Separate policy and simulator environments

OpenPI remote serving separates policy dependencies from robot or simulator dependencies. Its LIBERO example includes Docker and independent-environment workflows. Follow one pinned revision’s entry points and dependencies rather than merging the latest versions of every project.

The documented server entry point is:

uv run scripts/serve_policy.py --env LIBERO

Install the matching environment and obtain the required model resources first. Private host addresses, credentials, and display-server permissions are not reproduced. This command was not run on a target GPU here.

Official-code supplement: shape alone cannot validate images

The reviewed examples/libero/main.py reads the main and wrist images, applies [::-1, ::-1], makes them contiguous, resizes/pads, and converts to uint8. This belongs to that example and must not be applied blindly to another image source. State combines end-effector position, quaternion converted to axis-angle, and gripper joint positions.

Policy field	Meaning in this example	Inspect
observation/image	Main camera image	Orientation, RGB, size, uint8
observation/wrist_image	Wrist camera image	Camera identity and alignment
observation/state	Configured proprioception	Order, units, rotation representation
prompt	Task instruction	Task identity and update boundary
actions	Action chunk	Horizon, dimensions, normalization, gripper

Requests and action queues

from openpi_client import websocket_client_policy

client = websocket_client_policy.WebsocketClientPolicy(
    host="localhost", port=8000
)
# observation must follow the selected policy's preprocessing contract.
actions = client.infer(observation)["actions"]
if actions.ndim != 2:
    raise ValueError("expected [horizon, action_dim]")

Construct observation according to the table and selected configuration. This fragment is not a standalone experiment. Inspect the actual horizon and dimensions rather than inheriting a shape from a historical diagnostic. The reviewed example queues actions, consumes a prefix according to replan_steps, and then replans. Reset must clear the queue and history.

Validate one layer at a time

Check model loading, one request with finite outputs, and then action units and stopping on a fixed initial state before evaluating a suite. Record warmup, inference, transport, execution, and episode wall time separately. Connectivity, one successful task, or server latency does not establish a complete LIBERO benchmark score. Use paired evaluation to compare an outer agent with direct policy execution.