Submitting Jobs

Use execute() to submit a job and wait for the result:

from armnet_client import execute

result = execute(
    image="registry.example/my-runtime:tag",
    embodiment="lerobot/so-101",
    task="push_green_button",
    args={"calibrate": False},
    timeout_seconds=1200,
)

Arguments

All arguments are keyword-only.

Argument	Default	Purpose
image	(required)	Image ref, usually returned by Image.build(...).push().
embodiment	(required)	Robot embodiment used for cell routing.
task	None	Task used for cell routing. Omit to let any cell of that embodiment run the job.
args	None	JSON-serializable values exposed as ctx.args.
secrets	None	Mapping of env var name to armnet secret name.
detach	False	When False, disconnecting the client requests job cancellation.
stream_logs	True	Stream container stdout/stderr to your terminal.
use_rerun	False	Replay the job's ctx.log_rerun_data packets into a local Rerun viewer.
timeout_seconds	120	Max container runtime. Raise this for evals — the default is only enough for a smoke test.
allow_queue	True	Queue the job when no cell is free instead of failing immediately.

Cell Routing

Jobs are routed to robot cells by (embodiment, task). See Environments and Tasks for the task list and what each environment guarantees about scoring, resets, and availability.

If no matching cell is online the job queues rather than failing, so a manual-environment job can sit for days before an operator picks it up. Check availability first when that matters:

from armnet_client import OrchestratorClient

print(OrchestratorClient().get_cell_status(
    embodiment="lerobot/so-101",
    task="push_green_button",
))

Manual-environment tasks always queue, even when their cell is available. An operator must first stage the physical scene and then press Dispatch in the FMS. Use detach=True for these submissions. Passing allow_queue=False rejects a manual task rather than bypassing the setup requirement. Teleoperation and recording set skip_operator_dispatch so they start immediately instead.

For LeRobot policy evaluation, prefer the higher-level armnet-lerobot-eval command. It wraps Image.build, Image.push, and execute with a LeRobot-like CLI.