Learning-Based Locomotion Control
Ball-Balancing Locomotion on a Quadruped
A Unitree Go2 walks while keeping a free-rolling ball balanced on a plate mounted on its back — learned in simulation, transferred to the real robot.
- Period
- 2025 – 2026
- Platform
- Unitree Go2
- Methods
- Reinforcement Learning · Sim-to-Real Transfer · Domain Randomization
- With
- Hamidreza Raei (IIT), Arash Ajoudani (IIT), Panos Trahanias (FORTH)
Overview
Carrying a payload is easy; carrying an unstable payload is a control problem. In this project a quadruped robot walks while keeping a free-rolling ball balanced on a flat plate mounted on its back — a task that couples base motion, body attitude, and the ball’s dynamics at every step.
Approach
The controller is trained with reinforcement learning entirely in simulation, where the robot experiences thousands of randomized variations of the task, and is then transferred to a real Unitree Go2 — the sim-to-real recipe applied to a dynamics problem where the “payload” fights back. The task is a sharp benchmark for whole-body steadiness: any abrupt attitude change, foot slip, or jerky velocity tracking immediately shows up as ball motion.
Demo
The video below shows the real robot balancing the ball.
Publications
- Sim-to-Real Reinforcement Learning for Ball-Balancing Locomotion on Quadruped Robots. Changda Tian, Hamidreza Raei, Arash Ajoudani, Panos Trahanias. IEEE/ASME AIM 2026, Genova.