Skip to main content

Creature Lab logo

Creature Lab

Stress-test a real robot, see exactly why it fell, and prove the fix.
On a laptop, in a browser.

CI status PyPI version Python 3.11+ Apache-2.0 license

The Creature Lab app: a Unitree G1 humanoid takes a 150 N shove mid-stride and falls in slow motion while the side panel explains why. One click on 'A gentler shove' runs it again, and the two runs play side by side: at 100 N it keeps walking.

The real app on saved runs: a 150 N shove, why it fell, and one click to try it gentler. Watch the 1-minute tour. Unitree G1 model and policy by Unitree Robotics, BSD 3-Clause. Simulation.

Creature Lab runs published robots with their real trained walking policies. It shoves them at dozens of strengths and moments, walks them over stairs, ramps, rough ground and stepping stones, and tries a failing motor, late commands, slippery floors, loads and weak batteries. Then it explains every fall from the recorded evidence: where the weight went relative to the feet, how fast the feet slid, whether a foot caught a step or missed a stone, and whether the motors ran out of torque. Change one thing, or swap the policy, and see whether the fall is fixed. The physics runs on the CPU (MuJoCo). It is a simulation, not the hardware.

Try it

uvx --from "creature-lab[app]" creature-lab      # or: pip install "creature-lab[app]", then creature-lab

Or from a clone, which also gives you the classic editor and the tests:

git clone https://github.com/oney-erge/Creature-Lab.git
cd Creature-Lab
./run.sh            # macOS: ./run.command   Windows: run.bat

Either way it installs prebuilt packages only (no compiler, about 300 MB) and opens the app. Press Download and check up on the Unitree G1 and accept its license. On a laptop, the download and the whole check-up (40 shoves, then uneven ground, weak motors, late commands, slippery floors and loads) took 47 s from an empty cache.

What you can do

  • Watch a shove in 3D, slowed down at the shove and the fall. The app draws the floor area that holds the robot up and the point where its weight would land.
  • Read why it fell in plain words, with the numbers behind it.
  • Walk it over uneven ground: stairs, a ramp, rough ground and stepping stones, sized to the robot. A fall there is explained as a toe caught on a step, a foot down a gap or a slope too steep.
  • Ask "what if": shove from the other side, harder or gentler, on a slippery floor, slower, faster or carrying a load. It runs again in about a second and plays beside the original.
  • Compare policies on the same test. The 29-joint Unitree G1 has two policies trained for Creature Lab, one with small shoves and one with strong shoves; the Microduck has Pollen's walk-only and walk-and-get-up policies. A robot with more than two lets you pick which to compare.
  • Bring your own robot and policy: a MuJoCo model, an ONNX policy and a small robot.toml (robots validate, robots add; guide) get the same shoves and explanations.
  • Check for regressions from the command line: robots compare says which tests got better or worse between two check-ups (--fail-on-regression for CI), and robots card writes an SVG robustness card for a README or model card.
  • Build your own quadruped, hexapod, worm or humanoid with sliders, and run it.

The Creature Lab app showing the 29-joint Unitree G1 twice, side by side, taking the same 400 N sideways shove at the same moment. The small-shove policy falls at 4.6 s and stays down; the strong-shove policy keeps walking. The panel says the strong-shove policy stayed on its feet.

Same 400 N shove, same moment, two policies trained the same way except for how hard they were shoved in training (simulation). Over eight shoves at 400 N the small-shove policy fell 5 times, the strong-shove policy once. Unitree G1 model: Unitree Robotics, BSD 3-Clause.

The Creature Lab app showing the 29-joint Unitree G1 on stairs with 4 cm steps: a toe catches a step and it falls at 2.8 s, and the panel names the step it caught, with the force on the step's edge. Then the G1 on stepping stones 11 cm apart: a foot goes down into a gap and it falls.

Uneven ground (simulation): stairs with 4 cm steps and stepping stones 11 cm apart. Neither of the G1's policies was trained on anything but a flat floor, and both fall on every stair height the check-up tries. Unitree G1 model: Unitree Robotics, BSD 3-Clause.

From the command line

uv sync --inexact --extra robots
uv run creature-lab robots checkup unitree-g1     # or: microduck
uv run creature-lab robots compare BEFORE_DIR AFTER_DIR --fail-on-regression

A check-up writes an HTML report; add --film for a 1080p film. See Robot check-up for the tests, how falls are explained, licenses and limits.

Creature Lab started as a workbench for designing modular robots, and that is still here: a zoo of working and deliberately broken creatures, a part-by-part build editor with URDF import and export, root-cause diagnosis, and search for better bodies and gaits (creature-lab zoo, build, autopsy, evolve). PyBullet is optional for these (uv sync --inexact --extra sim); outside Linux it is built from source and needs a C++ compiler, so the launcher leaves it out and MuJoCo is used instead.

What it is not

A teaching and prototyping tool, not hardware qualification, a cloud service or large-scale GPU training. Check-ups are deterministic simulations of published models; a robot that passes here can still fall on a real floor.

Documentation

Guide Use it for
Browser app Robots, replays, "what if" and check-ups in the app
Robot check-up The tests, how falls are explained, licenses
Getting started A first experiment, and troubleshooting
CLI reference Every command and export
Known issues Current limitations
All documentation Build editor, zoo, failure lab, schemas, contributing

If the launcher fails, it prints what to do; uv run creature-lab doctor checks the setup. Docker serves the same app on localhost: ./run.sh docker.

Apache-2.0. Built by Oney Erge.

Metadata

Release files for creature-lab 0.4.2

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for creature-lab 0.4.2
File Size Uploaded Provenance Actions
creature_lab-0.4.2.tar.gz 8.7 MB With provenance Details for creature_lab-0.4.2.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for creature-lab 0.4.2
File Interpreter ABI Platform Actions
creature_lab-0.4.2-py3-none-any.whl Python 3 none any Details for creature_lab-0.4.2-py3-none-any.whl

Total release size: 12.5 MB

Release files / creature_lab-0.4.2.tar.gz

Download URL creature_lab-0.4.2.tar.gz
Size 8.7 MB
Tags Source
SHA-256 checksum
How to use checksums
5897025ddad2a37248fd5d5f2227d579404a66969a0be9100ed0a0705cdb41b0
BLAKE2b-256 checksum
How to use checksums
0d4dd8f50d9c3d82964d125d3c45ea106f0510716c73dbaede1ac16c9008b2f1
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 12, 2026.

Transparency log

Release files / creature_lab-0.4.2-py3-none-any.whl

Download URL creature_lab-0.4.2-py3-none-any.whl
Size 3.8 MB
Tags Python 3
SHA-256 checksum
How to use checksums
3227a8005113d1de79307d79f3a3cd50d3fa1eee189cd3930d67fa05f49f5238
BLAKE2b-256 checksum
How to use checksums
d2c436c4d5a4dac08b13a7360d795fc48811200b8cca5cc59d580bea5e8224ad
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.16

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 12, 2026.

Transparency log

Security for creature-lab 0.4.2

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

Full provenance

Both files in this release have been uploaded with provenance metadata.

Select from 2 files

Attestation for creature_lab-0.4.2.tar.gz

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Nov 12, 2026.

Transparency log

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page