{"slug": "aspire-agentic-skills-discovery-for-robotics", "title": "Aspire: Agentic /Skills Discovery for Robotics", "summary": "NVIDIA Labs released ASPIRE v1.0 in October 2026, a continual-learning robotics system that refines sensorimotor skills instead of using gradient descent, achieving 99% success on 80 LIBERO-Pro Goal, Object, Spatial and Long tasks and solving a Rubik's Cube. On 10 LIBERO-Pro tasks, ASPIRE success rises from 15% with GPT-6 Luna to 99% with Claude Opus 5.5, and Claude Opus 5.5 reaches 97–100% success across LIBERO-Pro, Robosuite and BEHAVIOR-1K. ASPIRE is coding-agent agnostic, packaged as skills for Claude Code and Codex, and is installed by cloning the NVlabs/ASPIRE repository and running the LIBERO-Pro Goal-Swap Quick Start.", "body_md": "ASPIRE is a new type of continual learning: \"training\" is skill refinement instead of gradient descent.\n\n\"Trained model\" is a repo of sensorimotor skills instead of floating weights.\n\n“Distributed training” is a panel of agents each practicing a different skill instead of sharded minibatches.\n\n- \n**[2026-10] ASPIRE v1.0 is out!** ASPIRE can now solve a Rubik's Cube, and it achieves 99% success on 80 LIBERO-Pro Goal / Object / Spatial / Long tasks and selected Robosuite, BEHAVIOR-1K tasks.\n- \n**[2026-06] ASPIRE is out!** We achieve SOTA results on LIBERO-Pro, Robosuite, and BEHAVIOR-1K with Claude Opus 4.6! Also check out our table bussing sim-to-real demos!\n\n| Cross-embodiment sim-to-real | Online real-world learning | \n|---|---|\n| table-bussing-sim-to-real-hq.mp4 | rubiks-cube-astra-hq.mp4 | \n| [Long Horizon Table Bussing](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/real/README.md)Claude Opus 4.6 (sim), GPT-5.5 (real) | [Rubik's Cube](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md#solve-a-rubiks-cube)GPT-6 Astra | \n\n| stack-cups.mp4 | pill-in-tube.mp4 | unscrew-cap.mp4 | \n| [Stack Cups](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md#inverted-cup-nesting)GPT-6 Astra | [Unscrew Bottle Cap](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md#unscrew-and-pour)GPT-6 Astra | [Pill in Bottle](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md#pill-insertion-and-bottle-closure)GPT-6 Astra | \n\n| Success vs. API cost | LIBERO-Pro, Robosuite, BEHAVIOR-1K | \n|---|---|\n| Stronger models discover better strategies |  | \n| case-study-book.mp4 | case-study-push.mp4 | \n| **Book under a shelf** Opus 5.5 and Astra let go of the book and push it under the low shelf. | **Push a plate** Opus 5.5 clears the clutter first; Astra lifts the plate over it. | \n| case-study-drawer.mp4 | case-study-mug.mp4 | \n| **Open the bottom drawer** Opus 5.5 and Astra tilt the gripper and hook the low handle, as the human does. | **Mug in the microwave** Opus 5.5 and Astra grasp the handle and slide the mug in upright. | \n\nASPIRE improves with the underlying model. On 10 LIBERO-Pro tasks, success rises from 15% (GPT-6 Luna) to 99% (Claude Opus 5.5), and the strongest models are not the most expensive: the green line marks the cost/success frontier. Across LIBERO-Pro, Robosuite and BEHAVIOR-1K, Claude Opus 5.5 reaches 97–100% success on every task.\n\nTo get started with ASPIRE, launch a coding agent such as Codex or Claude Code and enter the following prompt:\n\n```\nClone the main branch of https://github.com/NVlabs/ASPIRE/,\nread the iterative-debugging skill (.claude/skills/iterative-debugging/SKILL.md) and\ndocs/libero-pro.md, and run the complete ASPIRE LIBERO-Pro Goal-Swap Quick Start for\nall ten tasks in the libero_goal_swap suite.\n\nBefore executing, report the required GPUs, credentials, gated weights,\nservices, expected runtime, seed partitions, and output paths. Wait for\nmy confirmation before launching. Do not access real-robot code or instructions\n(AGENTS.md, docs/system2*.md, docs/yam-real.md, src/cap_harness/yam_real/).\n```\n\nBefore running, set up the host (CUDA, uv, and the Hugging Face token for the gated SAM3 weights) with [`docs/host-setup.md`](https://github.com/NVlabs/ASPIRE/blob/main/docs/host-setup.md).\n\n**Reference agent environments:** ASPIRE is coding-agent agnostic. The simulation workflow is packaged as agent skills for Claude Code ([`.claude/skills/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills)) and Codex ([`.codex/skills/`](https://github.com/NVlabs/ASPIRE/blob/main/.codex/skills)): `iterative-debugging` and `evosearch`. [` AGENTS.md`](https://github.com/NVlabs/ASPIRE/blob/main/AGENTS.md) holds the execution rules for the real-robot System 2 agent ([operator runbook](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md)) and is not needed for simulation runs.\n\nName the suite and experiment explicitly. If neither is named, the agent should present this table and stop for selection.\n\n| Suite | Experiment | Runbook | Video | \n|---|---|---|---|\n| LIBERO-Pro | Iterative-debugging loop | [`iterative-debugging/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills/iterative-debugging/SKILL.md) | fix-loop-iterative-debugging.mp4 | \n| Robosuite | Iterative-debugging loop | [`iterative-debugging/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills/iterative-debugging/SKILL.md) |  | \n| BEHAVIOR-1K | Iterative-debugging loop | [`iterative-debugging/behavior/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills/iterative-debugging/behavior/SKILL.md) |  | \n| LIBERO-Pro | Evolutionary Search | [`evosearch/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills/evosearch/SKILL.md) | Evolutionary_search.mp4 | \n| Robosuite | Evolutionary Search | [`evosearch/`](https://github.com/NVlabs/ASPIRE/blob/main/.claude/skills/evosearch/SKILL.md) |  | \n| LIBERO | Zero-Shot Transfer | [`libero/zeroshot-transfer/`](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/sim/.claude/libero/zeroshot-transfer/INSTRUCTIONS.md) | Skill_Library_Scaling_ZeroShot_Transfer.mp4 | \n| LIBERO-Long-Pro | Library-Size Scaling | [`libero/library-size-scaling/`](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/sim/.claude/libero/library-size-scaling/INSTRUCTIONS.md) |  | \n| LIBERO-Long-Pro | Inference-Time Scaling | [`libero/inference-time-scaling/`](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/sim/.claude/libero/inference-time-scaling/INSTRUCTIONS.md) |  | \n| Robosuite | Debugging-Token-Ablation | [`robosuite/training-law/`](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/sim/.claude/robosuite/training-law/INSTRUCTIONS.md) |  | \n\nBefore a paper-scale launch, the agent must report the selected tasks, seed schedule, expected trial count and runtime, GPU and credential requirements, services, and output paths, then wait for explicit confirmation.\n\nUse the [System 2 operator runbook](https://github.com/NVlabs/ASPIRE/blob/main/docs/system2.md#reproduce-the-manipulation-demos)\nto prepare a calibrated YAM station and launch a fresh agent with the appropriate\ntext-based task skill:\n\nThe [manipulation skill entrypoint](https://github.com/NVlabs/ASPIRE/blob/main/.agents/skills/manipulation-tasks/SKILL.md)\nroutes the agent to these recipes. They provide grasp strategies, recovery, and\nphysical completion checks without requiring episode logs or fixed-coordinate\nreplays. The three new recipes are distilled from demo sessions, not a new\nreal-robot validation; the pill recipe explicitly distinguishes a seated cap from\nverified final tightness. Re-ground the current scene and verify the actual outcome.\n\nWarning\n\nASPIRE executes language-model-generated Python with full import access. Trial processes and watchdogs are not a security sandbox. Run generated code on an isolated host without credentials or sensitive mounts, restrict network access, and never grant a simulation agent access to physical hardware. Real-robot work requires the controls in [`aspire/real/AGENTS.md`](https://github.com/NVlabs/ASPIRE/blob/v0.1/aspire/real/AGENTS.md) and separate operator authorization.\n\nStart with [CONTRIBUTING.md](https://github.com/NVlabs/ASPIRE/blob/v0.1/CONTRIBUTING.md). Third-party contributions must include a Developer Certificate of Origin sign-off.\n\nASPIRE material owned by NVIDIA or contributed under the project license is\navailable under the [Apache License 2.0](https://github.com/NVlabs/ASPIRE/blob/main/LICENSE). Third-party materials retain\ntheir original terms. See [NOTICE](https://github.com/NVlabs/ASPIRE/blob/main/NOTICE) and the central\n[licensing and compliance index](https://github.com/NVlabs/ASPIRE/blob/v0.1/LICENSES/README.md) before using or\nredistributing the full stack.\n\nIf you find ASPIRE useful in your research, please cite:\n\n```\n@article{lu2026aspire,\n  title   = {ASPIRE: Agentic /Skills Discovery for Robotics},\n  author  = {Runyu Lu and Yubo Wu and Ethan Kou and Haotian Lin and Letian Fu and Wenli Xiao and\n             Ajay Mandlekar and Yinzhen Xu and Guanya Shi and Ken Goldberg and\n             Ang Chen and Mosharaf Chowdhury and Jan Kautz and Yuke Zhu and Linxi \"Jim\" Fan and Guanzhi Wang},\n  year    = {2026},\n  journal = {arXiv preprint arXiv:2607.00272},\n  url     = {https://arxiv.org/abs/2607.00272}\n}\n```\n\nASPIRE was developed by researchers from NVIDIA, the University of Michigan, the University of Illinois Urbana-Champaign, UC Berkeley, and Carnegie Mellon University.", "url": "https://wpnews.pro/news/aspire-agentic-skills-discovery-for-robotics", "canonical_source": "https://github.com/nvlabs/aspire", "published_at": "2026-10-16 19:59:18+00:00", "updated_at": "2026-10-16 20:19:50.047201+00:00", "lang": "en", "topics": ["robotics", "ai-agents", "ai-research", "machine-learning", "developer-tools"], "entities": ["NVIDIA Labs", "ASPIRE", "LIBERO-Pro", "Robosuite", "BEHAVIOR-1K", "Claude Opus 5.5", "GPT-6 Astra", "Codex"], "also_reported_by": [], "alternates": {"html": "https://wpnews.pro/news/aspire-agentic-skills-discovery-for-robotics", "markdown": "https://wpnews.pro/news/aspire-agentic-skills-discovery-for-robotics.md", "text": "https://wpnews.pro/news/aspire-agentic-skills-discovery-for-robotics.txt", "jsonld": "https://wpnews.pro/news/aspire-agentic-skills-discovery-for-robotics.jsonld"}}