By Thomas Li, Team Lead and RL Engineer, Equilibrio — ETH Zurich Focus Project
Anyone who has ever tried riding a unicycle knows the feeling of those first few seconds: for a brief moment, you think it might just work—and then gravity delivers an uncompromising reality check. A unicycle is a textbook unstable system. There is no second wheel for support, no foot you can quickly put down, just a single point of contact and the constant need to react instantly before the system tips over.
This is precisely the challenge we are tackling with Wheeliam. Developed this year at the Robotic Systems Lab (RSL, ETH Zurich) as part of the Equilibrio Focus Project, Wheeliam is a self-balancing unicycle robot featuring a 3-DoF hip and a leg with closed 5-bar kinematics. This architecture allows it to crouch, stand upright, and ultimately jump over obstacles.
Behind the project is a team of twelve students: nine mechanical engineering students and one electrical engineer from ETH Zurich, along with two systems engineering students from ZHAW. Wheeliam weighs around 40 kg, stands approximately 1.5 m tall in a crouched stance, and extends up to 2.1 m. Our vision is a robot that navigates where conventional wheeled platforms and legged robots hit their limits: narrow paths, steep slopes, scree fields, and rugged terrain in between. To achieve this, we are not writing the controller by hand—we are letting the robot learn on its own.
Why Reinforcement Learning?
Classical control theory excels when a system can be accurately modeled. For Wheeliam, however, analytical modeling reaches its limits. A unicycle robot with a jump-capable leg, an actively stabilizing hip, and contact dynamics that shift with every surface cannot be easily captured in a compact mathematical model.
Reinforcement Learning (RL) provides an alternative: the robot is allowed to explore, fail, and iteratively improve. We define what desirable behavior looks like—such as tracking velocity commands, maintaining an upright posture, and operating energy-efficiently. The learning algorithm then optimizes a policy: a neural network that maps raw observations directly to motor commands.
The trade-off is sample efficiency: RL requires millions of trials. Crashing physical hardware at that scale is impossible. Therefore, we train inside simulation first and transfer the learned policy onto the physical hardware later. This sim-to-real transfer is one of the project’s most demanding phases, accounting for a major portion of our engineering effort.
Simulation Training and Hardware Transfer
We train Wheeliam’s policy using Proximal Policy Optimization (PPO) in NVIDIA Isaac Lab, simulating thousands of Wheeliams in parallel on a single GPU. The policy itself is a recurrent neural network: a GRU processes temporal dependencies, followed by an MLP that computes the joint setpoints.
The policy relies exclusively on proprioceptive observations. These include IMU measurements (orientation and angular velocity), joint positions and velocities, and a short history of past actions. From this data, it infers target commands for both the wheel and leg actuators. The policy runs at 50 Hz, while the underlying physics simulation operates at 200 Hz.
Bridging the sim-to-real gap requires several systematic techniques:
-
Domain Randomization: During training, no two simulated Wheeliams are identical. We randomize parameters like body masses, wheel-ground friction, restitution, and IMU sensor noise. This exposes the policy to a broad distribution of dynamics, ensuring the physical system feels like just another variation it already knows how to handle.
-
Privileged Information: The critic network receives state data during training that the actor cannot access at runtime, such as ground-truth linear velocities, contact forces, and exact segment masses. This privileged setup stabilizes training while keeping the deployed actor strictly bound to onboard sensory inputs.
-
Actuator Modeling with PACE: Real actuators deviate from ideal torque sources due to friction, inertia, damping, and bus latency. Using PACE—an RSL-developed system identification framework—we extract actuator parameters directly from logged hardware runs, aligning the simulated motor responses with the physical drive units.
-
Teacher-Student Pipeline: We also utilize a teacher-student framework, training an unconstrained teacher policy with full state access, then distilling its behavior into a recurrent student policy restricted solely to hardware-available inputs.
Current Milestones
In simulation, the policy demonstrates robust disturbance rejection and reliable velocity tracking. On real hardware, we have achieved promising initial results: the policy successfully stabilizes Wheeliam in place, balancing autonomously for roughly one minute. Given the system's aggressive underactuated dynamics, this represents an essential baseline.
In parallel, we are already putting the explosive dynamics and power delivery of the 5-bar leg kinematics to the test:
Hardware testing at the Robotic Systems Lab: Wheeliam demonstrates its vertical jumping power and leg shock absorption while tethered to the safety gantry.
Our immediate focus centers on refining the actuator models, broadening simulation randomization, and tuning observation processing to transition from stationary balance to controlled, forward locomotion.
Next Steps and the Collaboration with MYBOTSHOP
Wheeliam currently operates solely on proprioception. It tracks its own joint states and dynamic balance, but remains blind to external surroundings and absolute localization. This was a deliberate choice: solve dynamic balance first, then introduce environmental perception. However, true outdoor deployment demands spatial awareness; a unicycle robot unable to spot a curb cannot navigate complex terrain.
This is where MYBOTSHOP’s sponsorship proves essential. Their support equipped Wheeliam with a Robosense Airy LiDAR and a SIRIUS RTK GNSS Base Station. These sensors unlock the project's next phase: developing an exteroceptive policy that couples dynamic balancing with real-time mapping to handle challenging outdoor environments. We are grateful for MYBOTSHOP’s backing as we integrate these systems into our next hardware iteration.
To follow Wheeliam’s progress, visit www.equilibrio.ethz.ch or connect with Equilibrio on LinkedIn.
