NVIDIA ASPIRE Framework Improves Robot Learning

NVIDIA researchers introduced ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a novel continual learning system designed to write and refine robot control programs. This framework addresses the limitations of traditional robot programming and existing code-as-policy systems by providing a more granular feedback mechanism and a reusable skill library. ASPIRE operates through an open-ended learning loop with a coordinator-actor architecture, where a central coordinator manages a shared skill library and dispatches actor agents to perform tasks. Unlike previous systems, ASPIRE's actors do not share extensive histories but rather distilled skills, promoting efficiency and transferability.

The core innovation of ASPIRE lies in its closed-loop robot execution engine. This engine replaces coarse, task-level feedback with per-primitive multimodal traces. For every call to perception, planning, and control modules, it meticulously records inputs, outputs, return status, RGB keyframes, grasp candidates, object poses, and motion-planning results. When a task fails, the agent can inspect only the implicated calls, precisely localize the fault, and validate a repair through re-execution. This detailed feedback loop allows for much more targeted and effective error correction than simply knowing a task failed.

Furthermore, ASPIRE distills validated fixes into a reusable and transferable skill library. This library stores heterogeneous fixes, such as localization heuristics, rather than entire task programs. This approach ensures that the agent becomes more experienced with each task it completes, overcoming the issue of discarding learned fixes after a task ends. The system has demonstrated significant capabilities, achieving 31% zero-shot performance on long tasks within the LIBERO-Pro benchmark. This advancement represents a substantial step towards more adaptable and intelligent robotic systems capable of continuous self-improvement.