Repolode
Submit
NVlabs

NVlabs/ASPIRE

ASPIRE is a framework where coding agents iteratively refine reusable robot skills for simulated and real-robot manipulation tasks instead of training model weights.

View on GitHubWebsite
Stars
670
Forks
55
Open issues
15
Last commit
4 days ago

What is it?

ASPIRE (Agentic Skills Discovery for Robotics) treats robot learning as skill refinement: a coding agent writes and revises Python skills, and the resulting repository of sensorimotor skills serves as the learned artifact in place of neural network weights. Several agents can practice different skills in parallel.

The project provides agent skills for Claude Code and Codex, including an iterative-debugging loop and an evolutionary search workflow. It reports runs on LIBERO-Pro, Robosuite and BEHAVIOR-1K simulation suites, along with scripts and runbooks for reproducing experiments and real-robot manipulation demos on a YAM arm.

Generated code runs with full import access, so the project recommends an isolated host without credentials or sensitive mounts, and real-robot use requires separate operator authorization.

Key features

  • 01Refines robot skills through coding-agent loops instead of gradient-based training.
  • 02Ships Claude Code and Codex skills for iterative debugging and evolutionary search.
  • 03Supports LIBERO-Pro, Robosuite and BEHAVIOR-1K simulation benchmarks.
  • 04Includes runbooks for reproducing paper experiments and real-robot manipulation demos.
  • 05Provides text-based task recipes for grasping, recovery and completion checks.

Use cases

  • Researchers studying agent-driven skill discovery for robot manipulation.
  • Reproducing LIBERO-Pro, Robosuite or BEHAVIOR-1K results with coding agents.
  • Experimenting with real-robot manipulation on a calibrated YAM arm.

Quick start

  1. 1

    Set up the host (CUDA, uv, Hugging Face token for gated SAM3 weights) using the host setup guide

  2. 2

    Launch a coding agent such as Codex or Claude Code and enter the ASPIRE LIBERO-Pro prompt from the README

Things to know

  • Running generated code requires an isolated host, since ASPIRE executes model-generated Python with full import access and is not a security sandbox.
  • Full paper-scale runs need GPUs, credentials, gated model weights, and a Hugging Face token.
  • Strongest results depend on paid frontier coding-agent models such as Claude Opus.

Topics

#robotics#agents#skill-discovery#simulation#manipulation#reinforcement-learning#llm

FAQ

What is ASPIRE?
ASPIRE is a framework where coding agents iteratively refine reusable robot skills for simulated and real-robot manipulation tasks instead of training model weights. ASPIRE (Agentic Skills Discovery for Robotics) treats robot learning as skill refinement: a coding agent writes and revises Python skills, and the resulting repository of sensorimotor skills serves as the learned artifact in place of neural network weights. Several agents can practice different skills in parallel.
Is ASPIRE free to use? What license does it have?
ASPIRE is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is ASPIRE still maintained?
Actively maintained. The last commit to ASPIRE was on Oct 16, 2026.
What can I use ASPIRE for?
Researchers studying agent-driven skill discovery for robot manipulation. Reproducing LIBERO-Pro, Robosuite or BEHAVIOR-1K results with coding agents. Experimenting with real-robot manipulation on a calibrated YAM arm.
How do I install ASPIRE?
Set up the host (CUDA, uv, Hugging Face token for gated SAM3 weights) using the host setup guide Launch a coding agent such as Codex or Claude Code and enter the ASPIRE LIBERO-Pro prompt from the README
Where can I get ASPIRE?
The source code and installation instructions are on GitHub: https://github.com/NVlabs/ASPIRE

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects