Install and Download
Use Linux x86_64 and Python 3.12, with compatible GPU drivers and Vulkan support for simulation rendering. You also need Git and Docker to run the motion planner. See the simulation prerequisites for graphics setup.
1. Install the benchmark
Create a workspace and install the exact packages required by benchmark v1.0.0:
mkdir -p ~/magma-benchmark
cd ~/magma-benchmark
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install "magma_bench==2.0.1" "magma_core[simulation]==2.0.1" "magma_scenarios==2.1.1"
python -m pip check
This installs the benchmark runner and simulation dependencies. The benchmark files, agent server, and model weights are separate.
2. Download the benchmark
magma-bench fetch --version 1.0.0
The command prints the installed benchmark directory. By default it is ~/.cache/magma-bench/datasets/v1.0.0/, or $XDG_CACHE_HOME/magma-bench/datasets/v1.0.0/ if that environment variable is set.
Each version has its own directory. Fetching the same version again checks and reuses the complete download. If a download is damaged, fetch into a new empty directory with --path.
magma-bench fetch without --version downloads the latest stable benchmark release. Use an explicit version to keep your benchmark and package versions aligned.
Choose the download directory
To keep the benchmark in your workspace instead of the default cache:
magma-bench fetch --version 1.0.0 --path ./benchmark-v1.0.0
Use --benchmark-root ./benchmark-v1.0.0 when running this copy. The CLI option --dataset 1.0.0 selects benchmark v1.0.0 from the default cache. Evaluation reads local files and never downloads the benchmark itself.
3. Start the motion planner
In another terminal:
cd ~/magma-benchmark
git clone https://github.com/MAGMA-rob/magma-planner-mplib.git
cd magma-planner-mplib
bash scripts/launch_planner.bash -p 8000
Wait for server startup and leave this terminal running. The first launch builds the Docker image. Docker keeps the planner's dependencies separate from the simulation environment.
If a planner is already running, reuse its address. The default is http://localhost:8000.