Quickstart

Install

relay needs Python 3.10 or newer.

pip install relay-intelligence

Set the server URL

Put WALLE_URL in your environment or in a .env file in your working directory. {model} stays as written: the client replaces it with the model you connect to.

.env

WALLE_URL=https://<workspace>--walle-{model}.modal.run

Replace <workspace> with the Modal workspace the servers are deployed in. For a development server started with modal serve, add -dev before .modal.run.

Get actions

first_call.py

import numpy as np
from relay import PolicyClient

policy = PolicyClient("pi05")

frame = np.zeros((224, 224, 3), np.uint8)   # replace with your camera frames
actions = policy.infer({
    "observation.images.base_0_rgb": frame,
    "observation.images.left_wrist_0_rgb": frame,
    "observation.images.right_wrist_0_rgb": frame,
    "state": np.zeros(32, np.float32),        # replace with your joint state
    "prompt": "pick up the cup",
})

print(actions.shape)   # (50, 32): 50 future steps × 32 action dims
policy.close()

actions[0] is the action to execute now, actions[1] the one after, and so on.

Check what the server expects

Each server describes its model on connect. Read this before you build observations for a new model.

print(policy.get_server_metadata())
# {'image_resolution': [224, 224], 'max_cameras': 3, 'action_horizon': 50,
#  'action_dim': 32, 'action_space': 'joint_position', ...}

If a server has just been deployed or scaled up, the first connection waits while the model loads. That can take a few minutes. PolicyClient waits rather than timing out, and later connections take under a second.

Run the demo

The repository includes a demo. It sends synthetic frames, shows each action chunk as it arrives and reports latency percentiles at the end. It’s the quickest way to check your setup and see the latency from your network.

git clone https://github.com/WALLservE/robot-client && cd robot-client
pip install -e .
python examples/demo.py                 # π0.5, 500 steps
python examples/demo.py pi0 --steps 50  # another model, fewer steps

Replace get_observation() in examples/demo.py with your robot’s cameras and joint state to turn it into a real control loop.

Get started

Guides

Reference

Run the demo

Next steps

Observations

Control loops

​Run the demo

​Next steps

Observations

Control loops

Run the demo

Next steps