starvla_qwenoft_telearms_v5_random_1500

Final checkpoint: step 160,000.

StarVLA QwenOFT (Qwen2.5-VL-3B-Instruct backbone, full fine-tune including the vision encoder) 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 (front, left wrist, right wrist), no domain randomization.

One of five StarVLA action heads trained on this data with the same backbone and the same 15.36M-sample budget as the pi0.5 and MolmoAct2 Random-1500 runs.

  • Code: starVLA starVLA_dev @ 13d8ad4 plus a local examples/TelearmsYAM data config (cameras front/left/right, mixture telearms_v5_random_1500, robot type telearms_yam_sim).
  • 8 x H100, ZeRO-2, batch 12/GPU (global 96), 160,000 steps = 15.36M samples; this checkpoint has seen 15.36M.
  • Absolute 14-D joint actions (per arm 6 joints + gripper), 30-step chunks, q99 quantile normalization; 14-D proprioceptive state as input. dataset_statistics.json holds the statistics needed to un-normalize predicted actions.
  • LR 1e-5 (VLM) / 1e-4 (action head), 1,000 warmup steps, cosine to 5e-7, AdamW (0.9, 0.95).
  • Training loss (action_dit_loss) at step 160,000: 0.0214.
  • Files: pytorch_model.pt (StarVLA state dict), config.yaml / config.full.yaml (resolved training config).
  • Evaluation: not yet evaluated in simulation.
Downloads last month
-
Video Preview
loading

Model tree for helen9975/starvla_qwenoft_telearms_v5_random_1500

Finetuned
(933)
this model

Dataset used to train helen9975/starvla_qwenoft_telearms_v5_random_1500