PhyGile: Physics-Prefix Guided Motion Generation for Agile General Humanoid Motion Tracking
New framework eliminates retargeting artifacts, enabling robots to execute complex, text-described motions with physical stability.
A research team led by Jiacheng Bao has introduced PhyGile, a novel AI framework designed to solve a critical bottleneck in humanoid robotics: generating physically feasible motions from text descriptions. Current text-to-motion models are trained on human motion data, which assumes human biomechanics. When these motions are retargeted to robots with different mass and actuation, they often become physically impossible to execute. PhyGile closes this loop by performing physics-prefix-guided motion generation directly in a robot's native 262-dimensional skeletal space, bypassing the problematic retargeting step entirely and ensuring the generated trajectories respect real-world physics from the start.
Before adaptation, the system's General Motion Tracking (GMT) controller is trained using a curriculum-based mixture-of-experts scheme and post-trained on unlabeled motion data for robustness. The key innovation is the 'physics-prefix adaptation' phase, where the controller is fine-tuned with objectives derived from physics simulations. This teaches the AI the physical constraints of the real robot. The result is a system that can take a text prompt and generate a motion trajectory that is not only kinematically correct but also dynamically stable for a specific robot to perform, enabling agile and expressive whole-body movements previously unattainable.
- Eliminates retargeting artifacts by generating motions directly in a 262-dimensional robot-native skeletal space.
- Uses a 'physics-prefix' adaptation phase to fine-tune the controller with physics-derived constraints for real-world stability.
- Enables tracking of highly difficult, agile whole-body motions on real robots, far surpassing prior walking and low-dynamic capabilities.
Why It Matters
This bridges the sim-to-real gap for humanoids, enabling more dynamic, useful robots that can execute complex tasks described in natural language.