Downloads · 30 days
24
40% of all-time downloads
ASethi04/MolmoAct2-BimanualYAM-yam-ultrawide-teleop-ee20-measured-fk-12k
MolmoAct2-BimanualYAM-yam-ultrawide-teleop-ee20-measured-fk-12k is a robotics model from ASethi04. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot.
This policy was fine-tuned for 12,000 optimizer steps on 75 measured-YAM teleoperation episodes. The exact task is pick up oranges and place them in the bowl. Inputs are the current left and right wrist images plus th…
Downloads · 30 days
24
40% of all-time downloads
All-time downloads
60
Public
Parameters
5.6B
11.5 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors11.5 GB · 100%
How the weights are stored.
BF165.4B · 97%
From the Hugging Face model README
This policy was fine-tuned for 12,000 optimizer steps on 75 measured-YAM
teleoperation episodes. The exact task is pick up oranges and place them in the bowl. Inputs are the current left and right wrist images plus the current
identity-pose/current-gripper EE20 state.
The output shape is (24, 20). Each arm contributes xyz (3) + first two rotation-matrix rows (6) + absolute normalized gripper (1), left then right.
Every future row is independently query anchored:
DeltaT[k] = inverse(T[t]) @ T[t+k], for k = 1..24.
Training labels use measured YAM joints and authoritative I2RT FK at both the
query and future rows. At deployment, PR #15 reconstructs
T_target = T_query @ DeltaT, runs I2RT IK independently per arm, and fails
closed on the configured residual, joint-limit, and frozen-query seed gates.
Grippers are model-space [0, 1] (0 closed, 1 open); the robot adapter
applies the 85 mm / 96 mm driver mapping. No legacy tool-axis transform is
used. Execute exactly 24 actions per inference query.
brandonyang/yam-ultrawide-teleop@e0fae691e2eee74430dd463adf8e17180bc735e9.allenai/MolmoAct2-BimanualYAM@8dcbed66f2380e4393189c303ea72488eb9e63c2.allenai/MolmoAct2-FAST-Tokenizer@d45593b4c863d0bc1ca064f8b352fa16b75c38e8.7ed46f4e4e316133a0c39aa6cf34a73d2718e850, YAM + LINEAR_4310, authoritative grasp_site.4e439e9e62c9a65a58d22bafe5efcd4cba12f1cf.39af035ae1b9007c9c6d9c290d21fae8cb58b802.The checkpoint passed a complete finite-tensor scan and a clean-cache inference
smoke through the deployment code, producing a finite (1, 24, 20) action
chunk. On a deterministic 75-query training-observation replay, mean per-arm
errors were 14.10 mm translation and 0.02744 rad rotation, with zero predicted
gripper-domain violations. This replay is not held-out evaluation.
No held-out split was created. Offline replay is not evidence of generalization, closed-loop success, or hardware safety. A passing IK row does not replace collision checking, calibration, workspace limits, an emergency stop, or a low-speed no-motion/action probe. Use PR #15 at the deployment commit above, the exact prompt above, H24 inference, and an immutable Hub revision.