asRoBallet: Closing the Sim2Real Gap via Friction-Aware Reinforcement Learning for Underactuated Spherical Dynamics
First RL-based humanoid on a ball, mastering friction for real-world balance and dance.
Researchers from multiple institutions have unveiled asRoBallet, the first successful reinforcement learning (RL) deployment on a humanoid ballbot hardware—a robot balancing on a single sphere. Historically, ballbots have been a benchmark for underactuated control, but the complex friction between wheels, sphere, and ground created a reality gap that prevented RL from working on actual hardware. This study, accepted for RSS2026, closes that gap with a friction-aware RL framework.
The team built a high-fidelity MuJoCo simulation that explicitly models discrete roller mechanics of ETH-type omni-wheels, capturing parasitic vibrations and contact discontinuities previously ignored. Their RL policy masters coupled rolling, lateral, and torsional friction channels, enabling zero-shot sim-to-real transfer. The hardware was created by reconfiguring an overconstrained quadruped into a lightweight humanoid ballbot, and an iOS ecosystem allows intuitive control via natural motion. This opens doors for dynamic, expressive humanoid maneuvers in real-world settings.
- First RL-based controller deployed on a humanoid ballbot, achieving zero-shot sim-to-real transfer.
- Friction-aware framework models three friction channels (rolling, lateral, torsional) in MuJoCo simulation.
- Low-cost hardware built from repurposed quadruped components; controlled via an iOS app for expressive maneuvers.
Why It Matters
Bridges the sim-to-real gap for humanoid ballbots, enabling agile, low-cost robotics for real-world interaction and entertainment.