From Text to Torque
2026 — CS 224R · diffusion, PPO, MuJoCo Warp, Unitree G1
In March 2026, NVIDIA released Kimodo, a motion-diffusion model that turns a plain-English sentence into a full-body motion sequence. I saw the demo on Twitter and couldn’t let go of the obvious next question: if a model can imagine a motion from a sentence, can a robot learn to actually perform it?
Reference motions for humanoid policies usually come from people. At Stanford’s Movement Lab (Prof. Karen Liu), systems like TWIST retarget the movements of an operator wearing a motion-capture suit onto the robot — remarkable work, but every new behavior needs a human to suit up and perform it. Generating the reference from language takes the suit out of the loop: describe the motion, and the pipeline does the rest.
From Text to Torque is that pipeline, end to end. A natural-language command goes to Kimodo, which generates a kinematic reference motion; a PPO policy is then trained in MuJoCo Warp — across 1,024 parallel environments — to track that reference on a simulated Unitree G1 humanoid, turning imagined motion into torques.
Results
The interesting problem was early termination. Strict termination thresholds kill exploration before the policy can learn — default training survived an average of 13.5 steps on a squat task. Adaptive termination calibration, which tightens the thresholds as tracking improves, recovered full 250-step squat tracking.