Open Source RoboticsRobots, software & learning

ASPIRE

Software

Coding-agent framework for developing reusable robot manipulation skills in simulation, with a separate documented YAM hardware workflow. [1]

Installation & first result →Official websiteGetting started

For
Robotics researchers with Linux GPU infrastructure and coding-agent access [1] [2] [5]
Hardware
Simulated Panda in LIBERO-Pro, Calibrated YAM station for the separate real-robot runbook [1] [3]
Source rights
Named licenses are documented by part. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware

Requirements · Research details and dates

On this page

Overview

ASPIRE lets a coding agent develop reusable robot manipulation skills by trying programs, examining feedback and refining a skill library. Its repository provides simulation workflows and a separate route for a calibrated YAM robot station. For a first evaluation, the documented LIBERO-Pro simulation route is the narrower starting point. [1] [3]

Is this for you?

Workflows
  • Iterative debugging
  • Evolutionary skill search
  • Skill-library transfer and scaling experiments

[1]

Known limitations
  • SAM3 requires approved gated-model access and a Hugging Face token.
  • Third-party submodules and services are separate dependencies.
  • LIBERO structural validation does not measure task completion.
  • Saved initial states wrap around; larger seed numbers do not necessarily mean novel scenes.
  • Generated Python is not sandboxed by trial watchdogs; upstream requires host isolation and separation from physical hardware.

[2] [3] [4] [1]

What you need

Operating system
  • Linux x86_64 (uv environment scope)

[5]

Languages / runtimes
  • Python

[5]

Dependencies
  • uv
  • NVIDIA driver
  • SAM3 gated weights
  • Pinned LIBERO-PRO and robosuite forks
  • Provider-specific SAM3, PyRoki and cuRobo runtimes

[2] [3] [5]

Host computer

NVIDIA GPU host for the provider stack; exact minimum GPU VRAM for LIBERO is not specified in the reviewed host setup. BEHAVIOR has additional requirements.

[2]

Compute for the workload
  • Provider stack needs an NVIDIA driver/GPU; cuRobo bootstrap supplies its own nvcc wheel for the general route.
  • BEHAVIOR route additionally documents CUDA 12.x toolkit, glibc >=2.35, 32 GB RAM, about 12 GB environment storage and 40 GB dataset storage.

[2]

Skills
  • Python and Linux environment management
  • Simulation debugging and evaluation design

Editorial classification from the cited documented purpose and prerequisites; not a runtime assessment.

[2] [3] · Editorial assessment

Installation & first result

Open the upstream getting-started guide [3] · Project documentation [1]

  • Prepare uv, NVIDIA driver and approved SAM3 access.
  • Run the pinned LIBERO bootstrap and activate its runtime; configure EGL libraries and GPU selection.
  • Run doctor and the LIBERO smoke validation plan.
  • Run a documented example and inspect its recorded task outcome before attempting broader skill search.

[2] [3]

Installation

Use scripts/bootstrap_libero.sh and the frozen .venv-libero runtime; avoid unpinned package upgrades.

[3]

Hardware, models & data

Supported hardware
  • Simulated Panda in LIBERO-Pro
  • Calibrated YAM station for the separate real-robot runbook

[1] [3]

Interfaces
  • cap-harness CLI
  • Python Code-as-Policy APIs

[5]

Inputs and outputs

LIBERO adapter exposes RGB-D, seven Panda joints, gripper state and end-effector pose; programs issue actions and record run metadata.

[3]

File formats
  • Python programs
  • JSON run metadata
  • YAML validation plans

[3]

Algorithms and models
  • Code-as-Policy skill refinement
  • SAM3 perception
  • cuRobo motion planning

[1] [2] [5]

Evaluation

LIBERO smoke checks seed 1 and nightly checks seeds 1–3 for structural runtime behavior. Task success must be evaluated separately; saved states repeat beyond their finite set.

[3]

Project notes

What to prepare before trying it

The reviewed environment targets Linux x86_64 and Python 3.10–3.12. Install uv and provide a working NVIDIA driver and GPU. SAM3 weights require both approved Hugging Face access and a token; a successful bootstrap alone does not establish that the perception service can load them. The reviewed host guide does not specify a minimum GPU VRAM figure for LIBERO. [2] [5]

ASPIRE executes generated Python with full import access; its watchdogs are not a security sandbox. Upstream calls for an isolated host without sensitive mounts or exposed credentials, restricted network access, and no physical-hardware access for simulation agents. Keep necessary model authentication separate from unrelated credentials. [1] [2]

Follow the pinned LIBERO bootstrap, activate its frozen environment and configure EGL rendering. The driver alone is insufficient: the guide also requires the GLVND dispatch library, such as Debian's libegl1. On a multi-GPU host, set the MuJoCo EGL device as well as CUDA device visibility; the guide warns that contention on the default rendering GPU can resemble a policy failure. [3]

Start with the doctor check and smoke validation plan, then run one documented example and inspect its recorded outcome. Keep the bootstrap's dependency pins instead of repairing the environment with arbitrary upgrades. Coding-agent access, GPU time and model-service usage need a workload-specific budget; this review establishes no complete run price. [1] [3]

A passing smoke check is not a solved task

LIBERO validation checks reset metadata, schemas, movement ticks, clean shutdown and resumability. Its smoke plan uses seed 1; nightly uses seeds 1–3. These are structural checks, explicitly distinguished upstream from task-completion benchmarks. [3]

The default saved-state mode cycles through a finite set, typically 50 states. Seeds beyond that set can repeat earlier scenes. The procedural mode avoids that particular wraparound but still does not guarantee distinct layouts for every task. Record the initial-state mode, task and seeds when comparing results; do not treat a large seed count as evidence of equally broad variation. [3]

Choose the simulation workload before sizing the machine

The general provider setup builds cuRobo extensions using its own nvcc wheel. BEHAVIOR-1K adds a different requirement: a CUDA 12.x toolkit, glibc 2.35 or newer, 32 GB RAM, roughly 12 GB for its environment and 40 GB for datasets. It also introduces Isaac/Omniverse and dataset terms. These are BEHAVIOR requirements, not a stated minimum for every ASPIRE experiment. [2]

The real YAM workflow requires its own calibrated station and operator runbook. A working LIBERO example does not establish compatibility with an arbitrary arm or readiness to operate physical hardware. [1] [3]

What we checked

Source researched only; no installation, performance benchmark or hardware test performed. No affiliate links or reused media.

Updates

Upstream last updated
Oct 16, 2026 [6]
Latest release
Not established
Profile updated
Oct 17, 2026
Sources last checked
Oct 17, 2026
Next review
Nov 17, 2026
Reviewed revision
fa00f0ef2bc1d91d594b9aa502c22133c9f538f7 (commit snapshot; no GitHub release in checked listing) · software repository

Source files and licenses

Open robotics skill-discovery harness and documentation; external models, datasets and submodules have separate access and license requirements.

LayerAvailability / license
softwareavailable · Apache-2.0
Repository materials only; third-party submodules retain their own licenses. [7] [4]
model weightsnot found · License not fully assessed
Not present in the preserved ASPIRE source archive. External SAM3 weights require approved access; external model licenses are not covered by the ASPIRE root license. No weights were downloaded in this review. [2] [4]
datasetsnot found · License not fully assessed
Not present in the preserved ASPIRE source archive. Benchmark data are external; BEHAVIOR data terms must be accepted separately. No benchmark data bundle was downloaded in this review. [2] [4]
documentationavailable · Apache-2.0
Repository materials only; third-party submodules retain their own licenses. [7] [4]

Code license and review scope

ASPIRE's repository code and documentation carry Apache-2.0 terms. Its NOTICE explicitly separates third-party submodules and their licenses. Gated model weights, external model services and benchmark datasets are not made Apache-2.0 by that root license. The preserved source archive does not include those external dependencies. [4] [7] [2]

This profile reviews commit fa00f0ef2bc1d91d594b9aa502c22133c9f538f7, dated October 16, 2026. The README announces v1.0, but the checked GitHub releases listing is empty, so this is identified as a commit snapshot. We have not installed ASPIRE, reproduced its reported success rates or tested its hardware workflows. [1] [6]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5

Change history