UMI Client for OneTwoVLA

This repository provides the hardware client code for running the OneTwoVLA model on the Universal Manipulation Interface (UMI) platform. The training code for OneTwoVLA can be here, and the instructions for runnning policy server is available here.

---

Hardware Setup

Please follow the UMI hardware setup instructions.

Note:

• We remove the mirror and change the gripper design. See details here.
• TCP offset is updated in code (reference commit).

---

Run the UMI Client

The client script is umi_client.py.

To launch the client:

Activate your umi Python environment and start the client:

python umi_client.py \
    --robot_config=example/eval_robots_config.yaml \
    -o /path/to/output/dir/ \
    --frequency 5 -j \
    --temporal_agg -si 1 \
    -ins "<your instruction here>" \
    --state_horizon 3,15 \
    -action_down_sample_steps 3 \
    -getitem_type necessary \
    --remote_port 8000

Replace <your instruction here> with your task instruction. For example, for open-world visual grounding: -ins "Give me the object reminding me of the sea."