Inspect Robots
SoftwarePython framework for evaluating robot policies against scored tasks with compatible hardware, simulator, or mock adapters. [1] [2]
Installation & first result →Official websiteGetting started
- For
- Robotics researchers and engineers building repeatable policy evaluations. [1]
- Source rights
- Named licenses are documented by part. Review licenses
- Unknowns
- operating system, compute requirements not fully established. Review gaps
Source researched · Not tested by us
Sources checked · Research details
Overview
Inspect Robots is a Python framework for testing robot policies. It separates the task and scorer from the policy and the robot or simulator that runs it. You can keep the task fixed while changing one of those pieces, provided the replacement has a compatible adapter. The evaluation log records the configuration and scores so you can tell which run produced a result. [1] [2]
Is this for you?
- Workflows
- Run the scripted mock
- Review the evaluation log
- Configure a compatible rig and policy plugin
What you need
Installation & first result
Open the upstream getting-started guide [2]
- Create a Python 3.11+ virtual environment and install inspect-robots==1.0.0.
- Run the CubePick task with the scripted policy and cubepick embodiment.
- Inspect the saved evaluation log. Install and configure the relevant plugin for a physical or simulator rig.
Hardware, models & data
Project notes
Start with the scripted mock
Use Python 3.11 or newer and install inspect-robots==1.0.0 in a virtual environment. The upstream quickstart documents a CubePick mock with a scripted policy. It exercises task setup, scoring, and logging without a robot, GPU, simulator, or model API key. Use it to check the evaluation wiring before trying a learned policy. [1] [2]
``sh inspect-robots run --task cubepick-reach --policy scripted --embodiment cubepick ``
Use inspect-robots inspect with the resulting JSON log to examine the run or inspect-robots view to make an HTML report. Install the optional Rerun extra if you want the visualization workflow. [1] [2]
What changes on a real robot
The documented I2RT YAM route uses a Linux robot host, the YAM plugin, and camera and rig configuration. The MolmoAct2 policy client also needs a separately running model server on a GPU host. The LLM agent route uses a separate plugin and provider credentials. Confirm the combination of policy, adapter, and rig before moving to hardware. [1] [2]
Robocurve's RoboSpoof study is an example of the evaluation questions this tooling can support. Across five models and five tasks, the authors reported attempts in 81 of 300 trials with unaltered footage, 263 of 300 with blurred footage, and 291 of 300 with doctored footage. An AI judge classified attempts and refusals. The authors did not inspect the scene afterward to confirm physical completion. [4]
The core package declares an MIT license. Model servers, benchmarks, and external assets have their own terms and dependencies. This profile is based on documentation and package metadata. The mock, simulator, and hardware workflows were not run for this review. [1] [2] [3]
What we checked
Source research only. No upstream media reproduced.
Not yet established
- Requirements: operating system
Updates
- Upstream last updated
- Not established
- Latest release
- Nov 4, 2026 [1]
- Profile updated
- Nov 9, 2026
- Sources last checked
- Nov 9, 2026
- Next review
- Dec 9, 2026
- Reviewed revision
- PyPI 1.0.0 (November 4, 2026), source and quickstart at fc5ace923ce85117f9c86f35274122212fbcfe2c.
Source files and licenses
Robot-specific evaluation framework with released source, packaged install, a runnable mock workflow, and documented hardware adapters.
| Layer | Availability / license |
|---|---|
| software | available · MIT Scope is the core repository. Separately supplied benchmarks, models, and third-party assets require separate review. [5] [3] |
| documentation | available · MIT Scope is the core repository. Separately supplied benchmarks, models, and third-party assets require separate review. [2] [3] |
External model or benchmark assets are outside this framework profile.
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 Inspect Robots 1.0.0
Version 1.0.0 uploaded November 4, 2026. Python >=3.11, MIT, Rerun extra. The README describes the evaluation framework, Linux real-robot control, YAM plugin and separately hosted MolmoAct2 server. PyPI attestation links release workflow revision fc5ace923ce85117f9c86f35274122212fbcfe2c.Revision, date & where cited
Inspect Robots contributors via PyPI · pypi.org · Description, metadata, release files and provenance · Accessed Nov 9, 2026
Used for: Overview, What you need, Installation & first result, Hardware, models & data and 3 more - S002 Inspect Robots quickstart
CubePick scripted mock uses no robot or simulator. The CLI command is inspect-robots run --task cubepick-reach --policy scripted --embodiment cubepick. CLI inspect reads a saved JSON log; view writes an HTML report. YAM setup installs inspect-robots-yam and configures cameras. LLM agent route uses a separate plugin and API credentials. The quickstart was inspected at the PyPI release workflow revision.Revision, date & where cited
Robocurve · github.com · revision
fc5ace9· Install, first evaluation, command line, real robot, LLM and simulation · Accessed Nov 9, 2026
Used for: Overview, Is this for you?, What you need, Installation & first result and 4 more - S003 Inspect Robots repository license
MIT License. Copyright (c) 2026 robocurve. The grant covers the software and associated documentation and requires preserving the copyright and permission notice.Revision, date & where cited
Robocurve · github.com · revision
fc5ace9· Repository LICENSE · Accessed Nov 9, 2026
Used for: What changes on a real robot, Source files and licenses - S004 RoboSpoof study
The study ran 900 trials across five models and five tasks. It reports 81/300 attempts with unaltered footage, 263/300 with blur and 291/300 with doctored footage. Claude Fable 5.1 classified attempts and refusals. A call to done marked completion for an attempt and the scene was not checked afterward.Revision, date & where cited
Robocurve · robocurve.org · Results and methods, grading · Accessed Nov 9, 2026
Used for: Is this for you?, Hardware, models & data, What changes on a real robot - S005 Inspect Robots repository
The public tree contains src/inspect_robots, plugins, docs, examples and tests. The framework composes a policy, a compatible embodiment, scenes and a scorer, and produces EvalLogs. The README identifies benchmark catalog WorldEvals and backend plugin packages as separately scoped artifacts.Revision, date & where cited
Robocurve · github.com · revision
fc5ace9· Repository tree and README, install, quickstart, framework comparison · Accessed Nov 9, 2026
Used for: Hardware, models & data, Source files and licenses
Correction guidance · Discussion is not enabled · Profile format 0.7.0-draft.5