vovw/MolmoAct2-Pink-Towel-DAgger

MolmoAct2 Pink Towel DAgger

Action-expert fine-tune of MolmoAct2-BimanualYAM for the instruction “fold the towel” on a bimanual YAM robot.

Training dataset and visualization

• Training dataset: Dimios45/molmo-fold-pink-towel-dagger
• Visualize the dataset — episode 0
• LeRobot dataset viewer

Dataset revision: 4bdca5b4787c81ea90e016da349773d27434756a. The dataset contains 30 episodes and 142,422 frames at 30 FPS. All recorded actions were eligible for training, including policy rollouts, human interventions, failed attempts, and the retained aborted episode. Human interventions account for 20,916 frames (14.69%). This run uses ordinary behavior cloning of the mixed dataset; it does not mask the loss to human interventions only.

Training

• Base revision: 8dcbed66f2380e4393189c303ea72488eb9e63c2.
• 3,000 optimizer steps; effective batch size 1.
• AdamW action-expert learning rate: 5e-5; warmup: 200 steps.
• Action expert trained with FP32 master weights; vision-language backbone frozen and stored in BF16. Published inference weights are BF16.
• Action horizon: 30; one flow-matching timestep per training example.
• Single NVIDIA RTX 4090; training code from this repository's MolmoAct2 experiments package.

The final optimizer state is not included. Training loss is not a measure of autonomous robot success. No held-out evaluation or real-robot evaluation of this fine-tuned checkpoint has been performed.

Robot interface

State and absolute joint-position actions have 14 values, ordered as left joints 1–6, left gripper, right joints 1–6, right gripper. Camera keys are observation.images.top, observation.images.left_wrist, and observation.images.right_wrist. Use the normalization tag yam_pink_towel_dagger with the included norm_stats.json.

The model includes the MolmoAct2 custom Transformers implementation, tokenizer, processor, and normalization statistics. Load with trust_remote_code=True and use the MolmoAct2 continuous-action inference interface. The accompanying training/ directory records the training configuration and environment.

Dataset licensing is not specified by its uploader; see the linked dataset and base-model repositories for their respective terms and provenance.