stable-envs 0.1.0__tar.gz

stable_envs-0.1.0/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 GalilAI-group
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

stable_envs-0.1.0/PKG-INFO

@@ -0,0 +1,241 @@
+Metadata-Version: 2.4
+Name: stable-envs
+Version: 0.1.0
+Summary: Fast and Reproducible Environments with Controllable Variations
+Author: Randall Balestriero, Sami BuGhanem
+License-Expression: MIT AND LGPL-2.1-or-later
+Project-URL: Homepage, https://github.com/galilai-group/stable-envs
+Project-URL: Documentation, https://galilai-group.github.io/stable-envs/
+Project-URL: Issues, https://github.com/galilai-group/stable-envs/issues
+Project-URL: Source, https://github.com/galilai-group/stable-envs
+Keywords: reinforcement-learning,environments,rendering,pytorch,gpu
+Classifier: Development Status :: 3 - Alpha
+Classifier: Intended Audience :: Science/Research
+Classifier: Operating System :: MacOS
+Classifier: Operating System :: POSIX :: Linux
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: Python :: 3.10
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: <3.13,>=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+License-File: THIRD_PARTY_NOTICES.md
+License-File: licenses/APACHE-2.0.txt
+License-File: licenses/Bullet-zlib.txt
+License-File: licenses/Gymnasium-MIT.txt
+License-File: licenses/Gymnasium-Robotics-MIT.txt
+License-File: licenses/LGPL-2.1.txt
+License-File: licenses/Munk2D-MIT.txt
+License-File: licenses/NumPy-BSD-3-Clause.txt
+License-File: licenses/NumPy-random-NCSA-BSD-3-Clause.txt
+License-File: licenses/OGBench-MIT.txt
+License-File: licenses/PyFlyt-MIT.txt
+Requires-Dist: gymnasium>=1.0
+Requires-Dist: jax>=0.4.35
+Requires-Dist: loguru>=0.7
+Requires-Dist: numpy>=1.24
+Requires-Dist: tensordict>=0.6
+Requires-Dist: torch>=2.2
+Provides-Extra: classic
+Requires-Dist: pygame>=2.5; extra == "classic"
+Requires-Dist: numba>=0.60; extra == "classic"
+Provides-Extra: pusht
+Requires-Dist: pygame>=2.5; extra == "pusht"
+Requires-Dist: pymunk==7.2.0; extra == "pusht"
+Requires-Dist: opencv-python-headless; extra == "pusht"
+Requires-Dist: shapely; extra == "pusht"
+Requires-Dist: numba>=0.60; extra == "pusht"
+Provides-Extra: pusht-gpu
+Requires-Dist: pygame>=2.5; extra == "pusht-gpu"
+Requires-Dist: pymunk==7.2.0; extra == "pusht-gpu"
+Requires-Dist: opencv-python-headless; extra == "pusht-gpu"
+Requires-Dist: shapely; extra == "pusht-gpu"
+Requires-Dist: numba>=0.60; extra == "pusht-gpu"
+Requires-Dist: jax[cuda12]>=0.4.35; sys_platform == "linux" and extra == "pusht-gpu"
+Provides-Extra: pusht-reference
+Requires-Dist: gym-pusht; extra == "pusht-reference"
+Provides-Extra: navigation
+Requires-Dist: matplotlib>=3.10; extra == "navigation"
+Requires-Dist: numba>=0.60; extra == "navigation"
+Provides-Extra: navigation-gpu
+Requires-Dist: jax[cuda12]>=0.4.35; sys_platform == "linux" and extra == "navigation-gpu"
+Provides-Extra: ogbench
+Requires-Dist: mujoco<3.18,>=3.11; extra == "ogbench"
+Requires-Dist: ogbench<1.3,>=1.2.1; extra == "ogbench"
+Requires-Dist: numba>=0.60; extra == "ogbench"
+Provides-Extra: ogbench-manipulation
+Requires-Dist: ogbench==1.2.1; extra == "ogbench-manipulation"
+Requires-Dist: mujoco==3.17.0; extra == "ogbench-manipulation"
+Requires-Dist: dm-control>=1.0.50; extra == "ogbench-manipulation"
+Requires-Dist: numba>=0.60; extra == "ogbench-manipulation"
+Provides-Extra: ogbench-manipulation-gpu
+Requires-Dist: ogbench==1.2.1; extra == "ogbench-manipulation-gpu"
+Requires-Dist: mujoco==3.17.0; extra == "ogbench-manipulation-gpu"
+Requires-Dist: dm-control>=1.0.50; extra == "ogbench-manipulation-gpu"
+Requires-Dist: numba>=0.60; extra == "ogbench-manipulation-gpu"
+Requires-Dist: mujoco-warp==3.17.0; extra == "ogbench-manipulation-gpu"
+Requires-Dist: warp-lang==1.17.0; extra == "ogbench-manipulation-gpu"
+Provides-Extra: dmcontrol
+Requires-Dist: dm-control>=1.0.50; extra == "dmcontrol"
+Requires-Dist: mujoco<3.18,>=3.17; extra == "dmcontrol"
+Requires-Dist: numba>=0.60; extra == "dmcontrol"
+Provides-Extra: dmcontrol-gpu
+Requires-Dist: dm-control>=1.0.50; extra == "dmcontrol-gpu"
+Requires-Dist: mujoco<3.18,>=3.17; extra == "dmcontrol-gpu"
+Requires-Dist: mujoco-warp<3.18,>=3.17; extra == "dmcontrol-gpu"
+Requires-Dist: warp-lang>=1.16; extra == "dmcontrol-gpu"
+Requires-Dist: mujoco-mjx<3.18,>=3.17; extra == "dmcontrol-gpu"
+Requires-Dist: jax[cuda12]>=0.4.35; sys_platform == "linux" and extra == "dmcontrol-gpu"
+Provides-Extra: robotics
+Requires-Dist: gymnasium-robotics<1.5,>=1.4.2; extra == "robotics"
+Requires-Dist: mujoco<3.12,>=3.11; extra == "robotics"
+Requires-Dist: opencv-python-headless; extra == "robotics"
+Requires-Dist: numba>=0.60; extra == "robotics"
+Provides-Extra: mjwarp
+Requires-Dist: mujoco<3.18,>=3.11; extra == "mjwarp"
+Requires-Dist: mujoco-warp<3.18,>=3.11; extra == "mjwarp"
+Requires-Dist: warp-lang>=1.16; extra == "mjwarp"
+Provides-Extra: mshuman
+Requires-Dist: filelock>=3.0; extra == "mshuman"
+Requires-Dist: mujoco<3.18,>=3.17; extra == "mshuman"
+Requires-Dist: mujoco-warp<3.18,>=3.17; extra == "mshuman"
+Requires-Dist: warp-lang>=1.16; extra == "mshuman"
+Provides-Extra: rocket
+Requires-Dist: pillow; extra == "rocket"
+Requires-Dist: scipy; extra == "rocket"
+Requires-Dist: PyFlyt==0.29.0; extra == "rocket"
+Requires-Dist: pybullet==3.2.7; extra == "rocket"
+Requires-Dist: numba>=0.60; extra == "rocket"
+Provides-Extra: craftax
+Requires-Dist: craftax>=1.6.1; extra == "craftax"
+Requires-Dist: jax>=0.4.30; extra == "craftax"
+Provides-Extra: roadster
+Requires-Dist: mujoco<3.18,>=3.17; extra == "roadster"
+Requires-Dist: pillow; extra == "roadster"
+Provides-Extra: roadster-gpu
+Requires-Dist: mujoco<3.18,>=3.17; extra == "roadster-gpu"
+Requires-Dist: mujoco-warp<3.18,>=3.17; extra == "roadster-gpu"
+Requires-Dist: warp-lang>=1.16; extra == "roadster-gpu"
+Requires-Dist: pillow; extra == "roadster-gpu"
+Provides-Extra: craftax-gpu
+Requires-Dist: craftax>=1.6.1; extra == "craftax-gpu"
+Requires-Dist: jax[cuda12]>=0.4.30; sys_platform == "linux" and extra == "craftax-gpu"
+Provides-Extra: miniworld
+Requires-Dist: miniworld>=2.1; extra == "miniworld"
+Requires-Dist: pyglet<2,>=1.5.27; extra == "miniworld"
+Requires-Dist: numba>=0.60; extra == "miniworld"
+Provides-Extra: hf
+Requires-Dist: datasets>=2.14; extra == "hf"
+Provides-Extra: examples
+Requires-Dist: imageio; extra == "examples"
+Requires-Dist: imageio-ffmpeg; extra == "examples"
+Requires-Dist: opencv-python-headless; extra == "examples"
+Provides-Extra: docs
+Requires-Dist: mkdocs<2,>=1.6; extra == "docs"
+Requires-Dist: mkdocs-material<10,>=9.7; extra == "docs"
+Requires-Dist: mkdocstrings[python]>=0.27; extra == "docs"
+Requires-Dist: pymdown-extensions>=10.0; extra == "docs"
+Provides-Extra: dev
+Requires-Dist: build>=1.2; extra == "dev"
+Requires-Dist: pre-commit>=4.0; extra == "dev"
+Requires-Dist: pygame; extra == "dev"
+Requires-Dist: pytest>=8; extra == "dev"
+Requires-Dist: pytest-timeout>=2.3; extra == "dev"
+Requires-Dist: pytest-xdist>=3.6; extra == "dev"
+Requires-Dist: pyyaml>=6; extra == "dev"
+Requires-Dist: ruff>=0.12; extra == "dev"
+Requires-Dist: mypy>=1.15; extra == "dev"
+Requires-Dist: tomli>=2; python_version < "3.11" and extra == "dev"
+Requires-Dist: twine>=6; extra == "dev"
+Dynamic: license-file
+
+<img src="https://raw.githubusercontent.com/galilai-group/stable-envs/main/docs/assets/brand/banner.jpg" alt="stable-envs: frames from ten environment families" width="100%">
+
+<p align="center"><i>Fast and Reproducible Environments with Controllable Variations</i></p>
+
+<p align="center">
+  <a href="https://galilai-group.github.io/stable-envs/"><img alt="Docs" src="https://img.shields.io/badge/docs-galilai--group.github.io-black.svg"/></a>
+  <a href="https://pypi.org/project/stable-envs/"><img alt="PyPI" src="https://img.shields.io/pypi/v/stable-envs.svg?color=black"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/galilai-group/stable-envs/actions/workflows/ci.yml/badge.svg?branch=main"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/blob/main/pyproject.toml"><img alt="Python 3.10 to 3.12" src="https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12-black.svg"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/license-MIT%20%2B%20LGPL--2.1-black.svg"/></a>
+</p>
+
+<p align="center">
+  <a href="#installation"><b>Installation</b></a> ·
+  <a href="#quick-start"><b>Quick Start</b></a> ·
+  <a href="#environments"><b>Environments</b></a> ·
+  <a href="https://galilai-group.github.io/stable-envs/"><b>Documentation</b></a>
+</p>
+
+---
+
+stable-envs is a library of batched simulation environments that keep the full simulator state, so you can restore any step exactly and render images from it later, on CPU or GPU.
+
+Store states instead of pixels, replay or branch from any step, and render frames from a stored state as many times as you like. 80 environment classes in 11 families share the same API.
+
+## Installation
+
+```bash
+pip install stable-envs                       # or: uv add stable-envs
+pip install "stable-envs[pusht]"              # PushT simulation and CPU rendering
+pip install "stable-envs[dmcontrol,ogbench]"  # add family extras as needed
+stable-envs doctor                            # check what works on this machine
+```
+
+Install the `robotics` extra (Fetch) in its own environment, because it needs MuJoCo below 3.12, while `dmcontrol`, `mshuman`, `roadster` and `ogbench-manipulation` need 3.17. [Install](https://galilai-group.github.io/stable-envs/install/) lists every extra and the OpenGL setup.
+
+## Quick Start
+
+```python
+import torch
+import stable_envs as snv
+
+with snv.CartPole(num_envs=8, device="cpu") as env:
+    env.reset(seed=0)
+    actions = torch.randint(0, 2, (8, 10, 1))  # (B, T, A): ten steps for each environment
+    out = env.step(actions, frames=None)  # step without rendering
+    print(out["states"].batch_size)  # torch.Size([8, 10])
+    frames = env.render(out["states"][:, 4])  # render the state after five steps
+    print(frames.shape, frames.dtype)  # torch.Size([8, 400, 600, 3]) torch.uint8
+```
+
+`step` returns the state after every step, and `render` draws a batch of frames from any of them.
+
+The [Quickstart](https://galilai-group.github.io/stable-envs/quickstart/) covers state, frames, steps, save and restore, and variation. For policies, datasets and training, see [stable-worldmodel](https://github.com/galilai-group/stable-worldmodel).
+
+## Environments
+
+| Family | Classes | Extra |
+|---|---:|---|
+| [`classic_control`](https://galilai-group.github.io/stable-envs/environments/classic-control/) | 5 | none (`classic` for MountainCar, MountainCarContinuous and Pendulum) |
+| [`craftax`](https://galilai-group.github.io/stable-envs/environments/craftax/) | 4 | `craftax` |
+| [`dmcontrol`](https://galilai-group.github.io/stable-envs/environments/dmcontrol/) | 17 | `dmcontrol` |
+| [`fetch`](https://galilai-group.github.io/stable-envs/environments/fetch/) | 12 | `robotics` |
+| [`miniworld`](https://galilai-group.github.io/stable-envs/environments/miniworld/) | 1 | `miniworld` |
+| [`mshuman`](https://galilai-group.github.io/stable-envs/environments/mshuman/) | 1 | `mshuman` |
+| [`navigation`](https://galilai-group.github.io/stable-envs/environments/navigation/) | 5 | none (`navigation` for SimplePointMaze) |
+| [`ogbench`](https://galilai-group.github.io/stable-envs/environments/ogbench/) | 27 | `ogbench`, or `ogbench-manipulation` for Cube and Scene |
+| [`pusht`](https://galilai-group.github.io/stable-envs/environments/pusht/) | 2 | `pusht` |
+| [`pyflyt`](https://galilai-group.github.io/stable-envs/environments/pyflyt/) | 1 | `rocket` |
+| [`roadster`](https://galilai-group.github.io/stable-envs/environments/roadster/) | 5 | `roadster` |
+| **Total** | **80** | |
+
+## Examples
+
+- [`examples/`](https://github.com/galilai-group/stable-envs/blob/main/examples/README.md): short scripts for rendering, batching, action sequences and domain randomization.
+- [`examples/lewm/`](https://github.com/galilai-group/stable-envs/blob/main/examples/lewm/README.md): train and evaluate stable-worldmodel's LeWM world model on PushT, with frames that `env.render` draws on the training device from stored states.
+
+## Documentation
+
+The full documentation is at [galilai-group.github.io/stable-envs](https://galilai-group.github.io/stable-envs/).
+
+## Questions
+
+Open an [issue](https://github.com/galilai-group/stable-envs/issues) on GitHub.
+
+## License
+
+MIT, except the three PushT rasterizer files that port pygame's drawing routines, which are LGPL-2.1-or-later. See [LICENSE](https://github.com/galilai-group/stable-envs/blob/main/LICENSE) and [THIRD_PARTY_NOTICES.md](https://github.com/galilai-group/stable-envs/blob/main/THIRD_PARTY_NOTICES.md).

stable_envs-0.1.0/README.md

@@ -0,0 +1,88 @@
+<img src="https://raw.githubusercontent.com/galilai-group/stable-envs/main/docs/assets/brand/banner.jpg" alt="stable-envs: frames from ten environment families" width="100%">
+
+<p align="center"><i>Fast and Reproducible Environments with Controllable Variations</i></p>
+
+<p align="center">
+  <a href="https://galilai-group.github.io/stable-envs/"><img alt="Docs" src="https://img.shields.io/badge/docs-galilai--group.github.io-black.svg"/></a>
+  <a href="https://pypi.org/project/stable-envs/"><img alt="PyPI" src="https://img.shields.io/pypi/v/stable-envs.svg?color=black"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/galilai-group/stable-envs/actions/workflows/ci.yml/badge.svg?branch=main"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/blob/main/pyproject.toml"><img alt="Python 3.10 to 3.12" src="https://img.shields.io/badge/python-3.10%20%7C%203.11%20%7C%203.12-black.svg"/></a>
+  <a href="https://github.com/galilai-group/stable-envs/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/license-MIT%20%2B%20LGPL--2.1-black.svg"/></a>
+</p>
+
+<p align="center">
+  <a href="#installation"><b>Installation</b></a> ·
+  <a href="#quick-start"><b>Quick Start</b></a> ·
+  <a href="#environments"><b>Environments</b></a> ·
+  <a href="https://galilai-group.github.io/stable-envs/"><b>Documentation</b></a>
+</p>
+
+---
+
+stable-envs is a library of batched simulation environments that keep the full simulator state, so you can restore any step exactly and render images from it later, on CPU or GPU.
+
+Store states instead of pixels, replay or branch from any step, and render frames from a stored state as many times as you like. 80 environment classes in 11 families share the same API.
+
+## Installation
+
+```bash
+pip install stable-envs                       # or: uv add stable-envs
+pip install "stable-envs[pusht]"              # PushT simulation and CPU rendering
+pip install "stable-envs[dmcontrol,ogbench]"  # add family extras as needed
+stable-envs doctor                            # check what works on this machine
+```
+
+Install the `robotics` extra (Fetch) in its own environment, because it needs MuJoCo below 3.12, while `dmcontrol`, `mshuman`, `roadster` and `ogbench-manipulation` need 3.17. [Install](https://galilai-group.github.io/stable-envs/install/) lists every extra and the OpenGL setup.
+
+## Quick Start
+
+```python
+import torch
+import stable_envs as snv
+
+with snv.CartPole(num_envs=8, device="cpu") as env:
+    env.reset(seed=0)
+    actions = torch.randint(0, 2, (8, 10, 1))  # (B, T, A): ten steps for each environment
+    out = env.step(actions, frames=None)  # step without rendering
+    print(out["states"].batch_size)  # torch.Size([8, 10])
+    frames = env.render(out["states"][:, 4])  # render the state after five steps
+    print(frames.shape, frames.dtype)  # torch.Size([8, 400, 600, 3]) torch.uint8
+```
+
+`step` returns the state after every step, and `render` draws a batch of frames from any of them.
+
+The [Quickstart](https://galilai-group.github.io/stable-envs/quickstart/) covers state, frames, steps, save and restore, and variation. For policies, datasets and training, see [stable-worldmodel](https://github.com/galilai-group/stable-worldmodel).
+
+## Environments
+
+| Family | Classes | Extra |
+|---|---:|---|
+| [`classic_control`](https://galilai-group.github.io/stable-envs/environments/classic-control/) | 5 | none (`classic` for MountainCar, MountainCarContinuous and Pendulum) |
+| [`craftax`](https://galilai-group.github.io/stable-envs/environments/craftax/) | 4 | `craftax` |
+| [`dmcontrol`](https://galilai-group.github.io/stable-envs/environments/dmcontrol/) | 17 | `dmcontrol` |
+| [`fetch`](https://galilai-group.github.io/stable-envs/environments/fetch/) | 12 | `robotics` |
+| [`miniworld`](https://galilai-group.github.io/stable-envs/environments/miniworld/) | 1 | `miniworld` |
+| [`mshuman`](https://galilai-group.github.io/stable-envs/environments/mshuman/) | 1 | `mshuman` |
+| [`navigation`](https://galilai-group.github.io/stable-envs/environments/navigation/) | 5 | none (`navigation` for SimplePointMaze) |
+| [`ogbench`](https://galilai-group.github.io/stable-envs/environments/ogbench/) | 27 | `ogbench`, or `ogbench-manipulation` for Cube and Scene |
+| [`pusht`](https://galilai-group.github.io/stable-envs/environments/pusht/) | 2 | `pusht` |
+| [`pyflyt`](https://galilai-group.github.io/stable-envs/environments/pyflyt/) | 1 | `rocket` |
+| [`roadster`](https://galilai-group.github.io/stable-envs/environments/roadster/) | 5 | `roadster` |
+| **Total** | **80** | |
+
+## Examples
+
+- [`examples/`](https://github.com/galilai-group/stable-envs/blob/main/examples/README.md): short scripts for rendering, batching, action sequences and domain randomization.
+- [`examples/lewm/`](https://github.com/galilai-group/stable-envs/blob/main/examples/lewm/README.md): train and evaluate stable-worldmodel's LeWM world model on PushT, with frames that `env.render` draws on the training device from stored states.
+
+## Documentation
+
+The full documentation is at [galilai-group.github.io/stable-envs](https://galilai-group.github.io/stable-envs/).
+
+## Questions
+
+Open an [issue](https://github.com/galilai-group/stable-envs/issues) on GitHub.
+
+## License
+
+MIT, except the three PushT rasterizer files that port pygame's drawing routines, which are LGPL-2.1-or-later. See [LICENSE](https://github.com/galilai-group/stable-envs/blob/main/LICENSE) and [THIRD_PARTY_NOTICES.md](https://github.com/galilai-group/stable-envs/blob/main/THIRD_PARTY_NOTICES.md).

stable_envs-0.1.0/THIRD_PARTY_NOTICES.md

@@ -0,0 +1,235 @@
+# Third-party notices
+
+## envX TwoRoom
+
+`stable_envs/envs/navigation/tworoom_jax.py` adapts the collision rules and integration
+logic in `envx/tworooms/geometry.py` and `envx/tworooms/environment.py` from
+[fracapuano/envX](https://github.com/fracapuano/envX) at commit
+`d26c817c1124229c819be574f137cd68bda4380a` (Copyright 2026 Francesco Capuano, Apache-2.0).
+Modifications support batched Variation, boundary rounding and DLPack exchange with Torch.
+Shared navigation code handles masked steps, observations and reset/State/task contracts.
+envX is not a runtime dependency.
+The Apache license is included in `licenses/APACHE-2.0.txt` and distributed with the package.
+
+## envX Reacher
+
+`stable_envs/envs/dmcontrol/simulation/reacher_mjx.py` adapts `src/envx/reacher/env.py`
+from the envX commit above under Apache-2.0.
+
+## DeepMind Control Suite
+
+`stable_envs/envs/dmcontrol/simulation/mjwarp_tasks.py`, `batch.py`, `first_family.py` and `humanoid.py`
+adapt task observations and reward rules from the
+[DeepMind Control Suite](https://github.com/google-deepmind/dm_control/tree/main/dm_control/suite) of
+dm-control 1.0.50 (Copyright 2017 The dm_control Authors, Apache-2.0) to batched NumPy and Torch arrays.
+dm_control is a runtime dependency of the `dmcontrol` extra; these adaptations do not import its task code.
+The Apache-2.0 license text is distributed in `licenses/APACHE-2.0.txt`; the copyright notices
+above apply to these adaptations.
+
+## envX PushT
+
+`stable_envs/envs/pusht/simulation/jax_physics.py` adapts
+`src/envx/pusht/physics.py`
+from the envX commit above under Apache-2.0. `numba_physics.py` and `triton_physics.py` in the same
+directory port that adaptation to Numba and Triton.
+
+## Munk2D
+
+The PushT physics ports in `stable_envs/envs/pusht/simulation/` port parts of
+[Munk2D](https://github.com/viblo/Munk2D) at commit `ade7ed72849e60289eefb7a41e79ae6322fefaf3`, the
+Chipmunk2D fork that pymunk 7.2.0 builds: the wall contacts (`SegmentToPoly` and `ContactPoints` from
+`src/cpCollision.c`), the arbiter contact cache (`src/cpArbiter.c`, `src/cpSpaceStep.c`) and the
+bounding-box tree broadphase (`src/cpBBTree.c`, in `numba_broadphase.py` and `triton_broadphase.py`).
+Munk2D is MIT-licensed ("Copyright (c) 2025 Victor Blomqvist, Copyright (c) 2007-2024 Scott Lembcke and
+Howling Moon Software"). The license text is distributed in `licenses/Munk2D-MIT.txt`.
+
+## Craftax
+
+The `craftax` family (`stable_envs/envs/craftax/`) runs
+[Craftax](https://github.com/MichaelTMatthews/Craftax), which is MIT-licensed ("Copyright (c) 2024
+Michael Matthews"), as an external dependency. Its simulator, renderer and textures are not copied
+here; frames are the pixel observations Craftax computes. The per-lane key handling and the uint8
+frame scaling follow stable-worldmodel's `stable_worldmodel/envs/craftax/craftax.py` at commit
+`21446f1ede6d5284e981bd7b47f432b994e6d812` (MIT).
+
+## Miniworld
+
+The `miniworld` family (`stable_envs/envs/miniworld/`) runs `MovingBlockWorld` from
+[hlillemark/Miniworld](https://github.com/hlillemark/Miniworld), a fork of Farama's
+[Miniworld](https://github.com/Farama-Foundation/Miniworld) (Apache-2.0), as an external dependency, at
+the head of its `flowm_release` branch, the code frozen for the Flow Equivariant World Models paper. Its
+renderer and textures are not copied here. `stable_envs/envs/miniworld/simulation/kernel_cpu.py`,
+`kernel_torch.py` and `kernel_triton.py` port the step rules of `MiniWorldEnv.step` and
+`MovingBlockWorld.step` to Numba, Torch and Triton. Modifications: batched over environments, and the reset
+draws from a counter-based generator instead of the source's NumPy generator. The Apache-2.0 license text is
+distributed in `licenses/APACHE-2.0.txt`. The `flowm` preset and the heading handling follow the
+fork's `scripts/generate_videos.py` at commit `bc7972050155b67e242c07bfdccfe1eeddbf9882`.
+
+## pygame drawing routines (LGPL)
+
+`stable_envs/envs/pusht/raster.py`, `raster_cpu.py` and `raster_device.py` port the
+`draw_fillpoly`, `draw_circle_filled`, `draw_line_width` and `draw_aaline` routines of
+`src_c/draw.c` from [pygame](https://github.com/pygame/pygame) 2.6.1
+(Copyright (C) 2000-2001 Pete Shinners, GNU Library General Public License version 2 or later).
+Modifications: the routines are batched over frames, evaluated only at the pixels that the resize reads,
+and run in Numba and Torch. These three files are licensed under LGPL-2.1-or-later, not MIT; each file
+starts with an SPDX line that says so. The license text is in `licenses/LGPL-2.1.txt`.
+pygame is not a runtime dependency of these files.
+
+## SDL_gfx drawing rules
+
+`stable_envs/render/gfxdraw/` reimplements in Torch the integer rasterization and
+packed-channel alpha-blending rules of SDL_gfx by Andreas Schiffler, which pygame's
+`gfxdraw` module uses. Upstream [SDL2_gfx](https://www.ferzkopp.net/wordpress/2016/01/02/sdl_gfx-sdl2_gfx/)
+is licensed under the zlib license; the copy bundled in pygame (`src_c/SDL_gfx/`) is
+marked LGPL. No SDL_gfx source is included.
+
+## stable-worldmodel
+
+Some modules are adapted from
+[stable-worldmodel](https://github.com/galilai-group/stable-worldmodel/tree/21446f1ede6d5284e981bd7b47f432b994e6d812)
+at commit `21446f1ede6d5284e981bd7b47f432b994e6d812`. Its authors are Lucas Maes,
+Quentin Le Lidec, Randall Balestriero and contributors. It is MIT-licensed
+("Copyright (c) 2026 GalilAI-group").
+
+| Module in stable-envs | Source in stable-worldmodel |
+|---|---|
+| `stable_envs/spaces.py` | `stable_worldmodel/spaces.py` |
+| `stable_envs/envs/pusht/simulation/task.py` | `stable_worldmodel/envs/pusht/env.py`, `env_discrete.py` |
+| `stable_envs/envs/pusht/simulation/quantizer.py` | `stable_worldmodel/envs/quantizer.py` |
+| `stable_envs/envs/pusht/draw.py` | `stable_worldmodel/envs/utils.py` (drawing subset) |
+| `stable_envs/envs/rocket/simulation/task.py` | `stable_worldmodel/envs/rocket_landing/pyflyt_rocketlanding.py` |
+
+The adapted code has no runtime import of stable-worldmodel. Physics stays in the
+external dependencies under their own licenses: Pymunk/Munk2D (MIT) for PushT, and
+PyFlyt (MIT) with PyBullet (zlib) for Rocket. No native physics code is copied here.
+
+MIT permission notice for the adapted code follows:
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+
+## MS-Human-700
+
+The `mshuman` family (`stable_envs/envs/mshuman/`) loads the locomotion model of
+[MS-Human-700](https://github.com/LNSGroup/MS-Human-700) at commit
+`2d686957aefd5739cf4d2859a6acfd94c8f84400` (Apache-2.0). The MJCF and its assets are fetched at
+runtime, not vendored here; `MSHumanLocomotion` steps and draws them through MuJoCo Warp and MuJoCo.
+The Apache license is included in `licenses/APACHE-2.0.txt` and distributed with the package.
+
+## MuJoCo
+
+`stable_envs/render/mujoco_scene.py` and `stable_envs/render/mujoco_raster.py` follow MuJoCo's classic
+fixed-function OpenGL renderer (`src/render/render_gl3.c`, `src/render/render_context.c`) of
+[MuJoCo](https://github.com/google-deepmind/mujoco) 3.17.0 (Copyright 2021 DeepMind Technologies Limited,
+Apache-2.0). `stable_envs/envs/fetch/raster.py` and `stable_envs/envs/fetch/transparency.py` follow the same
+renderer at MuJoCo 3.11.0. Modifications: the scene extraction, lighting, shadow map and translucent-geom
+ordering are reimplemented in NumPy, Numba and Warp to draw batches of frames on CPU or CUDA. MuJoCo is a
+runtime dependency of the MuJoCo families; no MuJoCo C source is included. The Apache-2.0 license text is
+distributed in `licenses/APACHE-2.0.txt`.
+
+## MuJoCo Warp
+
+[MuJoCo Warp](https://github.com/google-deepmind/mujoco_warp) 3.12.0 and 3.17.0 (Apache-2.0; "Copyright 2025 The
+Newton Developers" in `smooth.py`, "Copyright 2026 The Newton Developers" in `render.py` and `render_util.py`) is
+adapted in these files:
+
+| Module in stable-envs | Source in MuJoCo Warp |
+|---|---|
+| `stable_envs/_execution/mjwarp_fused.py` | `mujoco_warp/_src/smooth.py` (`com_pos`, `crb`, `_rne_cfrc_backward`) |
+| `stable_envs/_execution/mjwarp_tree.py` | `mujoco_warp/_src/smooth.py` (kinematics, `com_pos`, `com_vel`, `crb`, RNE) |
+| `stable_envs/render/mjwarp/__init__.py` | `mujoco_warp/_src/render_util.py` (`create_warp_texture`), `mujoco_warp/_src/render.py` (`pack_rgba_to_uint32`) |
+| `stable_envs/envs/dmcontrol/trackcom_reflection.py` | `mujoco_warp/_src/render.py` (ray cast, lighting and texture sampling of the reflected rays) |
+| `stable_envs/render/cpu3d.py`, `stable_envs/render/cpu_raster/scene3d.py` | `mujoco_warp/_src/render.py` (shading rules) |
+
+Modifications: the tree passes are fused into one launch per pass or one thread block per world, the texture
+upload is shared between textures and can be downsampled, the RGB packing rounds instead of truncating, and the
+shading rules are reimplemented in Torch and Numba for CPU. MuJoCo Warp is a runtime dependency of the CUDA
+backends. The Apache-2.0 license text is distributed in `licenses/APACHE-2.0.txt`.
+
+## PyFlyt
+
+`stable_envs/envs/rocket/simulation/numba_physics.py` and `numba_cuda_physics.py` port the Rocket landing
+physics and task of [PyFlyt](https://github.com/jjshoots/PyFlyt) 0.29.0 (`PyFlyt/core/drones/rocket.py`,
+`PyFlyt/gym_envs/rocket_envs/rocket_base_env.py`, `rocket_landing_env.py`, MIT). Modifications: the forces
+and the task logic are rewritten as Numba CPU and CUDA kernels that step many environments in parallel.
+PyFlyt is a runtime dependency of the `rocket` extra for `simulation_backend="pybullet"`. The PyFlyt 0.29.0
+license file has no copyright line; its text is distributed in `licenses/PyFlyt-MIT.txt`.
+
+## Bullet, PyBullet and TinyRenderer
+
+`stable_envs/envs/rocket/tiny_raster.py` is an altered Numba port of the TinyRenderer of
+[Bullet](https://github.com/bulletphysics/bullet3) as built in pybullet 3.2.7
+(`examples/TinyRenderer/our_gl.cpp`, `TinyRenderer.cpp`, `model.cpp`,
+`TinyRendererVisualShapeConverter.cpp`) and of the convex
+hull that Bullet builds for URDF cylinders (`btShapeHull`, `HullLibrary`). `stable_envs/envs/rocket/raster.py`,
+`stable_envs/envs/rocket/render.py` and the TinyRenderer lighting in `stable_envs/render/nvdiff.py` follow the
+same renderer. `stable_envs/envs/rocket/simulation/numba_physics.py` and `numba_cuda_physics.py` follow
+Bullet's single-body Featherstone integration and multibody contact solver. These files are altered versions:
+they are not the original Bullet software. Modifications: the arithmetic is rewritten in Numba and
+CUDA to draw and step many environments in parallel. Bullet and pybullet are licensed under the zlib license
+(pybullet authors: Erwin Coumans, Yunfei Bai, Jasmine Hsu). pybullet is a runtime dependency of the `rocket`
+extra. The license text is distributed in `licenses/Bullet-zlib.txt`.
+
+## Gymnasium
+
+`stable_envs/envs/classic_control/simulation/batched.py`, `triton_step.py` and `device.py` port the `step`
+functions of the classic-control environments of [Gymnasium](https://github.com/Farama-Foundation/Gymnasium)
+1.3.0 (`gymnasium/envs/classic_control/`, MIT, "Copyright (c) 2016 OpenAI, Copyright (c) 2022 Farama
+Foundation"). `stable_envs/render/gfxdraw/` draws the frames of their `render` functions. Modifications: the
+steps run on batches in NumPy, Torch and Triton in the source's operation order. Gymnasium is a runtime
+dependency, and CPU resets stay in Gymnasium. The license text is distributed in `licenses/Gymnasium-MIT.txt`.
+
+## Gymnasium-Robotics
+
+`stable_envs/envs/fetch/simulation/batch.py`, `mjwarp.py` and `observations.py` port the step, reset and
+observation code of the Fetch environments of
+[Gymnasium-Robotics](https://github.com/Farama-Foundation/Gymnasium-Robotics) 1.4.2
+(`gymnasium_robotics/envs/fetch/`, MIT, "Copyright (c) 2022 Farama Foundation"). Modifications: the step and
+reset run on arrays of environments, with MuJoCo in one batched call or MuJoCo Warp on CUDA.
+Gymnasium-Robotics is a runtime dependency of the `robotics` extra. The license text is distributed in
+`licenses/Gymnasium-Robotics-MIT.txt`.
+
+## OGBench
+
+`stable_envs/envs/ogbench/simulation/manipulation_control.py`, `manipulation_batch.py`,
+`manipulation_warp.py` and `maze_batch.py` port the controller, step, observation and task code of
+[OGBench](https://github.com/seohongpark/ogbench) 1.2.1 (`ManipSpaceEnv`, the locomotion maze
+environments, MIT, "Copyright (c) 2024 OGBench Authors"). Modifications: the code runs on batches of
+environments, with MuJoCo in one batched call or MuJoCo Warp and Warp kernels on CUDA. OGBench is a runtime
+dependency of the `ogbench` extras. The license text is distributed in `licenses/OGBench-MIT.txt`.
+
+## NumPy random
+
+`stable_envs/envs/rocket/simulation/numba_physics.py` ports NumPy's `PCG64.next_uint64` and the ziggurat
+`Generator.standard_normal`, and `stable_envs/envs/ogbench/simulation/maze_batch.py` ports the MT19937
+`RandomState` seeding, generation, `uniform` and `randint`, from [NumPy](https://github.com/numpy/numpy)
+(`numpy/random/`). NumPy is BSD-3-Clause ("Copyright (c) 2005-2025, NumPy Developers"), and its
+`numpy/random` module is dual-licensed NCSA and BSD-3-Clause ("Copyright (c) 2019 Kevin Sheppard");
+MT19937 is by Makoto Matsumoto and Takuji Nishimura, PCG64 by Melissa O'Neill. Modifications: the generators
+are rewritten in Numba and NumPy to draw the same numbers for many environments at once. NumPy is a runtime
+dependency. The license texts are distributed in `licenses/NumPy-BSD-3-Clause.txt` and
+`licenses/NumPy-random-NCSA-BSD-3-Clause.txt`.
+
+## OpenCV
+
+`stable_envs/envs/pusht/raster.py` and `raster_cpu.py` reimplement the fixed-point
+`cv2.resize` `INTER_LINEAR` arithmetic of [OpenCV](https://github.com/opencv/opencv) 5.0
+(opencv-python-headless 5.0.0.93, Apache-2.0). Modifications: only the offsets and 11-bit weights are computed,
+in integer Torch and Numba code. OpenCV is a runtime dependency of the `pusht` extra; no OpenCV
+source is included. The Apache-2.0 license text is distributed in `licenses/APACHE-2.0.txt`.