asRoBallet: Closing the Sim2Real Gap via Friction-Aware Reinforcement Learning for Underactuated Spherical Dynamics
Researchers have introduced asRoBallet, recognized as the first end-to-end reinforcement learning locomotion policy successfully deployed on a humanoid ballbot hardware platform. This study addresses significant challenges in transitioning from simulation to reality for underactuated spherical dynamics, specifically targeting gaps in contact modeling, actuator latency, and safe hardware exploration. The team developed a high-fidelity MuJoCo simulation that explicitly models the discrete roller mechanics of ETH-type omni-wheels, capturing previously ignored parasitic vibrations and contact discontinuities. Furthermore, they created a Friction-Aware Reinforcement Learning framework that masters coupled rolling, lateral, and torsional friction channels, enabling zero-shot Sim2Real transfer. The physical platform was constructed through subtractive reconfiguration, repurposing components from an overconstrained quadruped into a new structural frame to ensure robustness and low cost. Additionally, a generalized iOS ecosystem was developed to transform consumer electronics into a low-latency interface, allowing single operators to control expressive humanoid maneuvers via intuitive natural motion. This work represents a significant advancement in robotics control systems and simulation fidelity.
Wire timeline
asRoBallet: Closing the Sim2Real Gap via Friction-Aware Reinforcement Learning for Underactuated Spherical Dynamics
Researchers have introduced asRoBallet, recognized as the first end-to-end reinforcement learning locomotion policy successfully deployed on a humanoid ballbot hardware platform. This study addresses significant challenges in transitioning from simulation to reality for underactuated spherical dynamics, specifically targeting gaps in contact modeling, actuator latency, and safe hardware exploration. The team developed a high-fidelity MuJoCo simulation that explicitly models the discrete roller mechanics of ETH-type omni-wheels, capturing previously ignored parasitic vibrations and contact discontinuities. Furthermore, they created a Friction-Aware Reinforcement Learning framework that masters coupled rolling, lateral, and torsional friction channels, enabling zero-shot Sim2Real transfer. The physical platform was constructed through subtractive reconfiguration, repurposing components from an overconstrained quadruped into a new structural frame to ensure robustness and low cost. Additionally, a generalized iOS ecosystem was developed to transform consumer electronics into a low-latency interface, allowing single operators to control expressive humanoid maneuvers via intuitive natural motion. This work represents a significant advancement in robotics control systems and simulation fidelity.
cs.AI updates on arXiv.org