: Continuous monopodal hopping in arbitrary directions, including lateral hops.
- Front Flips: Executing a full front flip by tucking the body to reduce moment of inertia, accelerating rotation to ~17 rad/s, and landing stably.
Robustness: The system demonstrated high repeatability in table jumps (15 consecutive jumps) and the ability to recover from disturbances during hopping.

5. Significance

The UMV represents a paradigm shift in robotic mobility by proving that mechanical underactuation combined with high-bandwidth learning-based control can yield athletic performance previously thought to require complex, heavy, or multi-track systems.

Efficiency vs. Agility: It bridges the gap between the energy efficiency of wheels and the obstacle negotiation of legs, offering a platform that is both fast and agile.
Scalability: With fewer actuators and a simpler mechanical structure than legged robots, the UMV offers advantages in manufacturing cost, weight, and maintenance.
Future Applications: The capabilities of the UMV make it highly suitable for applications requiring rapid traversal of rough terrain, such as urban courier services, search and rescue in confined spaces, and long-distance reconnaissance.
Control Theory: The work demonstrates that constrained RL can effectively handle complex, non-minimum phase dynamics and discover emergent strategies (like the "shimmy-turn") that are difficult to encode in classical control templates.

In summary, the Ultra Mobility Vehicle demonstrates that a bicycle-based morphology, when augmented with a reaction mass and controlled by advanced RL, can achieve a level of dynamic athleticism that rivals or exceeds current legged and hybrid robots, all while maintaining a compact footprint and high energy efficiency.