UIUC's InterEvolve Lifts Unitree G1 Task Success From 8% to 86.5%

InterEvolve, a new UIUC paper on Hugging Face, reports that goal-conditioned success on humanoid loco-manipulation tasks climbed from 8% with human-written rewards to 86.5% after test-time evolution. The frozen controller never changes; only the reward program does.

The system treats each task as a reward program with staged rewards, completion conditions, and tunable constants. "A large language model (LLM) agent revises the program structure in context, drawing on execution feedback and a skill library of verified programs," the authors write, while CMA-ES tunes the constants in a nested inner loop. Candidate programs get verified across parallel simulation scenarios, and the ones that work land in the library the agent draws on next time.

The behavioral foundation model itself is an object-aware extension of BFM-Zero: trainable object residuals read object features into frozen body networks, trained on large-scale human-object interaction data. Reference-tracking object error dropped to 24.92 cm from 30.80 cm, success rose to 72%, and the authors clock full evolution at 2.1 GPU-hours per task.

"Human-designed rewards leave much of the FB model's loco-manipulation competence untapped, whereas the programs InterEvolve evolves release it, sometimes through novel strategies," the authors write.

Evolved skills run autonomously on a physical Unitree G1 using FoundationPose for pose estimation from onboard cameras, landing a week after Stanford's HomeBody wired a different stack to the same robot.