All research
Embodied AIActive research

Generative control for a physical embodiment.

Adapting pretrained motion structure to the joint topology, observations, and control cycle of the Unitree G1.

FSQLatent priorsUnitree G1

The question

How can reusable motion knowledge become calibrated, task-directed behavior on a specific humanoid?

The approach

Verify the encoder–quantizer–decoder contract, canonicalize joint ordering and reference frames, and condition a latent prior on proprioceptive state. Explore supervised adapter initialization and parameter-efficient reinforcement-learning refinement.

Validation focus

  • Decoded actions correspond to the intended joints and coordinate frames.
  • Actuator limits, scaling, and observation histories agree across the pipeline.
  • Task adaptation remains close to a valid pretrained motion distribution.

Current boundary

This is an ongoing integration workstream. Public benchmark results, physical deployment demonstrations, and performance claims are not presented here.

Start a conversation

Bring us a hard
learning problem.

Discuss a research collaboration, explore a training pipeline, or request a walkthrough of our work.

Discuss your project