Rollout on hardware.

Run a hosted SO-101 policy through LeRobot's control loop. Start with a dry run, not motor execution.

Before connecting hardware#

• Set up and calibrate your SO-101 through LeRobot. Check serial port, joint order, camera positions, resolution and local limits.
• Known metadata mismatch: the legacy so101-degrees-gripper-percent-v1 label (still answered by stock-* rows; agi-* leases answer so101-normalized-v1) does not describe these checkpoints' numeric values. Arm state/actions are normalized −100…100 and gripper 0…100, not degrees. Preserve the handshake identifier, but verify your calibrated input/output transforms before motion. Do not bypass admission checks.
• ACT is a pick-and-place checkpoint, not language-conditioned. agi-smolvla/so101 is a fine-tune on one real-teleoperation orange-to-plate task (the third-party stock-smolvla/so101 is the simulation-trained one), not a demonstrated general instruction follower or guaranteed real-arm transfer.
• Keep a clear workspace and physical stop. Interactive mode delays policy playback, but connecting hardware can energize motors. Ctrl-C is not an emergency-stop system.

Inspect the configuration#

Install agi-diy in a dedicated Python 3.12 environment (LeRobot needs 3.12+; lerobot-rollout fails on 3.14 — agi rollout refuses both before leasing) and agi login using a dashboard key (quickstart). Replace the serial port and camera indices below. This command does not lease, install dependencies or move hardware.

agi rollout agi-act/so101 --dry-run -- \
  --robot.type=so101_follower --robot.use_degrees=false --robot.port=/dev/tty.usbmodemYOUR_ARM \
  --robot.cameras='{"top":{"type":"opencv","index_or_path":0,"width":640,"height":480,"fps":30},"wrist":{"type":"opencv","index_or_path":1,"width":640,"height":480,"fps":30}}' \
  --interactive=true --return_to_initial_position=false

After verifying the setup, remove --dry-run to start the hardware process. /start begins policy playback. Keep --return_to_initial_position=false to avoid a commanded return-to-start motion on exit.

The first real rollout installs the pinned LeRobot client and robot extras into the interpreter agi runs in. This can change dependency versions; use an isolated environment and allow time/disk space. AGI_NO_INSTALL=1 prints the install command instead.

What the lease fills in#

agi rollout <model> leases one slot, then runs python -m agi._rollout_main (= lerobot-rollout, same interpreter as agi, never a PATH lookup) with these flags in front of yours:

Everything after -- is yours and goes to lerobot-rollout unchanged. If you pass a flag the lease also sets, yours wins — except --fps, which must equal the lease's or admission fails, so agi rollout says so before leasing. The lease ends when the process exits, however it exits.

What you give:

Cameras#

The deployment names its cameras. agi-act/so101 expects top and wrist; agi-smolvla/so101 expects front and wrist — agi models --json lists them under features, and the lease carries the same list.

Name yours the same and nothing else is needed. Otherwise agi rollout matches by name, never by order (agi-diy 0.4.9 and later):

• a camera whose name contains a model camera name as a whole word keeps that slot — wrist_cam → wrist, top_cam → top, FrontCam → front (but laptop is not top);
• one camera left over takes the one slot left over — unless its own name says it is a different view (it contains top, front, wrist, side, overhead or gripper, or a model camera name, but not the free slot's) or names two views (top_wrist): then it stops and prints the --rename_map for you to confirm;
• two or more cameras it cannot place → it stops before leasing and prints the exact --rename_map to pass, so you decide which physical camera is which:

cameras: top_cam → top, wrist_cam → wrist
error: cannot tell which of your cameras (cam0, cam1) is agi-act/so101's top / wrist. Name them like the model's (top, wrist) or pass the mapping yourself, e.g. --rename_map='{"observation.images.cam0": "observation.images.top", "observation.images.cam1": "observation.images.wrist"}' (check that each one really is that view).

What that gives for agi-act/so101 (top, wrist) — every row is run against agi/rollout.py in CI:

agi-diy 0.4.8 kept exact names but mapped the rest by order, and 0.4.7 and earlier mapped everything by order — both can swap the wrist and scene views silently (wrist_cam → front). Upgrade before connecting hardware. A mapping only renames; it cannot tell which camera points where. Look at each view before motion.

If the count differs it stops before leasing:

error: agi-act/so101 expects 2 cameras (top, wrist); --robot.cameras names 1 (front).

Resolution must match too: the deployment's shapes are 480×640; declare "width": 640, "height": 480 on every camera.

SO-101#

--robot.type=so101_follower needs LeRobot's Feetech driver. agi rollout installs lerobot[feetech] with the rest when the robot type asks for it:

Calibration is LeRobot's and unchanged: run lerobot-calibrate --robot.type=so101_follower --robot.port=… once per arm; the file is reused on every rollout. The model's joint order (shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper) is the SO-101's; still verify calibration, normalized units and joint limits.

What you will see#

+ python -m agi._rollout_main --inference.type=remote --inference.endpoint=tcp/… --inference.deployment=d_… …
Remote server deployment=d_… instance=… protocol=1
Connecting robot (so101_follower)... Robot connected
Remote session ready: deployment=d_… mode=chunk merge=aligned action_rate=30.0 Hz horizon=3.333s; contract verified
Remote rollout admitted
Base strategy control loop started
Remote progress: accepted=2 rejected=0 … turnaround mean/max=0.197/0.329s … queued playback=1.900s

Without interactive mode, policy playback starts after admission. On exit LeRobot may command a return to the initial pose unless --return_to_initial_position=false was set. Ctrl-C requests shutdown; it is not a guaranteed physical stop. With --interactive=true the arm holds still until you type /start; /subtask <text> changes the instruction, /reset returns to the starting pose, /stop ends the run. Every chunk the policy returns is billed at one token per action, visible in agi usage after metering updates.

Historical transport measurements (2026-10-24, laptop → public endpoint, two JPEG cameras). These are not physical-task success rates or latency guarantees:

Read those turnarounds against 0.75 s, not against the chunk. agi rollout asks for the next chunk when 0.75 s of motion is left (--inference.refill_seconds, set because the public wire's measured tail — smolvla 527 ms, 2026-10-25 — exceeded LeRobot's 0.5 s default), so that is the whole budget for a round trip — a 3.3 s chunk buys no more time than a 1.7 s one, it only means you pay for a call less often. Against that threshold, on the measurements above: agi-act/so101 has roughly 170 ms of margin at its worst observed turnaround, and agi-smolvla/so101 has almost none (494 ms of 500). A turnaround past it does not break the run — the arm holds its last target until the next chunk lands — but it holds on every chunk, and nothing arriving within the 3 s starvation grace faults the run.

The other way a run ends is the arm playing a chunk for too long. Every action in a chunk comes from ONE observation, and LeRobot revokes motion — a terminal fault, "Dispatched action source observation is too old" — once that observation passes --inference.max_observation_age_s. Its own default is 5.0 s, which against agi-act/so101's 3.333 s chunk leaves about a second, so agi rollout derives the value from the lease instead (chunk + grace + refill: 7.083 s for act, 5.883 s for the flagship, never less than LeRobot's 5.0 s). This matters when your control loop runs slower than the model's rate: at half rate a 100-step chunk takes 6.7 s to play, and the observation that produced it is stale long before the end. If you see that error, your loop is not keeping --fps; check that first, then widen the flag.

agi-wfm/so101 is the exception, and it is the one that can starve the queue. Its chunk is 32 actions at 15 Hz — half the rate of the others, so agi rollout injects --fps=15 and the chunk is 2.13 s of motion rather than 3.3 s. The turnaround is the problem, and it is not close: measured on the serving box on 2026-10-25, the policy alone takes 643 ms at the 4 denoise steps the service runs — past the 500 ms budget before any network — against 418 ms at 1 step, which would leave about 80 ms. Round trips measured from a client have been 1.4–2.1 s. So at the served setting this policy holds the arm briefly on every chunk; it moves with where you call from, so read the current measurement on /models/agi-wfm-so101 before you plan a loop. What that means at the arm: LeRobot's look-ahead is 0.75 s of playback under agi rollout (LeRobot's default is 0.5 s), so if a call takes longer than the remaining queue the arm holds its last target and, if nothing arrives within the 3 s starvation grace, the run faults (the same path as a dropped link: what LeRobot does) — the hold and the release of torque are the same path as a network drop, not a different one. Keep a request in flight ahead of the one that is playing, or run the flagship in --fake / simulation and keep real hardware on the 30 Hz default tier until your own measurement says otherwise.

Without an arm#

agi rollout <model> --fake -- --task="…" runs the real lerobot-rollout against the real policy with a robot that has no hardware: it holds position, reports a still state and black frames at the deployment's camera shapes, and accepts every action. Use it to check your install, your network and your bill before plugging in — the actions are real and billed.

agi rollout <model> --dry-run -- <your flags> prints the composed command and the camera mapping and leases nothing:

$ agi rollout agi-act/so101 --dry-run -- --robot.type=so101_follower --robot.use_degrees=false --robot.port=/dev/tty.usbmodem1 --robot.cameras='{"top_cam": …, "wrist_cam": …}'
cameras: top_cam → top, wrist_cam → wrist
installs: lerobot[remote,dataset,feetech] if missing
/…/bin/python -m agi._rollout_main --inference.type=remote '--inference.endpoint=<leased>' '--inference.deployment=<leased>' --inference.zenoh_mode=client --inference.encoding=jpeg --inference.semantics=so101-degrees-gripper-percent-v1 --inference.handshake_timeout_s=30 --inference.action_timeout_s=30 --inference.startup_timeout_s=30 --inference.action_starvation_grace_s=3.0 --inference.refill_seconds=0.75 --inference.max_observation_age_s=7.083 '--rename_map={"observation.images.top_cam": "observation.images.top", "observation.images.wrist_cam": "observation.images.wrist"}' --fps=30 --robot.type=so101_follower --robot.use_degrees=false --robot.port=/dev/tty.usbmodem1 '--robot.cameras={…}'

Errors#

One sentence, naming the fix. From agi rollout before anything is leased:

From LeRobot once the hardware is up (its own text, unchanged):