NVIDIA ASPIRE turns robot debugging into a reusable skill library

MarkTechPost reports that NVIDIA introduced ASPIRE, a self-improving robotics framework for code-as-policy agents. The system analyzes failed robot executions, refines control programs, and distills fixes into a reusable skill library instead of discarding them after a task. In the reported LIBERO-Pro evaluation, ASPIRE reached 31% zero-shot success on long-horizon tasks, far above prior methods that were near 4%.

Key details: ASPIRE writes and refines robot-control programs, Failed rollouts are turned into reusable skills, NVIDIA reports 31% zero-shot success on LIBERO-Pro long tasks.

Why it matters: Robotics agents only become useful if they learn from real execution failures, and reusable skills are a practical route beyond one-task demos.