Software Setup

This setup is split by machine. Currently we are using 3 computers. Keep OpenPI policy dependencies separate from DROID robot-control dependencies.

4090 Workstation: OpenPI

cd /home/kindred/Desktop/repo/openpi
uv sync
GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .

The policy server is started from this repo:

uv run scripts/serve_policy.py \
  --port=8000 \
  policy:checkpoint \
  --policy.config=pi05_droid \
  --policy.dir=gs://openpi-assets/checkpoints/pi05_droid

For a fine-tuned checkpoint, use the fine-tuning config and local checkpoint directory:

uv run scripts/serve_policy.py \
  --port=8000 \
  policy:checkpoint \
  --policy.config=pi05_droid_finetune \
  --policy.dir=/path/to/checkpoint_step

Control Laptop: DROID + OpenPI Client

Install the lightweight OpenPI client into the DROID environment:

conda activate droid
cd /home/kindred/Desktop/repo/openpi/packages/openpi-client
pip install -e .
pip install tyro

The control laptop uses the DROID checkout:

cd /home/kindred/Desktop/repo/droid

Key scripts:

Script	Purpose
scripts/collect_openpi_realsense.py	Collect OpenPI-compatible RealSense DROID demonstrations.
scripts/main.py	Run VLA policy rollouts against the 4090 policy server.
scripts/tests/check_nuc_control.py	Check laptop to NUC to Franka communication.

NUC: DROID Low-level Server

On the NUC:

ssh hcilab@192.168.4.4
cd /home/hcilab/Desktop/github/droid
bash scripts/server/launch_server.sh

The server binds tcp://0.0.0.0:4242. The control laptop connects through DROID ServerInterface.

Hugging Face CLI

Use Hugging Face as the lab dataset registry:

python -m pip install -U "huggingface_hub[cli]"
hf auth login
hf auth whoami

For this lab workflow, do not put credentials in submit files or committed scripts.