Docker Installation for MME-VLA Policy Learning

This guide explains how to set up Docker and NVIDIA GPU support so you can build and run the MME-VLA image.

1) Install Docker Engine

Skip this if you have already installed Docker.

Follow Docker’s official instructions for Ubuntu:

• Docker Engine install guide: https://docs.docker.com/engine/install/ubuntu/

After installing, make sure the service is running:

docker run --rm hello-world

2) Install NVIDIA Container Toolkit (GPU support)

Skip this if you have already installed nvidia-ctk.

Install the toolkit (Ubuntu):

curl -fsSL https://nvidia.github.io/libnvidia-container/gpgkey | \
  sudo gpg --dearmor --batch --yes -o /usr/share/keyrings/nvidia-container-toolkit-keyring.gpg

curl -fsSL https://nvidia.github.io/libnvidia-container/stable/deb/nvidia-container-toolkit.list | \
  sed 's#deb https://#deb [signed-by=/usr/share/keyrings/nvidia-container-toolkit-keyring.gpg] https://#g' | \
  sudo tee /etc/apt/sources.list.d/nvidia-container-toolkit.list

sudo apt-get update
sudo apt-get install -y nvidia-container-toolkit

Configure Docker to use the NVIDIA runtime and restart Docker:

sudo nvidia-ctk runtime configure --runtime=docker
sudo systemctl restart docker

Verify GPU access inside a container:

docker run --rm --gpus all nvidia/cuda:12.8.0-base-ubuntu24.04 nvidia-smi

3) Build the MME-VLA Docker image

From the repository root:

docker build -t <image_name>:<tag> .
# e.g., run `docker build -t mme_vla:cuda12.8 .`

Start the container:

export PORT=8001
docker run --rm -it --gpus all \
  -e NVIDIA_DRIVER_CAPABILITIES=compute,graphics,utility,video \
  -v "$PWD/runs:/app/runs" -v "$PWD/data:/app/data" \
  -p $PORT:$PORT \
  mme_vla:cuda12.8

-e sets an environment variable inside the container (e.g., NVIDIA_DRIVER_CAPABILITIES).
-v mounts a host path into the container as a bind mount. Here we mount the host ./runs and ./data directories to /app/runs and /app/data inside the container.

Because these are bind mounts, files you create/modify inside the container will be visible to you on the host (and vice versa) as they are the same underlying directories. -p publishes a container port to a host port (port mapping).

Permissions note: this image runs as root by default, so any new files created in the mounted runs/ or data/ dirs will become root-owned on the host.

To fix ownership on the host (run outside the container): sudo chown -R "$USER:$USER" runs data

Alternatively, run the container as your UID/GID to keep created files owned by you:

docker run --rm -it --gpus all \
  --user "$(id -u):$(id -g)" \
  -e NVIDIA_DRIVER_CAPABILITIES=compute,graphics,utility,video \
  -v "$PWD/runs:/app/runs" -v "$PWD/data:/app/data" \
  -p $PORT:$PORT \
  mme_vla:cuda12.8

If you use --user, apt-get update inside the container may fail because non-root users typically can’t write to /var/lib/apt.

4) Evaluate the policy

# terminal 0
CUDA_VISIBLE_DEVICES=0 uv run scripts/serve_policy.py --seed=7  --port=$PORT policy:checkpoint --policy.dir=runs/ckpts/mme_vla_suite/perceptual-framesamp-modul/79999 --policy.config=mme_vla_suite

# terminal 1 
eval "$(micromamba shell hook --shell bash)"
micromamba activate robomme
CUDA_VISIBLE_DEVICES=1 python examples/robomme/eval.py --args.model_seed=7 --args.port=$PORT --args.policy_name=perceptual-framesamp-modul --args.model_ckpt_id=79999

5) Other Hints

To stop the container:

docker ps
docker stop <container_id_or_name>

Alternatively, inside the container shell you can stop the session with exit (or Ctrl-D).

To rebuild the Docker image:

docker build --no-cache -t <image_name>:<tag> .

To detach from the running container (without stopping it), press Ctrl-p then Ctrl-q.

To re-attach the session:

docker ps
docker exec -it <container_id_or_name> bash

To start a detached container directly, use -d:

docker run -d --rm -it --gpus all ...

To install additional packages inside the container, run:

apt-get update
apt-get install <package>