Event

ASPIRE Enables Robots to Autonomously Write and Iterate Control Programs

1 reports · 1 sources

Runyu Lu and other researchers propose ASPIRE, a continual learning system that lets robots autonomously write and iterate control programs under a "code-as-policy" paradigm, accumulating experience into a reusable skill library. The paper reports improvements of up to 77% over existing methods on perturbed manipulation tasks, 72% on the Robosuite bimanual handover task, and up to 32% on long-horizon household tasks; on a representative unseen long-horizon task, the success rate is 31%, higher than the 4% of existing methods. When using Claude Opus 5.5, ASPIRE achieves an average success rate of 99% across all 80 LIBERO-Pro tasks, and reaches 100% on some Robosuite and BEHAVIOR-1K tasks. The researchers also state that skills learned in simulation can reduce the cost of programming real robots, and that skills acquired through online interaction with a physical robot enable ASPIRE to progressively solve the Rubik's Cube with increasing efficiency.