| [build-system] | |
| requires = ["uv_build>=0.8.19,<0.9.0"] | |
| build-backend = "uv_build" | |
| [project] | |
| name = "mjlab-microduck" | |
| version = "0.1.0" | |
| description = "RL training environments for the Microduck robot, built on mjlab" | |
| readme = "README.md" | |
| license = "Apache-2.0" | |
| license-files = ["LICENSE"] | |
| # <3.13: bam (better-actuator-models) pins requires-python <3.13, and jobs must | |
| # run the same interpreter we test locally (3.12) — a floating upper bound let | |
| # HF jobs pick 3.13.14 (2026-07-21). | |
| requires-python = ">=3.12, <3.13" | |
| dependencies = [ | |
| "mjlab==1.3.0", | |
| "warp-lang==1.12.0", | |
| # BAM distribution name is `better-actuator-models`; the import stays `bam`. | |
| "better-actuator-models", | |
| "onnxruntime>=1.24.4", | |
| "rustypot>=1.4.2", | |
| "huggingface_hub>=0.27.0", | |
| "matplotlib>=3.10.9", | |
| # mjlab 1.3.0 imports scipy (terrains/heightfield_terrains.py) but forgets | |
| # to declare it — without this line a fresh `uv sync` (e.g. on HF jobs) | |
| # can't even `import mjlab_microduck` (found 2026-07-21). | |
| "scipy>=1.16", | |
| # Direct dep ONLY so [tool.uv.sources] can bind torch to the CUDA index on | |
| # aarch64 (see the torch source below) — uv applies sources to DIRECT | |
| # dependencies only, so as a purely transitive dep (mjlab, rsl_rl) the | |
| # source entry is silently ignored. | |
| # Pinned to the exact version uv.lock already resolved from PyPI so this | |
| # changes only the SOURCE of the wheel on aarch64, not the version: a | |
| # floating `>=` lets the CUDA index (which carries newer builds than the | |
| # PyPI pin) drag torch 2.9.1 -> 2.13.0, an unvetted bump for mjlab/rsl_rl. | |
| "torch==2.9.1", | |
| ] | |
| [project.entry-points."mjlab.tasks"] | |
| mjlab_microduck = "mjlab_microduck.tasks" | |
| [project.scripts] | |
| # Shadows mjlab's `train` entry point: identical behavior, plus a --hf-jobs | |
| # flag that submits the run to Hugging Face Jobs instead (see train_cli.py). | |
| train = "mjlab_microduck.train_cli:main" | |
| [tool.ruff] | |
| src = ["src"] # Helpful for recognizing first-party imports. | |
| indent-width = 4 | |
| exclude = [ | |
| "src/mjlab/third_party", | |
| "typings", | |
| ] | |
| [tool.uv] | |
| override-dependencies = [ | |
| # NOTE: do NOT override mujoco — mjlab 1.3.0's mujoco-warp pins a compatible | |
| # mujoco (3.10.x); an override here previously forced mujoco down to 3.4.0, | |
| # which lacks mjDSBL_MULTICCD that mujoco-warp 3.8.1 imports → import crash. | |
| # bam pins protobuf<4.0 (for its zmq/dynamixel messaging, which we don't | |
| # use — we only import bam.model/bam.actuator). Without this override the | |
| # downgrade cascades onnx down to 1.17.0, which has no wheel and fails to | |
| # build from source. | |
| "protobuf>=4.0,<7.0", | |
| # Keep onnx on a version with prebuilt wheels (1.17.0 has no py3.13 wheel and | |
| # fails to build from source). Pulling in bam otherwise nudges it downward. | |
| "onnx>=1.20.1", | |
| # bam depends on the PyPI stub `zmq==0.0.0`, whose prebuilt wheel is invalid | |
| # (missing .dist-info) -> `uv sync` fails on a FRESH install (e.g. HF Jobs | |
| # remote build), even though a warm local cache tolerated it. We only use | |
| # bam.mjlab (no zmq/erob messaging), so drop zmq via an always-false marker. | |
| "zmq ; python_version < '3.0'", | |
| ] | |
| [tool.uv.sources] | |
| # mjlab now comes from PyPI (==1.3.0, pinned in [project.dependencies]); no git source. | |
| # Official BAM actuator model. Switch to `main` (or a tag) once the | |
| # mjlab-facing API lands there. | |
| better-actuator-models = { git = "https://github.com/Rhoban/bam.git", branch = "mjlab_frictionloss" } | |
| # For local BAM development, comment the line above and use: | |
| # better-actuator-models = { path = "/home/antoine/Rhoban/bam", editable = true } | |
| # On linux-aarch64 (DGX Spark / GB10) PyPI's torch wheel is CPU-ONLY: | |
| # torch.__version__ == "2.9.1+cpu", torch.version.cuda is None, so | |
| # torch.cuda.device_count() == 0 and mjlab's select_gpus() indexes an empty | |
| # list -> `IndexError: list index out of range` before training even starts | |
| # (mjlab/utils/gpu.py:70). Route torch to PyTorch's CUDA index there. | |
| # cu129 (not cu130) matches the CUDA toolkit warp 1.12.0 bundles, so the | |
| # zero-copy warp<->torch interop stays on one runtime major version. | |
| # The marker keeps x86_64 (HF Jobs) on PyPI, where the wheel already bundles | |
| # CUDA via its nvidia-*-cu12 deps — that resolution is unchanged. | |
| torch = [ | |
| { index = "pytorch-cu129", marker = "sys_platform == 'linux' and platform_machine == 'aarch64'" }, | |
| ] | |
| [[tool.uv.index]] | |
| name = "pytorch-cu129" | |
| url = "https://download.pytorch.org/whl/cu129" | |
| # explicit: only packages that name this index resolve from it. | |
| explicit = true | |