Installation

PACE on mjlab is supported on 64-bit Linux with Python 3.10–3.13. CUDA is strongly recommended for fitting; CPU is suitable only for smoke tests. Plan for an NVIDIA driver compatible with the installed PyTorch/CUDA runtime, at least 16 GB system RAM, and roughly 8–10 GB free disk space for the virtual environment and kernel cache.

Install uv first, then create the project-local environment:

git clone https://github.com/fan-ziqi/pace-sim2real-mjlab.git
cd pace-sim2real-mjlab
uv sync --group dev

The environment is stored at .venv. Use uv run <command> or activate it:

source .venv/bin/activate

The project pins mjlab==1.5.2 and uses its MuJoCo-Warp CUDA runtime. Check the environment with uv run python scripts/list_envs.py; it must list Isaac-Pace-Anymal-D-v0.

For CPU-only debugging, append --device cpu to PACE scripts. Production fitting should use CUDA because the CMA-ES population is evaluated in parallel. The first CUDA launch compiles MuJoCo-Warp kernels and may take several minutes.

The lock file is generated against the public PyPI registry. If your network requires an internal mirror, configure uv at your organization level and then run uv lock so that its source is explicit in your own lock file.

Published documentation is available at fan-ziqi.github.io/pace-sim2real-mjlab.