molmoact2_busybox_push_green_button

Full-component fine-tune of MolmoAct2 for busybox_push_green_button on SO101 data.

Policy	MolmoAct2 (policy.type=molmoact2, train_mode_vlm=fft)
Init checkpoint	allenai/MolmoAct2
Dataset	villekuosmanen/busybox_push_green_button
Task	busybox_push_green_button
Action dim	6 (single-arm)
Cameras	top, wrist, front
Training	Vast.ai 4× RTX PRO 6000 S, full-component, 30k steps, global batch 4, chunk/action horizon 30, QUANTILES ACTION/STATE
W&B project	molmoact2_busybox_push_green_button
W&B run	6scxbrtc

Checkpoints

main is step 30,000. Intermediate steps are git tags:

Step	Revision
30000	step-30000 (main)
25000	step-25000
20000	step-20000
15000	step-15000
10000	step-10000
5000	step-5000

Usage

from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy
policy = MolmoAct2Policy.from_pretrained("pravsels/molmoact2_busybox_push_green_button")
# earlier step: revision="step-25000"