OpenArm LeRobot Policy Server

LeRobot policy servers for OpenArm inference with dora-rs.

They serve a pretrained LeRobot policy over a Unix domain socket to the policy server nodes used in dora-openarm-evaluation:

Setup

git clone git@github.com:enactic/openarm_lerobot_policy_server.git
cd openarm_lerobot_policy_server
uv sync
source .venv/bin/activate
# uv pip install --torch-backend=cu128 --reinstall torch torchvision  # for CUDA 12.8

Policy

The policy is enactic/act-openarm-2-cell-pick_up_cube_mujoco by default. To use another one, pass a Hugging Face Hub repository ID or a local directory with --policy, or set it to the OPENARM_LEROBOT_POLICY environment variable. The policy must take observation.state and the observation.images.{ceiling,head_left,head_right,wrist_left,wrist_right} camera images as its inputs.

Local inference

openarm-lerobot-policy-server-local /dev/shm/policy-server.socket
# openarm-lerobot-policy-server-local /dev/shm/policy-server.socket --policy ${POLICY}

Then run dataflow-local-inference.yaml in dora-openarm-evaluation with the same SOCKET.

Docker inference

Pull a released Docker image for your PyTorch backend:

docker pull ghcr.io/enactic/openarm_lerobot_policy_server:latest

${BACKEND}-${VERSION} such as cu128-1.0.0 is also available for each release. main and ${BACKEND}-main are built from the main branch.

Or build the Docker image:

docker build -t ghcr.io/enactic/openarm_lerobot_policy_server:latest .

The policy and its weights are downloaded at build time, so the container runs offline. Use --build-arg POLICY=${POLICY} to bake another policy into the image:

docker build \
  --build-arg POLICY=${POLICY} \
  -t ghcr.io/enactic/openarm_lerobot_policy_server:latest \
  .

PyTorch for CPU is installed by default. Use --build-arg TORCH_BACKEND=${BACKEND} to install PyTorch for another backend such as cu128 (CUDA 12.8). Make sure to match the CUDA version with your host machine:

docker build \
  --build-arg TORCH_BACKEND=cu128 \
  -t ghcr.io/enactic/openarm_lerobot_policy_server:latest \
  .

Then run dataflow-docker-inference.yaml or dataflow-docker-evaluation.yaml in dora-openarm-evaluation.

Tests

uv run pytest

License

Licensed under the Apache License 2.0. See LICENSE for details.

Copyright 2026 Enactic, Inc.

Code of Conduct

All participation in the OpenArm project is governed by our Code of Conduct.