Installation

RGBench supports three install paths. The one-shot script is the recommended default; the manual conda path is for users who want fine control over each dependency; the Docker path is a future option.

Path 1 — one-shot setup.sh (recommended)

git clone https://github.com/hwk0809/RGBench
cd RGBench
bash setup.sh

What it does:

1. Confirms Python ≥ 3.10.
2. Creates a virtualenv at .venv/ if one isn't already active.
3. Editable install with the [all] extras (PyBullet, MuJoCo, the segmentation pipeline, dev tools).
4. Downloads SAM + GroundingDINO checkpoints from the original authors into third_party/grounded_sam/checkpoints/.
5. Downloads the smoke-test sample from Hugging Face into data/sample/.
6. Runs examples/pybullet_quickstart.py and exits non-zero if it fails.

Path 2 — manual conda environment

git clone https://github.com/hwk0809/RGBench
cd RGBench
conda create -n rgbench python=3.11 -y
conda activate rgbench
pip install -e ".[pybullet,mujoco,segment,dev]"
python scripts/download_checkpoints.py
python scripts/download_data.py --sample-only
python examples/pybullet_quickstart.py

Optional extras groups:

Group	Adds
pybullet	pybullet
mujoco	mujoco
segment	groundingdino-py, segment-anything, pykeops (GPU chamfer)
dev	pytest, ruff, build
all	everything above

Path 3 — Isaac Sim

Isaac Sim cannot be installed via pip; install it from NVIDIA's Omniverse Launcher first, then point at its bundled Python:

export ISAACSIM_PYTHON=~/isaacsim/python.sh
make install-isaacsim
# or, equivalently:
"$ISAACSIM_PYTHON" -m pip install -r requirements-isaacsim.txt

Then run the benchmark with make benchmark sim=isaacsim — run_batch.py detects isaacsim as the backend and dispatches the child process under the Isaac interpreter.

Path 4 — Docker (planned for v0.2)

A Dockerfile and pre-built image will be published alongside the v0.2 release. For now use Path 1 or Path 2.

Troubleshooting

OSError: libGL.so.1: cannot open shared object file — install system OpenGL: sudo apt install libgl1-mesa-glx libegl1. setup.sh exports MUJOCO_GL=egl + LIBGL_ALWAYS_SOFTWARE=1 to avoid the most common GL issues on headless servers.

huggingface_hub 401 / private dataset — the dataset is public; if you see auth errors, log in with huggingface-cli login or point at a mirror by setting RGBENCH_HF_DATASET=<your-org>/<repo>.

Chamfer distance is slow — install pykeops (in the segment extras) for the GPU code path. CPU fallback via SciPy is correct but ~50× slower for large point clouds.

Isaac Sim crashes on launch — verify your driver and Isaac Sim versions match the matrix at https://docs.isaacsim.omniverse.nvidia.com/.