Fine-tuned VLA-JEPA on the MolmoAct2-BimanualYAM-Dataset with 3 cameras packed top → left → right, unfrozen VLM, and world-model loss on (blocks_v2 recipe).

Checkpoint details

Setting	Value
Training step	100000
Cameras	top → left → right
Action dim	14 (absolute joints + grippers)
Chunk size	30 (@ 30 Hz)
VLM	unfrozen
World model	on
Base model	lerobot/VLA-JEPA-Pretrain
Gripper range	[0, 1] (continuous)

Usage (LeRobot)

python

from lerobot.policies.vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
from lerobot.policies.factory import make_pre_post_processors

ckpt = "Jiafei1224/molmoact2-yam-vla-jepa"
policy = VLAJEPAPolicy.from_pretrained(ckpt)
preprocessor, postprocessor = make_pre_post_processors(policy.config, pretrained_path=ckpt)

Open-loop smoke (2 held-out episodes, stride 30, reset-each-step)

MSE: 0.00672
MAE: 0.0596