XNews

ASPIRE: Robot Skill Library Ships Know-How Instead of Weights, Speeds Transfer 10x

ASPIRE: Robot Skill Library Ships Know-How Instead of Weights, Speeds Transfer 10x

ASPIRE introduces a self-evolving robot skills library where coding agents search over control programs and distill successful know-how into an ever-growing repository, replacing gradient descent with skill refinement. It achieves sim-to-real and cross-embodiment transfer by shipping know-how rather than pixels or weights, sidestepping visual and physics gaps. Transfer learning tokens are cut up to ~10x. The system yields 150+ tasks and 90+ self-taught skills, with a full open-source release planned.

Source posts