3-Axis Unicycle Robot
Learning to Balance: A Reaction-Wheel Unicycle Robot
Capstone project by Tristan Lee, Julian Lapenna, Jackson Fraser, Kyle Mackenzie, and Simon Ghyselincks
UBC Engineering Physics 2025
Overview
This is Broomy, a fully autonomous, self-balancing reaction-wheel unicycle robot, built as a platform to benchmark classical and reinforcement learning (RL) control algorithms in real hardware.
We extended the well-known Wheelbot design to add:
- Additional axis of control (yaw)
- Supervised point-to-point navigation
- Full telemetry and remote control
- Direct sim-to-real testing for RL policies
The robot balances on a single wheel with internal reaction wheels for roll, pitch, and yaw stabilization.
Features
- Mechanical
- Compact 3D-printed PLA chassis (~28cm tall, ~2.05kg)
- Steel-ring flywheels for optimal inertia-to-mass ratio
- Self-righting geometry
- Electrical
- T-Motor Antigravity MN6007II brushless motors
- Moteus-n1 BLDC drivers
- 6S LiPo battery system for onboard power
- Software
- Jetson Orin Nano running Ubuntu 22.04 (PREEMPT-RT patched)
- 80Hz Python-based control loop
- MQTT telemetry, InfluxDB logging, Grafana dashboards
- Simulation in NVIDIA Isaac Sim
Control Algorithms
Implemented and benchmarked controllers:
- LQR (Linear Quadratic Regulator)
- Best performance in real-world tests
- Handles roll and pitch effectively
- Yaw remains uncontrolled (future work)
- Reinforcement Learning (RL)
- Trained in Isaac Gym with PPO
- Direct sim2real deployment
- Good performance in sim, but suffers from sim2real gap in physical testing
- Fast Model Predictive Control (Fast-MPC)
- Prototype stage; paused due to minimal benefit over LQR
Comparison of Balance Control
LQR holds balance well; RL drifts and falls.
Results
- LQR: Stable balancing indefinitely in hardware.
- RL: Learned balancing in simulation, with ~14s average balance time in real hardware (Table 2). Sim2real transfer works but not robust.
-
Stand-up from rest: LQR succeeds consistently, RL can stand but fails to stabilize.
- Disturbance rejection:
- LQR is solid, rejects major roll disturbances.
- RL shows potential but fails at high-magnitude disturbances.
LQR disturbance rejection.
RL disturbance rejection.
Deliverables
- Source Code
- CAD Models
- Isaac Sim RL Environment
- Physical prototype (assembled robot)
- Telemetry server and database
- Spare parts and prototyping components
Limitations
- Yaw axis not actively controlled yet. Expect uncontrolled spins.
- RL policy underperforms in real hardware due to sim2real gap.
- Chassis vibrations from flywheel imbalance affect IMU accuracy (band-aid: weighted IMU averaging).
Future Improvements
- Active yaw control implementation
- Chassis reinforcement to dampen vibrations
- Improved domain randomization for better sim2real RL transfer
- Positional setpoint control for navigation
- Physics-informed machine learning models
Citations
See full 2454 report for detailed references and appendices.
Acknowledgements
Sponsored by UBC Engineering Physics Project Lab, with support from Cyrus Neary.