Robots that behave better by thinking longer with an LLM, not retraining.
InterEvolve
Test-Time Evolution of Reward Programs for Humanoid Loco-Manipulation.
01 / Input
Open-vocabulary text in, whole‑body loco‑manipulation out.
Type a task in plain words. An LLM agent turns it into a reward program, and a frozen controller carries it out, with no retraining.
02 / Method
Evolve the task, not the controller.
A task arrives as plain text, such as “Kick the box to the mark, feet only”. An LLM agent writes a staged reward program, a frozen controller runs it in parallel simulation, and a fixed verifier's feedback drives the next revision.
03 / Evolve
Evolve reward programs at test time.
04 / Discover
Discover unseen skills.
Novel use of skills from OMOMO.
05 / Adapt
Adapt fast from the skill library.
Large-box programs stored in the skill library evolve into the same skills for a plastic box, a small box, and a suitcase.
06 / Generalize
Generalize to novel goals and initializations.
Each selected program runs unchanged as the target moves or the box starts elsewhere.
07 / Compose
Compose strategies for complex scenarios.
08 / Arrange
Long-horizon box arrangement.
09 / Grasp
Grasp smaller objects.
Inspire hands.
10 / Deploy
Deploy autonomously on a real G1.
Onboard perception only, except where noted.
Kick after kick.
On a real G1, the same kicking program kicks the box in succession. Wherever one kick sends the box, the next kick follows it. Box pose from motion capture.