Reported by QbitAI (量子位), “Embodied AI's Skill moment! NVIDIA open-sources a robot skill library; Jim Fan: the paradigm has changed.”1(republished on ifeng3. Every key fact below is traced to a first-hand source (NVIDIA GEAR's official project page / the paper). Use the “Key facts” index on the right to jump by category.
🧠 Core tech & paradigm
NVIDIA's GEAR lab has open-sourced a robot skill library called ASPIRE (Agentic Skills Discovery for Robotics). Its core idea: have the robot distill every successful debugging fix into a named, executable, retrievable skill — so the next time a similar failure appears, it simply pulls the skill back out and reuses it.2
It follows the “Code as Policy” route: instead of having the model output robot actions directly, a large model writes executable robot control programs — so robot behaviour is visible code rather than implicit weights hidden inside a neural network.1
The system forms an open-ended learning loop from three parts: a closed-loop execution engine (records the multimodal trace of each perception, planning and grasping call), a continually expanding skill library (distills validated fixes into transferable knowledge), and an evolutionary search (generates many candidate programs in parallel and iterates toward the best).2
Training changes with it: multiple agents can each practise different skills, then merge their experience into one shared skill library — so the output of training is no longer just a pile of floating-point weights, but an ever-growing “sensorimotor skill library.”1
🏢 Lead org & team
The project is led by NVIDIA's GEAR lab — NVIDIA's team building foundation models for embodied agents across real and virtual worlds.2
Contributors also include labs at the University of Michigan, the University of Illinois Urbana-Champaign (UIUC), UC Berkeley, and Carnegie Mellon University (CMU).2
👤 Key people
Jim Fan (Linxi Fan) — NVIDIA's robotics lead and GEAR co-lead — frames this as a shift in the robot-learning paradigm: the trained “model” now corresponds not to a bundle of floating-point weights, but to a continually expanding robot skill library.1
Yuke Zhu co-leads the GEAR lab with Jim Fan and is a project lead on this work.2
The paper's equal-contribution first authors are Runyu Lu, Yubo Wu, and Ethan Kou, with Guanzhi Wang and others contributing, and Yuke Zhu and Jim Fan as project leads.2
📊 Results
On Robosuite's two-arm object handover task, ASPIRE lifted the success rate from 20% to 92% through iterative debugging.2
Skills learned on LIBERO-90 carry over to unseen long-horizon tasks: the richer the skill library, the higher the coding agent's success rate on LIBERO-Long.2
Evaluation spans Robosuite, LIBERO-Pro, and BEHAVIOR-1K, covering contact-heavy tabletop manipulation and long-horizon mobile manipulation with navigation.2