Remote (WebSocket inference)

create_policy("remote", endpoint="ws://gpu-box:8765") returns a RemotePolicy - a drop-in Policy that forwards observations to a remote PolicyServer and returns the action chunk it computes. Use it when the robot host cannot run a large policy locally and a GPU box can.

from strands_robots import create_policy

policy = create_policy("remote", endpoint="ws://gpu-box:8765")
# smart string is equivalent:
policy = create_policy("ws://gpu-box:8765")

Config key	Default	Meaning
endpoint	-	Full server URL (ws://host:port); wins over host/port
host	127.0.0.1	Server host (when endpoint is omitted)
port	8765	Server port (when endpoint is omitted)
connect_timeout	10.0	Seconds to wait for the WebSocket handshake
request_timeout	60.0	Seconds to wait for each inference reply

The client mirrors the server policy's requires_images, execution_horizon, actions_per_step and supports_rtc, and forwards the Real-Time Chunking observed-delay count on every request, so a remote rollout behaves like a local one. Install with pip install 'strands-robots[inference]'.

See Remote Policy Inference for the full two-machine setup, the server CLI, and the wire protocol.