Windows x64 experimental preview — October 30, 2026

Creature Lab — Windows x64 experimental preview

This archive contains the current source snapshot at
0c2e0c1572abfb439a9ade6b4ec2911d34d356a0, including project controller,
backend and seed settings, saved-run replay, and baseline comparison controls.

Download creature-lab-windows-x64-portable-unverified.zip, extract it to a
writable folder, and open START.bat. Close its terminal to stop the editor.
It bundles Python, runtime dependencies and a locally compiled PyBullet wheel;
it is a portable archive, not an installer. The Python wheel is included for
developers. Optional CMA-ES, media export, and trained-model dependencies are
not included.

gait-comparison-example.html is a historical comparison from an earlier local
0.2.0 candidate: one Windows/Python 3.12/PyBullet 3.2.7 run per controller at
seed 7. The curated and sinusoid scores were 1.9703 and 0.5736 for the
crawl-forward task. It shows recorded paths and scores, not live simulation.
The original raw run directories are not included, and the current preview was
not run to verify the result. The separate .sha256 file covers this example.

This archive is unsigned and unverified. The app, simulations, browser
playback, and clean-machine first run were not exercised for this preview. No
tests or CI results were used. Diagnosis is heuristic, and physics can differ
across backends.

The existing cross-platform preview
contains older source for Linux and macOS. macOS previews use MuJoCo and may
behave differently from this PyBullet build. See BUILD-INFO.md and
SHA256SUMS.txt for the reused runtime and artifact hashes.

Creature Lab 0.2.0 — experimental prerelease

Creature Lab 0.2.0 — experimental prerelease

Design a robot, simulate it, inspect a candidate explanation, and preserve the
recorded evidence.

Try a recording: download creature-lab-recorded-example.html and open it
in a browser. It is a previously recorded quadruped replay; it does not run
physics in the browser.

Windows x64: download creature-lab-windows-x64-portable-unverified.zip,
extract it to a writable folder, and open START.bat. It contains Python, the
application, locked runtime dependencies, and a locally compiled PyBullet wheel,
so a compiler is not needed to assemble the included payload. It is a portable
archive, not an installer. Close its terminal to stop the local server. Python
wheel and source archives are also supplied for developers.

Linux x86-64: creature-lab-linux-x64-portable-unverified.tar.gz includes
the official PyBullet 3.2.7 Linux wheel and the runtime. Extract it to a writable
directory and run run.sh. It requires glibc 2.28 or newer. This unsigned preview
has not been run on Linux.

macOS Apple Silicon: creature-lab-macos-arm64-mujoco-portable-unverified.tar.gz
contains a runtime and starts the editor using the existing MuJoCo backend. It
does not include PyBullet: gaits and scores can differ, the default gait may move
weakly, and PyBullet-only or cross-backend features are unavailable. Extract to a
writable folder and open START.command. This unsigned archive has not been run
on macOS.

macOS Intel: creature-lab-macos-x64-mujoco-3.10-portable-unverified.tar.gz
uses the official MuJoCo 3.10.0 wheel, rather than the project's locked MuJoCo
3.17.0, because 3.17.0 has no Intel Mac wheel. It also excludes PyBullet and has
the same backend limitations. This unsigned archive has not been run on macOS.

This prerelease and the recorded example have not been opened or run as a
clean-machine/browser check. They are unsigned and may have problems. No tests or
CI results were used for this candidate. Results and diagnosis are heuristic;
different physics backends can behave differently. A separate Python installation
from source may still require a C++ toolchain for PyBullet on Windows or macOS.
The portable downloads do not include optional CMA-ES, media-export, or trained-model
dependencies, so features requiring those extras need a separate development setup.

The editor now skips an indistinguishable controller counterfactual, retains
the reference and perturbation recordings in verified packs, and compares 3D
recordings by simulation time with their own terrain.