HF Paper InterEvolve Lets Humanoid Robots Evolve Reward Programs at Test Time, Demoed on Unitree G1

InterEvolve pairs an object-aware behavioral foundation model with reward programs refined at test time by an LLM agent, allowing a humanoid to repurpose existing skills and retain what it learns without retraining. The authors demonstrate the system on simulated scenarios and a physical Unitree G1, arguing pre-trained controllers contain latent competencies that standard reward design does not unlock.