Zero-Shot Real-World Deployment. RAI Institute Creates 5-Motor Bicycle Robot: 8 m/s, Half-Second Flips, 1-Meter Jump

“Fall 11 to 19 times to flip half a second before the empty flip.”

In half a second, 5 motors. A bicycle robot without a rider completed a front flip.

It has no flywheel, no springs, no control moment gyroscopes, and no mechanical device specifically designed for balance. The entire robot has only 5 driving degrees of freedom, less than half of the typical 12 found in quadruped robots.

Relies on these 5 motors to jump onto a 1-meter-high platform, reach a maximum speed of 8 meters per second, and also perform single-leg jumps on the rear wheel, perform a head-up drive, and maintain balance in place.. All motion policies are trained in simulation and transferred directly to the physical robot through zero-shot transfer, without on-site parameter tuning or retraining.

This is the UMV (Ultra Mobility Vehicle, super mobility vehicle) robot published by the US RAI Institute in Science Robotics. It does not take the old path of 'stacking hardware for performance,' but instead actively reduces mechanical complexity and then compensates for dynamic capabilities through reinforcement learning.

01.

A children's bicycle with a 23.5-kilogram head

The appearance of UMV is easy to describe. It looks like a bicycle, but the handlebar position is occupied by a large “head.” The entire robot weighs 23.5 kilograms and is only 0.8 meters tall when fully crouched. The skeleton is based on a commercially available Specialized Hotwalk Carbon children's carbon fiber balance bike.

The head contains batteries, a computing unit, and a jump motor. The design logic is simple, to make the robot jump high, flip fast, and balance well, the most effective method is to concentrate most of the mass on the topmost link.

When the head accelerates upward rapidly through the linkage, according to Newton's third law, it exerts a downward reaction force on the bicycle linkage, causing the rear wheel to forcefully push against the ground, and the entire robot is thus propelled forward. This is similar to how a person jumps by swinging their arms, except that the UMV's 'arms' are much heavier.

Balance is also the same logic. When landing on one wheel, if the robot tilts to the left, its head swings to the left, generating a counteracting torque to the right, pulling the center of mass back to the balanced position, the principle is exactly the same as when a human rides a unicycle and swings their arms.

The front flip is even more clever. After takeoff, pulling the head in reduces the moment of inertia, increasing the rotational speed; when landing, extending the head out increases the moment of inertia, slowing down and cushioning the landing for a smooth touchdown.

When jumping onto a 1-meter high platform, the motor power peaks at 4.5 kilowatts, for 0.25 seconds. What does this number mean? The power density of a 23.5 kg robot at the moment of takeoff is close to that of an electric vehicle motor.

The design philosophy of UMV can be summarized in one sentence: Cut out all auxiliary stabilization devices, and hand the problem over to the algorithm.

No springs, no reaction wheels, no control moment gyroscopes, single rail narrow body, no wide support surface from multiple rails. Stabilization, energy storage, and cushioning—functions originally handled by mechanics—are now entirely entrusted to reinforcement learning strategies.

02.

The algorithm “invented” the techniques of human drivers on its own

To address hardware shortcomings, a constraint-enhanced reinforcement learning framework is used, with PPO (Proximal Policy Optimization) algorithm at its core.

The research team trained four main strategies using this algorithm: driving, wheelie jumps, flipping, and jump ramps. The entire training process was conducted in the NVIDIA Isaac Lab simulation environment. The approach is similar to teaching a child to ride a bike, where if they fall, they start over, and they are rewarded for doing it correctly.

The team did not write any “crouch, then push, then retract” action scripts for the robot, nor did they specify exact contact timing. They only did three things:

First, design a reward function to tell the robot what is good.

The rewards are divided into three categories: task rewards have the highest weight, for example, tracking target speed, jumping toward the target position; style rewards encourage stable posture; regularization rewards penalize excessive torque, action jitter, and overly forceful impacts. The robot explores by trial and error, figuring out how to score high.

Second, replace soft penalties with hard termination to simulate real hardware damage.

Once the landing speed exceeds the limit, power overload occurs, joints hit their limits, or the body collides with the ground, the current training round is immediately terminated and restarted. In the early stages of training, constraints are loose, and as the strategy improves, they gradually tighten, allowing AI to first learn core actions, then learn to act within the hardware safety boundaries.

Third, full randomization is used to force the strategy to become robust.

Randomly altering almost all parameters during training: aircraft mass, tire friction, center of mass position, motor power, controller gains, drive delay. Sensor readings are added with noise, and the robot is occasionally pushed by a random external force. The policy learns to complete tasks in various 'imperfect' scenarios, rather than only under a set of precise parameters.

The finally trained strategy can be directly copied to a real robot and used without any on-site fine-tuning or additional training, achieving zero-shot transfer.

More interestingly, there are two self-emerged actions, which engineers never designed; they were discovered by the AI itself in the reward space.

The first is oscillating turning in place. Lifting the front wheel, the robot rapidly swings its head left and right, causing the front wheel to oscillate horizontally, enabling the robot to turn around the rear wheel. It can turn without moving forward, which is entirely a control strategy spontaneously found by the policy;

The second is hopping and rolling balance on the rear wheel. When the rear wheel makes a single-point jump and contacts the ground, it first rolls backward a little before pushing forward, relying on forward and backward rolling to maintain balance, which is highly similar to the skills of a unicycle rider.

These two actions are very difficult to manually code using traditional control methods.

03.

Seven months, dozens of crashes, half a second of aerial flips

The experimental data in the paper brings the capabilities of this extremely simple robot to a very intuitive level.

• Maximum speed:8 m/s (approximately 28.8 km/h), exceeding the steady-state cruising speed of conventional quadruped robots;
• Obstacle clearance height:1 meter, reaching 130% of its crouched height;
• Continuous reliability:15 autonomous jump tests, with trajectory height matching during the takeoff phase, with fluctuations only occurring during the landing impact phase.

The most difficult action is the front aerial flip, which is also the part with the highest engineering cost.

UMV has a heavy lower body and light upper body, making it difficult for the lightweight chassis to withstand the impact from the upper heavy block upon landing, which easily leads to direct damage. Jin Zhenghuan, a Ph.D. student at Georgia Tech, specifically tackled this problem during his internship at RAI.

The method he used is called Iterative Motion Imitation (IMI). The core idea is to first generate a rough, physically even infeasible flipped trajectory as a “sketch,” letting the reinforcement learning policy imitate it. With each iteration, it optimizes a little, gradually converging to actions that the real robot can safely execute.

The process was extremely harsh. During two semesters of internship, the prototype was broken dozens of times, and each time it broke, the hardware team had to disassemble and repair it. It took nearly five months to achieve safe landing without damaging the hardware; another two months to achieve a smooth transition to the running state.

Seven months of iteration, dozens of hardware damages, resulting in less than half a second of stable aerial flip.

To what extent is this strategy stable? Shamel Fahmi, a RAI researcher, said in an interview that now it is a visitor demonstration project in the lab, very reliable, and has never had any failures.

04.

It is not replacing legs, but opening up a new design space

UMV is not intended to replace quadruped or biped robots. Terrain adaptability, load capacity, and static stability, legged robots still have irreplaceable advantages.

The value of UMV is that it has proven a new path—Mechanics do subtraction, computation does addition.

Previously, the industry assumed that to make robots more agile, you needed more joints, stronger actuators, and more precise transmission. UMV proved that in dynamic motion scenarios, as long as the learning control is strong enough, extremely simplified mechanical structures can also achieve many actions that were previously only possible with complex legged robots.

Of course, its limitations are also very clear. Currently, lab testing still relies on Optitrack motion capture system for state estimation, and there is still a distance to fully autonomous outdoor operation; the gap between simulation and reality remains the core challenge for such methods. The team is currently testing LiDAR inertial odometry, aiming to detach from motion capture and achieve fully autonomous operation.

But the key significance of this paper is that it raised a new design proposition; When computation is sufficiently cheap, simulation is sufficiently accurate, and reinforcement learning is sufficiently robust, will the first-order question in robot design shift from 'What structure do I need to add?' to 'What algorithm do I need to let learn?'

If this path proves viable, the cost reduction for the next generation of mobile robots may not come from cheaper hardware, but from a smarter brain.

Paper link: https://www.science.org/doi/10.1126/scirobotics.aed3627Return to the previous page for more information