Dongkkka/molmoact2_dashboard_0904_10k_16bs

MolmoAct2: Cyclo dashboard, 10,000 steps, batch size 16

Intermediate 10,000-step checkpoint for pick up the bottle and place it into basket. Contains the saved policy, processors, normalization statistics, original training configuration and checkpoint verification report. The source run targets 20,000 steps and was still training when this snapshot was selected.

Training

• Batch size 16, 10,000 completed optimizer steps in this checkpoint, seed 42, BF16.
• Planned run length: 20,000 steps; train_config.json records that original target.
• Base: allenai/MolmoAct2, revision e432d85f6e039edca44afb93c262f3084ab72a9c.
• VLM LoRA: rank 64, alpha 16, dropout 0.05; full action expert training.
• All 28 episodes / 5,133 frames used for training, no held-out split.
• Three cameras: left head, left wrist, right wrist.
• State/action dimension 22; action chunk 30; continuous inference.
• First 19 action channels are absolute joint positions; last three are base velocities.
• Saved quantile normalization includes grippers; padded action dimensions are masked.

Loading

from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy
from lerobot.policies.factory import make_pre_post_processors

repo_id = "Dongkkka/molmoact2_dashboard_0904_10k_16bs"
policy = MolmoAct2Policy.from_pretrained(repo_id, strict=True).eval()
preprocessor, postprocessor = make_pre_post_processors(policy.config, pretrained_path=repo_id)

Use the Cyclo LeRobot environment used for training, including its MolmoAct2 support. The policy and processors reference the original AllenAI base/tokenizer repositories; these dependencies are fetched separately when needed. Restore the saved processors alongside the policy. Original local paths in train_config.json are provenance. Optimizer/resume states remain in the local checkpoint.

Verification

The saved checkpoint records step 10,000 and batch size 16. The exported checkpoint loaded strictly and produced finite (1, 30, 22) predictions from one training observation without ground-truth actions. The 22-to-32 action padding mask was checked. Robot task success and held-out generalization have not been evaluated.