Policy Inference Quickstart

Policy inference runs across two machines on the same network, exactly like data collection:

Main host (the upper computer) — runs the policy server and the autonomous control loop.
ZED box (the Jetson sender) — streams the three ZED-X One cameras.

Both machines must agree on the time (via PTP) so the camera observations fed to the policy are aligned with the joint state. Each machine runs two long-lived processes, each in its own terminal, that stay up for the entire rollout session.

Start the clock-sync daemon on both machines before launching run-policy. If PTP isn’t running, the receiver’s capture/decode skew check in ZedCamera.connect() warns and the observations sent to the policy will not be time-aligned with joint state. See zed.sync-clocks.

Network setup

The two machines must be able to reach each other over the network. Pick the topology that matches your wiring — it determines whether you let axol assign static IPs or you point it at an existing address. This is the same setup as the Data Collection quickstart.

Direct ethernet
Shared router / switch

A single cable between the two NICs, no router in between. Neither machine gets a DHCP address on that link, so axol assigns static IPs for you: the ZED box takes 192.168.10.1 (via --setup-ip) and the main host takes 192.168.10.2 (via --zed_iface). The receiver’s default zed_host is already 192.168.10.1, so no IP flag is needed on run-policy.

Both machines plug into the same LAN and get their addresses from DHCP (or static config you manage). Do not pass --setup-ip / --zed_iface here — they would overwrite the address the network already assigned. Instead, find the ZED box’s IP (e.g. ip addr) and point the receiver at it with --robot_config.zed_host <zed-box-ip>.

Interface names vary by machine — the link might be eth0, enp3s0, eno1, and so on. Run ip link (or ip addr) on each machine to find the interface carrying the link, and substitute it for <iface> in the commands below. In both cases PTP runs over whichever interface carries the link (zed.sync-clocks --iface). If the two machines are not on the same L2 segment (e.g. separated by a router rather than a switch), add --transport udpv4 to the clock-sync commands.

Prerequisites

The lerobot extra installed on the main host. For GPU inference also add the cuda extra: uv sync --extra lerobot --extra cuda (see Installation).
A trained checkpoint (local path or HuggingFace repo) and its policy type (act, smolvla, pi0, …).
CAN set up on the main host (can.setup) and motors verified (motor.info).
pyzed installed on the ZED box (zed.install) and the three camera serial numbers on hand.
A network link between the two machines — either a direct ethernet cable or a shared router/switch (see Network setup).

ZED box (Jetson)

Open two terminals on the ZED box.

Start clock sync (slave)

axol zed.sync-clocks --role slave --iface <iface>

Replace <iface> with the ZED box’s link interface (find it with ip link). Leave this running for the whole session.

Stream the cameras

Direct ethernet
Shared router / switch

axol zed.stream \
    --overhead 12345678 \
    --left-arm 23456789 \
    --right-arm 34567890 \
    --setup-ip <iface>

--setup-ip <iface> assigns the sender IP 192.168.10.1/24 to that interface before streaming.

axol zed.stream \
    --overhead 12345678 \
    --left-arm 23456789 \
    --right-arm 34567890

Drop --setup-ip so the box keeps its network-assigned address. Note that address (e.g. via ip addr) — you’ll pass it to the main host as --robot_config.zed_host.

Use the same --resolution and --fps the policy was trained on so the ZedCamera validation on the main host passes. See zed.stream.

Main host (upper computer)

Open two terminals on the main host.

Start clock sync (master)

axol zed.sync-clocks --role master --iface <iface>

Replace <iface> with the main host’s link interface (find it with ip link). Leave it running for the whole session.

Run the policy

Direct ethernet
Shared router / switch

axol run-policy \
    --policy_path myorg/pick-place-policy \
    --policy_type act \
    --task "Pick the red cube" \
    --zed_iface <iface>

--zed_iface <iface> assigns the receiver IP 192.168.10.2/24 to that interface; the cameras are reached at the default zed_host 192.168.10.1.

axol run-policy \
    --policy_path myorg/pick-place-policy \
    --policy_type act \
    --task "Pick the red cube" \
    --robot_config.zed_host 192.168.1.42

Replace 192.168.1.42 with the ZED box’s address. Omit --zed_iface so the host keeps its network-assigned address.

A PolicyServer child process is launched on localhost; the parent streams observations to it and applies the returned action chunks. --fps must match the rate the policy was trained on. For CPU inference add --device cpu. See run-policy for aggregation, chunking, and dataset-saving fields, and Command configuration for the draccus override syntax.

Controlling a rollout

While run-policy runs, control the episode from stdin in the run-policy terminal:

Key	Action
`s`	Save the rollout and end the episode
`r`	Discard and re-record
`q`	Discard and quit

--episode_time_s is a safety cap (default 120 s) that falls back to the same [Enter]=save / r / q prompt if no key is pressed. If --repo_id is supplied, each saved episode is appended to a LeRobot-format dataset. Between episodes the arms return to the rest pose via a collision-aware IK trajectory.

Policy Inference Quickstart

Network setup

Prerequisites

ZED box (Jetson)

Main host (upper computer)

Controlling a rollout

Next steps

Data Collection

run-policy reference

Documentation Index

​Network setup

​Prerequisites

​ZED box (Jetson)

​Main host (upper computer)

​Controlling a rollout

​Next steps

Data Collection

run-policy reference

Network setup

Prerequisites

ZED box (Jetson)

Main host (upper computer)

Controlling a rollout

Next steps