ASPIRE README

项目的 README,来自 github.com/NVlabs/ASPIRE。合并为一个 markdown 文件:README.md。

ASPIRE: Agentic /Skills Discovery for Robotics

Project Page  |  Paper

ASPIRE discovering /skills across simulation and real-robot tasks

ASPIRE is a new type of continual learning: "training" is skill refinement instead of gradient descent.

"Trained model" is a repo of sensorimotor skills instead of floating weights.

“Distributed training” is a panel of agents each practicing a different skill instead of sharded minibatches.

News

Real-World Results

Cross-embodiment sim-to-real Online real-world learning
Long Horizon Table Bussing
Claude Opus 4.6 (sim), GPT-5.5 (real)
Rubik's Cube
GPT-6 Astra
Stack Cups
GPT-6 Astra
Unscrew Bottle Cap
GPT-6 Astra
Pill in Bottle
GPT-6 Astra

Model Scaling

Success vs. API cost LIBERO-Pro, Robosuite, BEHAVIOR-1K
ASPIRE success vs. API cost across twelve coding-agent models ASPIRE success rates on LIBERO-Pro, Robosuite and BEHAVIOR-1K by model
Stronger models discover better strategies
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.
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.

ASPIRE 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.

Quick Start

Run with a coding agent

To get started with ASPIRE, launch a coding agent such as Codex or Claude Code and enter the following prompt:

Clone the main branch of https://github.com/NVlabs/ASPIRE/,
read the iterative-debugging skill (.claude/skills/iterative-debugging/SKILL.md) and
docs/libero-pro.md, and run the complete ASPIRE LIBERO-Pro Goal-Swap Quick Start for
all ten tasks in the libero_goal_swap suite.

Before executing, report the required GPUs, credentials, gated weights,
services, expected runtime, seed partitions, and output paths. Wait for
my confirmation before launching. Do not access real-robot code or instructions
(AGENTS.md, docs/system2*.md, docs/yam-real.md, src/cap_harness/yam_real/).

Before running, set up the host (CUDA, uv, and the Hugging Face token for the gated SAM3 weights) with docs/host-setup.md.

Reference agent environments: ASPIRE is coding-agent agnostic. The simulation workflow is packaged as agent skills for Claude Code (.claude/skills/) and Codex (.codex/skills/): iterative-debugging and evosearch. AGENTS.md holds the execution rules for the real-robot System 2 agent (operator runbook) and is not needed for simulation runs.

Reproduce full paper results

Name the suite and experiment explicitly. If neither is named, the agent should present this table and stop for selection.

Suite Experiment Runbook Video
LIBERO-Pro Iterative-debugging loop iterative-debugging/
Robosuite Iterative-debugging loop iterative-debugging/
BEHAVIOR-1K Iterative-debugging loop iterative-debugging/behavior/
LIBERO-Pro Evolutionary Search evosearch/
Robosuite Evolutionary Search evosearch/
LIBERO Zero-Shot Transfer libero/zeroshot-transfer/
LIBERO-Long-Pro Library-Size Scaling libero/library-size-scaling/
LIBERO-Long-Pro Inference-Time Scaling libero/inference-time-scaling/
Robosuite Debugging-Token-Ablation robosuite/training-law/

Before 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.

Reproduce the real-robot manipulation demos

Use the System 2 operator runbook to prepare a calibrated YAM station and launch a fresh agent with the appropriate text-based task skill:

The manipulation skill entrypoint routes the agent to these recipes. They provide grasp strategies, recovery, and physical completion checks without requiring episode logs or fixed-coordinate replays. The three new recipes are distilled from demo sessions, not a new real-robot validation; the pill recipe explicitly distinguishes a seated cap from verified final tightness. Re-ground the current scene and verify the actual outcome.

[!WARNING] ASPIRE 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 and separate operator authorization.

Contribution Guidelines

Start with CONTRIBUTING.md. Third-party contributions must include a Developer Certificate of Origin sign-off.

License

ASPIRE material owned by NVIDIA or contributed under the project license is available under the Apache License 2.0. Third-party materials retain their original terms. See NOTICE and the central licensing and compliance index before using or redistributing the full stack.

Citation

If you find ASPIRE useful in your research, please cite:

@article{lu2026aspire,
  title   = {ASPIRE: Agentic /Skills Discovery for Robotics},
  author  = {Runyu Lu and Yubo Wu and Ethan Kou and Haotian Lin and Letian Fu and Wenli Xiao and
             Ajay Mandlekar and Yinzhen Xu and Guanya Shi and Ken Goldberg and
             Ang Chen and Mosharaf Chowdhury and Jan Kautz and Yuke Zhu and Linxi "Jim" Fan and Guanzhi Wang},
  year    = {2026},
  journal = {arXiv preprint arXiv:2607.00272},
  url     = {https://arxiv.org/abs/2607.00272}
}

ASPIRE was developed by researchers from NVIDIA, the University of Michigan, the University of Illinois Urbana-Champaign, UC Berkeley, and Carnegie Mellon University.

ASPIRE (NVlabs/ASPIRE)

正在搜索数据库…

这是什么?ASPIRE(Agentic /Skills Discovery for Robotics)是一个研究框架。机器人通过不断完善由编程智能体驱动的技能库来提升能力,而不是训练神经网络权重。它在 LIBERO-Pro、Robosuite 和 BEHAVIOR-1K 基准测试中达到了最先进水平,甚至能还原魔方。
这是可以下载的训练好的模型吗?不提供任何模型权重。这里的“训练”是指一组编程智能体反复完善一个包含 Python 技能脚本的代码仓库,因此输出的是不断增长的感知与运动技能库,而不是浮点权重。
我该怎么把它跑起来?请先按照 docs/host-setup.md 配置主机(CUDA、uv,以及用于受限 SAM3 权重的 Hugging Face 令牌),然后启动 Codex 或 Claude Code 等编程智能体,并把 README 中的快速入门提示词交给它。
它支持哪些编程智能体?ASPIRE 不依赖特定的编程智能体。仿真工作流以技能的形式提供给 Claude Code(.claude/skills/)和 Codex(.codex/skills/),包括 iterative-debugging 和 evosearch。
直接运行生成的代码安全吗?不安全。ASPIRE 会以完整的导入权限执行语言模型生成的 Python 代码,而且试验看门狗并不是安全沙箱,因此必须在隔离的主机上运行。该主机不应包含凭据、敏感挂载点,也不应有网络或硬件访问权限。
我得有真机器人才能用吗?不需要。主要工作流在仿真环境中运行(LIBERO-Pro、Robosuite、BEHAVIOR-1K)。魔方和倒液体这类真机演示需要单独校准的 YAM 工作站,并需操作人员明确授权。
这个机器人能完成哪些真实世界的任务?System 2 真机任务方案涵盖还原魔方、拧开瓶盖后倒液体、放入药片并盖上瓶盖,以及将杯子倒扣嵌套。
不同模型之间的表现为什么差别这么大?成功率取决于底层大语言模型的能力。在 10 个 LIBERO-Pro 任务上,成功率从使用 GPT-6 Luna 时的 15% 提升到使用 Claude Opus 5.5 时的 99%。而在成本与成功率的权衡曲线上,表现最强的模型不一定是最贵的。
作者
NVlabs
大小
56.1 MB
星标
1K
更新
提交Oct 16
时长
4个月
提交 · 90 天