ASPIRE (robotics)

ASPIRE (paper title "ASPIRE: Agentic /Skills Discovery for Robotics"; the paper expands the acronym as Agentic Skill Programming through Iterative Robot Exploration) is a continual learning system for robot manipulation from NVIDIA's GEAR lab, built with researchers from the University of Michigan, the University of Illinois Urbana-Champaign, UC Berkeley and Carnegie Mellon University.^[1][3] A large language model coding agent writes robot control programs in the code-as-policy style, runs them, reads detailed execution traces to find out why they failed, repairs them, and saves the validated fixes as reusable text "skills" that later tasks can retrieve.^[1] The paper went on arXiv on 30 June 2026 (arXiv:2607.00272). An updated "v1.0" paper and the v1.0 open-source code release under the Apache License 2.0 were announced on 16 October 2026; v1.0 added results with newer frontier models and a real-robot demo that solves a Rubik's Cube.^[1][3][4][7]

The project leads are Runyu Lu, Yuke Zhu, Linxi "Jim" Fan and Guanzhi Wang.^[3] Jim Fan has described ASPIRE as "a new type of continual learning: 'training' is skill refinement instead of gradient descent", in which the "trained model" is "a repo of sensorimotor skills instead of floating weights".^[5][4]

Background

Code-as-policy methods have a language model write executable programs that call perception, planning and control functions, rather than outputting motor actions directly. The 2022 paper "Code as Policies" by Jacky Liang and colleagues is an early example.^[10] The ASPIRE paper groups itself with that line of work, which it contrasts with end-to-end vision-language-action models.^[1] Its main baseline is CaP-Agent0, from the CaP-X framework (arXiv:2603.22435), a benchmark and agent framework for code-as-policy robot manipulation whose authors include Letian Fu, Ken Goldberg, Yuke Zhu and Linxi "Jim" Fan.^[8] The paper says CaP-Agent0 "uses visual differencing, a predefined skill library, and per-episode test-time retries".^[1]

The authors argue that existing robotic coding agents get only coarse task-level feedback (a rollout either succeeded or failed) and throw away what they learned once a task is done, so that "the agent solving its hundredth task is effectively no more experienced than the agent solving its first".^[1] The skill-library idea follows earlier LLM-agent work such as Voyager, an open-ended Minecraft agent co-authored by Guanzhi Wang and Linxi Fan that ASPIRE cites.^[1][11] On X, Fan called ASPIRE the second work in a series he calls "Physical AutoResearch", following a project named ENPIRE.^[6]

Method

ASPIRE uses a coordinator-actor architecture. A central coordinator manages a shared skill library and sends actor coding agents to individual tasks, so several tasks can be learned in parallel. Actors do not share chat histories or raw rollouts with each other; what they learn passes between them only through the skill library.^[1] The system has three main components:^[1][2]

The paper's worked example is a BEHAVIOR-1K task in which a mobile robot must navigate to a radio and pick it up. The trace showed that perception found the radio, but repeated navigate_to_pose calls returned PLANNING_ERROR because the chosen target poses fell inside the table's collision-avoidance buffer. The agent wrote a routine that tries other approach angles around the object, and the coordinator saved this as a reusable "Multi-Angle Approach" navigation skill rather than as a radio-specific program.^[1][3]

The v1.0 paper adds a curriculum step: the coordinator picks which task to try next based on the current library. Once every task is done, it reruns low-scoring early tasks with the finished library, which the paper calls backward skill transfer.^[3]

Evaluation setup

In the June paper, every simulation experiment used Claude Code with Claude Opus 4.6 and a 1M-token context window as the coding agent. The agent wrote Python programs against the CaP-X framework, which is built on MuJoCo Playground.^[1] The v1.0 paper no longer names CaP-X; it says the agent writes programs in "ASPIRE's unified code-as-policy harness", with one set of perception, geometry and motion-planning APIs shared across all benchmarks. In v1.0, Claude and open-weight models ran through Claude Code and GPT models through the Codex CLI.^[3]

ASPIRE learns on a small set of "debug" seeds and then reports success on a separate set of held-out seeds, using one generated program per task for LIBERO-Pro and Robosuite. CaP-Agent0, by contrast, writes a new program for each seed with test-time reasoning and retries.^[1] The benchmarks were:^[1]

LIBERO-Pro is a 2025 extension of the LIBERO benchmark. It perturbs objects, initial states, task instructions and environments; its authors found that models scoring over 90% on standard LIBERO collapsed to 0% under these perturbations.^[9] The vision-language-action baselines were OpenVLA, π0 and π0.5.^[1]

Results with Claude Opus 4.6 (June 2026)

All of these results come from the authors' own evaluations.

LIBERO-Pro (macro-averaged success, 10 tasks per suite, 50 held-out seeds):^[1]

Averaged across the two perturbation types, the paper reports gains over the strongest baseline of 77 points on Object, 41.5 on Goal and 42.5 on Spatial.^[1]

Robosuite (100 held-out trials per task):^[1]

BEHAVIOR-1K (navigation / task success, 25 held-out seeds): for soda-can pickup, human-written programs scored 80% / 72%, CaP-Agent0 84% / 72% and ASPIRE 92% / 88%. For radio pickup, the scores were 88% / 36%, 80% / 56% and 100% / 88% respectively.^[1]

Zero-shot transfer. With a library built from all 90 LIBERO-90 tasks and no further debugging, ASPIRE averaged 23% on LIBERO-Pro Long under position perturbations and 38% under task perturbations (30.5% overall). CaP-Agent0 scored 3.8% and π0.5 5% overall. Success rose as the library grew from 0 to 25, 50 and 90 source tasks (4.7%, 13.7%, 21.5%, 30.5% overall). Individual tasks moved non-monotonically, which the authors attribute to stale or misleading library entries.^[1]

Ablations. On LIBERO-Pro, the zero-shot agent with neither the execution engine nor evolutionary search averaged 14%. Adding the execution engine raised this to 62%, and adding evolutionary search raised it to 72%.^[1]

Real-robot experiments

Cross-embodiment skill transfer. The authors took three skills that ASPIRE learned in Franka-arm simulation (soda-can pickup, bowl-on-plate placement, drawer push/pull) and gave them as in-context guidance to a coding agent controlling a real bimanual YAM station. That robot has a different body and a different API. The agent was OpenAI Codex with GPT-5.5 in "reasoning-xhigh" mode. The paper stresses that this is not direct policy deployment: the real-robot agent still has to debug through real-world feedback.^[1]

For the drawer, the no-skill run used up its token budget without producing a successful program.^[1] The project page's headline demo is cross-embodiment long-horizon table bussing, with Claude Opus 4.6 in simulation and GPT-5.5 on the real robot.^[2][4]

Online learning: Rubik's Cube (v1.0). In v1.0 the coding agent runs as a "System 2" on the real robot. It connects through the Model Context Protocol to a code-as-policy runtime that exposes observation, object-localization, end-effector, joint and gripper tools.^[3] In one GPT-6 Astra session on the YAM platform with soft grippers, the agent was given a randomly scrambled cube and human-written task guidance. It identified the cube's colors from camera images, called the Kociemba solver for a move sequence, and then had to carry out and verify the physical turns, writing down what it learned as text skills.^[3] Across three consecutive episodes:^[3]

Estimated model cost per turn, at OpenAI API prices, fell from $16.85 in the first episode to $2.89 in the last.^[3] The README also shows GPT-6 Astra unscrewing a bottle cap, putting a pill in a bottle and screwing on the cap, and stacking cups upside down. It notes that the published recipes for these three tasks were "distilled from demo sessions, not a new real-robot validation".^[4]

Model comparison (v1.0)

The v1.0 paper ran ASPIRE with twelve different language models, keeping the environment, robot API, seeds and protocol fixed. The test was iterative debugging, without evolutionary search and starting from an empty skill library, on the 10 LIBERO-Pro Goal tasks under position perturbation. Costs are estimates at list API prices.^[3][2]

The paper concludes that "stronger models are not necessarily more expensive": Sonnet 5.5 and Opus 5.5 each reached 99% for an estimated $121 and $183 across the ten tasks, compared with $597 for Opus 5 at 91%.^[3]

The paper also compared Opus 5, Opus 5.5 and GPT-6 Astra against Opus 4.6 across all of LIBERO-Pro, this time carrying the skill library across tasks. With iterative debugging alone, average success on Object, Goal and Spatial was 96% for Opus 5, 98% for Astra and 98.9% for Opus 5.5, compared with 62% for Opus 4.6. On LIBERO-Pro Long, the scores were 97.9% for Opus 5.5, 87.9% for Astra and 74.6% for Opus 5 (82.2% after targeted reruns). On Robosuite, the seven-task mean rose from 81% with Opus 4.6 to 99% with Opus 5 and Astra and 100% with Opus 5.5. The v1.0 paper reports that Opus 5.5 reached 100% navigation and task success on both BEHAVIOR-1K tasks, although the project page says Opus 5.5 was not evaluated on BEHAVIOR-1K.^[3][2] The paper summarizes this as 99% average success across all 80 LIBERO-Pro tasks and 100% on representative Robosuite and BEHAVIOR-1K tasks with Claude Opus 5.5.^[3] With these stronger models, evolutionary search changed Object, Goal and Spatial success by at most 2 points, compared with up to 34 points for Opus 4.6.^[3]

A case study found that stronger models switch strategies instead of repeating the default grasp-and-move. To slide a book under a low shelf, Opus 5.5 knocked the book flat without grasping it and pushed it under (68% held-out success). Astra rolled the book flat, dropped it and pushed it (22%). Opus 4.6 and Opus 5 scored 0% and 2%. The human demonstration's tilt-and-slide move was not found by any model.^[3]

Open-source release

The GitHub repository (NVlabs/ASPIRE) was created on 8 July 2026; its commit history shows a "Public release" commit on 21 August 2026 and an "ASPIRE v1.0 release" commit on 15 October 2026. Its README marks the October 2026 entry as "ASPIRE v1.0 is out!"^[4] NVIDIA-owned material is released under the Apache License 2.0; third-party components keep their own licenses.^[4] According to the README, ASPIRE works with any coding agent. The simulation workflow ships as two agent skills, iterative-debugging and evosearch, packaged for both Claude Code and Codex (see Agent Skills). Setup needs CUDA, uv and a Hugging Face token for the gated SAM3 weights.^[4] The README warns that ASPIRE "executes language-model-generated Python with full import access" and that its trial processes "are not a security sandbox". It says generated code should run on an isolated host, and a simulation agent should never be given access to physical hardware.^[4]

Claims around the v1.0 release

Jim Fan announced the open-source v1.0 on X on 16 October 2026. He wrote that "on the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night", that "Opus 5.5 slightly edges out Astra on our robot simulation bench", and that "Sonnet 5.5 hits the Pareto frontier of cost vs. robot capability, $12/task". He also said that "Astra was a 0->1 tsunami on many tasks" and invited readers to "feel the Physical Bitter Lesson at home" (see The Bitter Lesson).^[7]

The paper's tables support some of these numbers. Sonnet 5.5's $12.11 per task at 98.8% success is the $12 figure. Opus 5.5 scores above Astra on most, but not all, of the paper's model comparisons. Astra is slightly ahead on LIBERO-Pro Object under position perturbation (99.0% vs 98.2%), and Opus 5.5's largest suite-average lead is on LIBERO-Pro Long (97.9% vs 87.9%).^[3] The paper's headline 99% figures are for Claude Opus 5.5. Astra's reported results are 94.4% on the ten-task model comparison, about 98% on Object, Goal and Spatial, 87.9% on LIBERO-Pro Long, a 99% seven-task Robosuite mean and 96% on both BEHAVIOR-1K tasks.^[3]

When ASPIRE was first announced in June, Fan said the website hosted a gallery of "150+ tasks and 90+ skills the robots taught themselves". He also claimed that ASPIRE "achieves up to ~10x cut in 'transfer learning' tokens"; the paper's largest reported reduction is on the soda-can task, from 61.94M to 6.58M total tokens (about 9.4 times).^[5][1]

Limitations

The authors list five limitations:^[1][3]

• Not a fully autonomous real-world learner. Real deployment still needs robust success detection, safe resets, safety monitoring and calibration upkeep, all of which are cheap and programmatic only in simulation.
• Dependence on a frozen frontier LLM. The June paper says the authors "have not verified that smaller or weaker LLMs can sustain the same debugging loop". In v1.0, this limitation now reads that weaker models (GPT-6 Luna, Claude Opus 4.6, Qwen 3.8 Max, GPT-6 Sol) reach only 15% to 59% on the ten-task comparison, the best open-weight model (Kimi K3) reaches 78%, and the best results use proprietary frontier models.
• A fixed primitive API. The predefined perception, planning and control primitives keep debugging tractable and safe, but limit which behaviors the agent can express.
• Incomplete long-term memory management. As the library grows, entries can become stale, overly specific, redundant or misleading.
• Compute cost. The debugging and evolutionary-search loop uses many LLM calls and rollouts per task.

Authors

The v1.0 paper lists sixteen authors: Runyu Lu, Yubo Wu, Ethan Kou, Haotian Lin (equal contribution), Letian Fu, Wenli Xiao, Ajay Mandlekar, Yinzhen Xu, Guanya Shi, Ken Goldberg, Ang Chen, Mosharaf Chowdhury, Jan Kautz, Yuke Zhu, Linxi "Jim" Fan and Guanzhi Wang. Lu, Zhu, Fan and Wang are project leads.^[3] The arXiv v1 of 30 June 2026 lists fourteen authors; Haotian Lin (thanked in the v1 acknowledgments) and Jan Kautz were added later.^[1][3] The project page's team list gives Letian Fu as "Max Fu".^[2] The acknowledgments thank NVIDIA GEAR, the UMich SymbioticLab, UsesysLab, the UC Berkeley AUTOLab and the CMU LeCAR Lab.^[3]

See also

• Embodied AI
• Physical AI
• Robot learning
• Sim-to-real transfer
• AlphaEvolve

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25Runyu Lu et al., "ASPIRE: Agentic /Skills Discovery for Robotics", arXiv:2607.00272 (v1, 30 June 2026). arxiv.org/...2607.00272
2. ^1 ^2 ^3 ^4 ^5NVIDIA GEAR, "ASPIRE: Agentic /Skills Discovery for Robotics" project page. research.nvidia.com/...aspire
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22Runyu Lu et al., "ASPIRE: Agentic /Skills Discovery for Robotics" (v1.0 paper PDF, October 2026), NVIDIA. research.nvidia.com/...aspire.pdf
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8NVlabs, "ASPIRE" GitHub repository and README. github.com/...ASPIRE
5. ^1 ^2Jim Fan (@DrJimFan), post introducing ASPIRE, X, 30 June 2026. x.com/...2072004190856212902
6. ^Jim Fan (@DrJimFan), "ENPIRE -> ASPIRE, our 2nd work in the series for Physical AutoResearch", X, 1 July 2026. x.com/...2072337488782475390
7. ^1 ^2Jim Fan (@DrJimFan), post announcing the ASPIRE v1.0 open-source release, X, 16 October 2026. x.com/...2111128797181567466
8. ^Letian Fu et al., "CaP-X: A Framework for Benchmarking and Improving Coding Agents for Robot Manipulation", arXiv:2603.22435. arxiv.org/...2603.22435
9. ^Xueyang Zhou et al., "LIBERO-PRO: Towards Robust and Fair Evaluation of Vision-Language-Action Models Beyond Memorization", arXiv:2510.03827. arxiv.org/...2510.03827
10. ^Jacky Liang et al., "Code as Policies: Language Model Programs for Embodied Control", arXiv:2209.07753. arxiv.org/...2209.07753
11. ^Guanzhi Wang et al., "Voyager: An Open-Ended Embodied Agent with Large Language Models", arXiv:2305.16291. arxiv.org/...2305.16291