NVIDIA AI Launches ASPIRE: A Self-Improving Robotic Framework Achieves 31% Zero-Shot on LIBERO-Pro Long Tasks

Programming traditional robots is difficult to scale. It requires multimodal vision planning, physical communication skills, multitasking, and manual dexterity. Coding systems as a policy allow language models to compile these into usable robotic programs. That makes the robot’s behavior testable, programmable, and adjustable.

But robotic coding agents work in nonsensical use cases. They only get solid, job-grade feedback. A failed release indicates that the job failed, not why. The root cause can be vision, movement planning, grasping, communication skills, or long-term communication. These systems also discard repairs when the job is done. So an agent who solves his hundredth task is less knowledgeable than he was at the beginning.

A team of researchers from NVIDIA, University of Michigan, UIUC, UC Berkeley, and CMU presented ASPIRE (Agentic Skill Programming through Iterative Robot Exploration). It is a continuous learning program that writes and refines robotics control systems. It also integrates proven fixes into a reusable, transferable skills library.

How ASPIRE works

ASPIRE uses a three-part open learning loop. It uses a coordinator–actor architecture. A central coordinator manages a library of shared skills and deploys coding agents to tasks. The characters don’t exchange full stories of dialogue or raw ways. Only distilled abilities flow through them.

A closed-loop robot performance engine: This replaces coarse-grained output feedback with classical multimodal tracking. For each view, schedule, and control call, it stores input, output, and return status. It also stores RGB keyframes, overlays, grip candidates, object orientation, and motion editing effects. The agent only checks calls that are affected by the failure. It then locates the error and verifies the fix by redoing it.

A library of skills: Re-usable information is rarely a system-wide system. Therefore the library maintains various preparations. These include localization heuristics, detection commands, hold limits, motion primitives, and debugging workflows. Each skill is a unified guide to content. It contains the signature of the failure, the application criteria, the remediation strategy, and usually the code diagram. The connector only accepts patterns that pass debug validation and API policy checks.

Natural search: Guided debugging alone can fall into local debugging loops. The agent keeps repeating the same failed strategy. To extend the evaluation, ASPIRE proposes K candidate programs each cycle. Students are in the context of previous high-performance programs and their remaining tracks of failure. The next round explores different strategies rather than refining a single solution.

In the simulation, the code agent is Claude Code with Claude Opus 4.6 and a 1M token context window. The programs are written in CaP-X, an open source code-as-policy framework developed at MuJoCo Playground. The agent cannot learn the simulated ground truth. Reading the state of the physics-engine or similar property files .bddl, .xmlor .urdf it is forbidden. The rule is simple. If a real robot with a camera can do it, it’s allowed.

Interactive Descriptor