Researchers have introduced InterEvolve, a framework for test-time evolution in humanoid loco-manipulation. The work targets a demanding setting: solving tasks a controller was never trained for by repurposing its existing skills, improving from its own attempts, and retaining what it learns — all without retraining.
The central insight is that a broad controller already possesses much of the competence a new task requires. That competence becomes accessible through an interface between planning and control that is expressive enough to specify contact-rich, multi-stage interactions, yet executable and measurable enough for execution feedback to guide planning from experience.
InterEvolve builds this interface from two components. First, an object-aware forward-backward (FB) behavioral foundation model whose object residuals on a frozen body prior turn a new reward about the body or objects into loco-manipulation behavior at test time. Second, tasks are specified as reward programs: staged rewards with completion conditions and tunable constants.
An LLM agent revises the program structure in context, drawing on execution feedback and a skill library of verified programs, while a numerical optimizer tunes its constants. Each candidate is verified across parallel simulation scenarios, allowing the program to explore new ways to induce, repurpose, and compose the controller's existing motor competence and improve over iterations.
Experiments show that 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 approach also produces behaviors for diverse tasks, complex scenes, and long-horizon compositions in simulation, and evolved skills run autonomously on a physical Unitree G1 using egocentric onboard perception.