Overview
RobotDancing is a reinforcement learning (RL) framework designed to achieve robust long-horizon, high-dynamic motion tracking on humanoid robots. The system addresses the challenge of dynamic inconsistency between kinematically plausible retargeted reference motions and the actual capabilities of the robot. By employing a residual-action approach, the policy learns to compensate for discrepancies such as actuation limits, latency, friction, and inertia rather than re-synthesizing the entire motion.
Research Context
The problem of robustly tracking long-horizon, high-dynamic motions on humanoids is characterized by inherent brittleness. Retargeted reference motions, while often kinematically sound, frequently exhibit dynamic inconsistencies when applied to a robot. These inconsistencies can lead to the accumulation of small tracking errors, which ultimately compromise control stability and lead to failure in motion execution.
Approach
RobotDancing employs a single-stage reinforcement learning recipe. A key aspect of its methodology is the parameterization of actions as residual corrections. Specifically, the policy predicts reference-conditioned residual joint targets. This means the policy's primary function is to compute adjustments on top of the provided reference motion, allowing it to focus on mitigating the dynamic mismatch between the reference motion and the robot's physical characteristics. This residual action formulation is intended to enable the policy to compensate for factors such as actuation limits, system latency, friction effects, and the robot's inertia.
The training regimen for RobotDancing involves training one policy per reference sequence. Despite this per-sequence training, a common training and deployment recipe is reused across different motions and various robot platforms. This standardization aims to streamline the development and deployment process.
Findings
- RobotDancing was evaluated using eight LAFAN1 dance motions.
- Primary evaluation was conducted on the Unitree G1 humanoid robot.
- The resulting G1 policies were able to execute long-horizon, high-energy behaviors on hardware.
- These policies operated without requiring test-time modifications.
- Cross-platform feasibility was qualitatively illustrated through experiments on H1 and H1-2 robot platforms.
Why This Matters
The ability to track complex, dynamic motions robustly over extended periods without test-time modifications could enhance the versatility and performance of humanoid robots. Addressing the dynamic inconsistencies between motion references and robot hardware is a foundational step towards more agile and reliable robotic systems in various operational contexts.