@dair_ai
NEW paper from NVIDIA. They discuss robot programming that compounds experience instead of throwing it away. Traditional robot programming forces you to orchestrate perception, contact dynamics, diverse configurations, and constant execution failures by hand. Most learned approaches then bury what they learned in opaque weights. ASPIRE runs a code-as-policy loop that autonomously writes and refines control programs. A closed-loop execution engine exposes fine-grained multimodal traces, so the system diagnoses its own failures, synthesizes repairs, and validates them. Validated fixes distill into a reusable skill library, and evolutionary search explores diverse task sequences beyond single-trajectory tuning. ASPIRE gains up to 77 percent on LIBERO-Pro under perturbation, 72 percent on Robosuite bimanual handover, and 32 percent on BEHAVIOR-1K. On LIBERO-Pro Long it hits 31 percent zero-shot versus 4 percent for prior methods, with early sim-to-real transfer across embodiments. Paper: https://t.co/zDzVNC2NXn Learn to build effective AI agents in our academy: https://t.co/LRnpZN7L4c