DriveRL Environment

Use one Python 3.11 environment for DriveRL and the nuPlan companion.

component	version
Python	3.11
NumPy	1.26.4
PyTorch	2.7.1
Hydra / OmegaConf	1.3.2 / 2.3.0
Ray	2.51.1
Linux GPU CUDA	12.8

GPU installation

python3.11 -m venv .venv_driverl
source .venv_driverl/bin/activate
python -m pip install --upgrade pip

DRIVERL_REPO=/path/to/DriveRL
python -m pip install "torch==2.7.1" \
  --index-url https://download.pytorch.org/whl/cu128 \
  -c "$DRIVERL_REPO/constraints.txt"
python -m pip install -r "$DRIVERL_REPO/requirements.txt" \
  -c "$DRIVERL_REPO/constraints.txt"
python -m pip install -e "$DRIVERL_REPO" --no-deps
python -m pip install -e "$DRIVERL_REPO/nuplan-devkit" --no-deps

Verify the selected PyTorch build:

python - <<'PY'
import torch
assert torch.__version__.split("+", 1)[0] == "2.7.1"
assert torch.version.cuda == "12.8"
print(f"torch={torch.__version__} cuda={torch.version.cuda}")
PY

CI installs the same version from the PyTorch CPU index.

Source selection

Select the DriveRL and companion source roots for nuPlan evaluation:

PYTHONPATH="/path/to/DriveRL/nuplan-devkit:/path/to/DriveRL/src" \
  python -m nuplan.planning.script.run_simulation --help

Ray and NuBoard

For distributed evaluation, provide a private RAY_REDIS_PASSWORD. Ray's dashboard is disabled by default. Set DRIVERL_EVAL_RAY_INCLUDE_DASHBOARD=1 for local diagnostics and keep its host at 127.0.0.1.

Set a short writable directory for Ray session files:

export DRIVERL_EVAL_RAY_TEMP_DIR=/path/to/short-runtime-tmp/driverl-ray-n${MLP_ROLE_INDEX:-0}

NuBoard uses loopback by default. A reverse proxy can provide an explicit NUPLAN_NUBOARD_WEBSOCKET_ORIGINS list.