🤖 Robot Learning Breakthrough: Genius Fixes! ✨

SYSTEMATIC ROBOT PROGRAMMING THROUGH CONTINUOUS LEARNING
The current approach to robot programming is often inefficient, requiring extensive manual orchestration of multimodal perception, physical contact dynamics, and diverse configurations. Code-as-policy systems offer a potential solution by allowing language models to compose executable robot programs, enabling inspectability, editability, and debuggability. However, existing robotic coding agents typically operate within naive execution environments, receiving only coarse, task-level feedback, which hinders root cause analysis.

THE CHALLENGES OF NAIVE ROBOT EXECUTION
Existing robotic coding agents rely on coarse rollout feedback, signaling only task failure without pinpointing the underlying cause – which could stem from perception, motion planning, grasping, or long-horizon coordination. Furthermore, these systems discard fixes once a task concludes, preventing the agent from learning and improving over time. A core issue is the lack of persistent experience, where an agent solving its hundredth task is no more knowledgeable than when it began.

INTRODUCING ASPIRE: AGENTIC SKILL PROGRAMMING
Researchers at NVIDIA, University of Michigan, UIUC, UC Berkeley, and CMU have developed ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a continual learning system designed to write and refine robot control programs. ASPIRE distills validated fixes into a reusable, transferable skill library, streamlining the learning process. This system utilizes a coordinator-actor architecture to manage and deploy skills effectively.

ASPIRE’S CORE COMPONENTS: A CONTINUOUS LEARNING LOOP
ASPIRE operates through a three-component continuous learning loop. First, a central coordinator manages a shared skill library and dispatches actor coding agents to specific tasks. Second, actors exchange only distilled skills, avoiding the transfer of full chat histories or raw trajectories, thereby minimizing computational overhead. Finally, a closed-loop robot execution engine provides per-primitive multimodal traces, capturing detailed inputs, outputs, and return statuses for each perception, planning, and control call.

MULTIMODAL TRACES AND DETAILED INSPECTION
This execution engine replaces coarse rollout feedback with a rich dataset of multimodal traces, including RGB keyframes, grasp candidates, object poses, and motion-planning results. The agent inspects only the calls implicated by a failure, localizes the fault, and validates a repair through re-execution. This targeted approach significantly reduces the time and effort required for debugging.

THE SKILL LIBRARY: A REPOSITORY OF REUSABLE FIXES
The skill library within ASPIRE stores heterogeneous fixes—localization heuristics, perception prompts, grasping constraints, motion primitives, and debugging workflows. Each skill is compact, providing in-context guidance, and incorporates a failure signature, a when-to-apply condition, a repair strategy, and often a code sketch. The coordinator admits only patterns that pass rigorous debug validation and API-policy checks, ensuring the quality and reliability of the skill library.

EVOLUTIONARY SEARCH: BROADENING THE EXPLORATION SPACE
To mitigate the risk of local repair loops, ASPIRE employs evolutionary search. The agent generates K candidate programs each round, conditioned on top-performing prior programs and their remaining failure traces. This strategy promotes exploration of distinct strategies rather than refining a single solution, accelerating the learning process.

SIMULATION ENVIRONMENT AND CODING AGENT
The coding agent within ASPIRE is Claude Code with Claude Opus 4.6 and a 1M-token context window. Programs are written in CaP-X, an open-source code-as-policy framework built on MuJoCo Playground. A critical constraint is that the agent cannot directly access simulator ground truth, preventing reliance on pre-programmed knowledge. Only actions that a real robot with a camera could perform are permitted.

THE BEHAVIOR-1K TASK: A TEST CASE FOR ASPIRE
Consider the BEHAVIOR-1K task, where a robot must pick up a radio near a table. Repeated navigate_to_pose calls fail, with the target goal located within approximately 20 centimeters of the table edge, resulting in a PLANNING_ERROR from cuRobo. The agent analyzes the trace, identifies the failure as target infeasibility (not perception or grasping), and then writes a repair that samples standoff poses around the radio.

REUSABLE NAVIGATION-RECOVERY SKILL
This repair, where one side of the object is blocked, another is often open (e.g., a 180-degree pose clearing the buffer), is validated and admitted as a reusable navigation-recovery skill. ASPIRE demonstrates the ability to transfer skills accumulated on LIBERO-90, achieving approximately 31% success on held-out LIBERO-Pro Long tasks, a significant improvement over prior methods that saturate near 4%.

EVALUATION AND COMPARATIVE RESULTS
ASPIRE’s performance is evaluated across three benchmark families: LIBERO-Pro, Robosuite, and BEHAVIOR-1K. The primary coding-agent baseline is CaP-Agent0, which utilizes visual differencing, a predefined skill library, and per-episode test-time retries. Comparative analyses also include end-to-end vision-language-action policies: OpenVLA, π0, and π0.5.

LIBERO-PRO PERFORMANCE
On LIBERO-Pro, ASPIRE achieves up to 77 points on the Object suite, averaging both perturbation axes over the strongest baseline. Gains are also observed on Goal (41.5 points) and Spatial (42.5 points).

ROBUSUITE PERFORMANCE
In Robosuite, bimanual handover rises from 20% to 92%.

BEHAVIOR-1K PERFORMANCE
On BEHAVIOR-1K, the radio pickup task improves from 56% to 88%.

REAL-WORLD VALIDATION AND SKILL TRANSFER
The research team tests three simulation-discovered skills on a real bimanual YAM station using OpenAI Codex GPT-5.5. The embodiment and API differ from simulation, and transferred skills reduce debugging cost. Specifically, soda-can lifting improves from 13/20 to 19/20, while drawer opening moves from 0/20 to 11/20, where the no-skill baseline never succeeded.