pi05_telearms_v5_full

Intermediate checkpoint: step 122,500 of 195,313. Training is still running.

pi0.5 (lerobot/pi05_base) full fine-tune on mliu48/telearms-yam-cube-v5-realistic (revision 052e5550): 864,347 simulated bimanual YAM demonstrations over 104 task groups after the upstream motion filter, 1.05B frames at 30 Hz, three cameras rendered like the real RealSense cameras (16:9) and letterboxed to 224x224.

  • Code: helen9975/RoboEval-telearms PI05_training @ 9dc9e9f (tools/training_diagnostics/train_v5.sh settings, 200M samples) plus local changes for a billion-frame dataset (compact frame index, vectorised load check, linear motion filter, float32 timestamp tolerance 4e-3) and the NaN guards / PyAV logging fix; LeRobot 0.6.1 @ e40b58a8 with the repo's patch.
  • Full fine-tune (LLM + vision + action expert), bf16, gradient checkpointing; per-task prompt variants; no episodes held out.
  • 16 x H100 (2 nodes x 8), batch 64/GPU (global 1,024), 195,313 steps = 200M samples; this checkpoint has seen 125M.
  • AdamW (0.9, 0.95, weight decay 0.01), peak LR 1.414e-4, 1,000 warmup steps, cosine decay; 50-step action chunks; relative actions with absolute grippers.
  • These files are the checkpoint's pretrained_model/.
  • Evaluation: not yet evaluated in simulation.
Downloads last month
11
Safetensors
Model size
4B params
Tensor type
F32
·
BF16
·
Video Preview
loading

Dataset used to train helen9975/pi05_telearms_v5_full