Conventional robotic programming is tough to scale. It requires orchestrating multimodal notion, bodily contact dynamics, various configurations, and execution failures by hand. Code-as-policy methods let language fashions compose these into executable robotic packages. That makes robotic habits inspectable, editable, and debuggable.
However current robotic coding brokers run in naive execution environments. They obtain solely coarse, task-level suggestions. A failed rollout indicators that the duty failed, not why. The foundation trigger could be notion, movement planning, greedy, contact dynamics, or long-horizon coordination. These methods additionally discard fixes as soon as a job ends. So the agent fixing its hundredth job is not any extra skilled than at its first.
A crew of researchers from NVIDIA, College of Michigan, UIUC, UC Berkeley, and CMU introduces ASPIRE (Agentic Talent Programming by means of Iterative Robotic Exploration). It’s a continuous studying system that writes and refines robotic management packages. It additionally distills validated fixes right into a reusable, transferable ability library.
How ASPIRE works
ASPIRE runs an open-ended studying loop with three elements. It makes use of a coordinator–actor structure. A central coordinator manages the shared ability library and dispatches actor coding brokers to duties. Actors don’t trade full chat histories or uncooked trajectories. Solely distilled expertise transfer between them.
Closed-loop robotic execution engine: This replaces coarse rollout suggestions with per-primitive multimodal traces. For every notion, planning, and management name, it shops inputs, outputs, and return standing. It additionally shops RGB keyframes, overlays, grasp candidates, object poses, and motion-planning outcomes. The agent inspects solely the calls implicated by a failure. It then localizes the fault and validates a restore by means of re-execution.
Talent library: Reusable information is never a complete job program. So the library shops heterogeneous fixes. These embrace localization heuristics, notion prompts, greedy constraints, movement primitives, and debugging workflows. Every ability is compact in-context steering. It holds a failure signature, a when-to-apply situation, a restore technique, and infrequently a code sketch. The coordinator admits solely patterns that go debug validation and API-policy checks.
Evolutionary search: Hint-guided debugging alone can collapse into native restore loops. The agent retains patching the identical failed technique. To broaden exploration, ASPIRE proposes Okay candidate packages every spherical. Candidates situation on top-performing prior packages and their remaining failure traces. The following spherical explores distinct methods reasonably than refining one resolution.
In simulation, the coding agent is Claude Code with Claude Opus 4.6 and a 1M-token context window. Applications are written in CaP-X, an open-source code-as-policy framework constructed on MuJoCo Playground. The agent can’t learn simulator floor reality. Studying physics-engine state or asset information like .bddl, .xml, or .urdf is forbidden. The rule is straightforward. If an actual robotic with a digital camera may do it, it’s allowed.
Interactive Explainer
