Models · WFM (world foundation model, Cosmos3-Edge post-trained) · flagship · preview

agi-wfm/so101

checking availability…

Our checkpoint cagataydev/agi-wfm-so101 (weights private for now), for the SO-101 arm. It answers one observation with the next 32 actions. It is not a general-purpose controller; the task sentence is passed to the model, but only the task it was trained on is expected to work. The older id wfm/so101 still works and leases the same row.

flagship · preview.post-training in progress. The numbers below are held-out offline scores, not a success rate on a real arm — what the service can and cannot tell you.

Price

per call: $0.0032 · 32 tokens
per action: $0.0001
billed: every returned action, whether or not your robot plays it

What you send

observation.state [6]: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper · normalized: arm -100..100, gripper 0..100
observation.images.front [480, 640, 3]: front camera, RGB uint8 (JPEG on the wire)
observation.images.wrist [480, 640, 3]: wrist camera, RGB uint8 (JPEG on the wire)
task text: passed to the policy

Read the camera names out of your lease’s features (or the wire’s hello) rather than this table: a slot answers with its own.

What comes back

32 actions × 6 joints, to be played at 15 Hz — 2.13 s of motion per call.

Units before motion.Values are LeRobot normalized (arm −100…100, gripper 0…100), not degrees — whatever the semantics label says. That label echoes whichever runner answered your lease, so both values are in circulation: never branch on it (units, explained). LeRobot’s so101_follower defaults to degrees: run it with --robot.use_degrees=false. What the arm does if the connection drops.

Held out

ours: 14.3 chunk / 3.51 first step

RMSE over 2.13 s of motion (32 actions at 15 Hz) / first step (normalised units) — every gate below is measured over the same 2.13 s · lerobot/svla_so101_pickplace (apache-2.0), episodes 45–49 held out · Cosmos3-Edge post-trained, 1,500 iterations, H200, 62.6 min; longer runs in progress. Lower is closer to the recorded human demonstration; it is not a task success rate.

The chunk here is its own 32 steps — 2.13 s of motion at 15 Hz. Do not read it against another model’s figure unless both the step count and the seconds of motion match: drift grows with time, so a chunk covering less of it is an easier average.

The first action

this checkpoint: 3.51
our ACT checkpoint: 4.82

One action rather than an average over a chunk, so the two are measured on the same thing. This is the part of the flagship that is already as good as our ACT checkpoint; the gap over a whole chunk is drift across it.

Does it use its cameras?

as recorded: 14.66
cameras swapped: 20.2
frames blanked: 34.88
a different episode: 55.8

The same checkpoint over the same episodes, with the images interfered with. Swapping the two cameras makes it a little worse, blanking them makes it much worse, and feeding frames from a different episode is worse still — so the policy is reading the images, not replaying a memorised trajectory.

Try it

One billed call with synthetic inputs ($0.0032). No robot moves.

Code

pip install agi-diy
agi login
agi run agi-wfm/so101 --probe -n 1

Needs Python 3.12 — LeRobot is installed on first use, requires 3.12 or later and does not run on 3.14 yet.

Or run a test call from the dashboard, or drive the simulated SO-101.

Licences and data

Checkpoint	cagataydev/agi-wfm-so101 · weights private	OpenMDW-1.1 (base nvidia/Cosmos3-Edge)
Base model	nvidia/Cosmos3-Edge	OpenMDW-1.1
Training data	lerobot/svla_so101_pickplace (apache-2.0), episodes 45–49 held out	as declared on the Hub

Licences as declared on the Hub. A lease says which of these answered it — pool and checkpoint on the session. Policies, explained.