molmoact2_telearms_v5_random_1500

Final checkpoint: step 120,000.

MolmoAct2 full fine-tune of allenai/MolmoAct2-BimanualYAM on mliu48/telearms-yam-cube-v5-exp-random-1500: 15,000 simulated bimanual YAM demonstrations (10 tasks x 1,500), 14,340,545 frames at 30 Hz, three 224x224 cameras, no domain randomization. Same data and 15.36M-sample budget as the pi0.5 and StarVLA Random-1500 runs.

  • Trainer: allenai/molmoact2 experiments/launch_scripts/train_lerobot.py (upstream 6070080), converted with olmo.hf_model.convert_molmoact2_to_hf.
  • Full fine-tune (LLM, ViT, connector, LM head, action expert), unpacked + dynamic sequence length.
  • 8 x H100, batch 16/GPU (global 128), 120,000 steps = 15.36M samples; LRs: LLM 1.414e-5, ViT 7.07e-6, connector 7.07e-6, action expert 7.07e-5 (upstream full-FT recipe x sqrt(128/64)).
  • Mixture tag telearms_yam_sim_v5: absolute joint-pose control, 14-D state/action (per arm 6 joints + gripper), cameras [camera_front, camera_left, camera_right] (BimanualYAM's [top, left, right] order), 30-step action chunks. norm_stats.json holds this dataset's absolute action/state statistics.
  • Evaluation: not yet evaluated in simulation.
Downloads last month
8
Safetensors
Model size
5B params
Tensor type
F32
·
Video Preview
loading

Model tree for helen9975/molmoact2_telearms_v5_random_1500

Finetuned
(47)
this model

Dataset used to train helen9975/molmoact2_telearms_v5_random_1500