pi05_telearms_v5_exp_elo_1500

Final checkpoint: step 30,000.

π₀.₅ (lerobot/pi05_base) full fine-tune on mliu48/telearms-yam-cube-v5-exp-elo-1500 (revision 940f41ef): 15,000 simulated bimanual YAM demonstrations of 10 tasks, 10,273,656 frames at 30 Hz, three 224×224 cameras, no domain randomization.

Selection. 1,500 demonstrations per task drawn at random from the top third of the task by the player's ELO at recording time, within all successful demonstrations of the task in the full dataset that have a demo score and a rated player ELO.

One of eight v5 experiment runs (Random-50/100/500/1000/1500, Elo-1500, Score-1500, DemInf-1500) trained with the same recipe and the same 15.36M-sample budget so they can be compared.

  • Code: helen9975/RoboEval-telearms branch PI05_training @ 9dc9e9f plus local changes (NaN guards, PyAV logging fix for data-loading speed, configurable minimum episodes per task), LeRobot 0.6.1 @ e40b58a8 with the repo's patch (tools/install_lerobot_patch.py).
  • Full fine-tune (LLM + vision + action expert), bf16, gradient checkpointing; per-task prompt variants (meta/task_prompt_variants.json); no episodes held out, so there is no validation loss.
  • 8 × H100 (1 node), batch 64/GPU (global 512), 30,000 steps = 15.36M samples; this checkpoint has seen 15.36M of them.
  • AdamW (β = 0.9, 0.95, weight decay 0.01), peak LR 1e-4 (2.5e-5 × sqrt(512/32)), 1,000 warmup steps, cosine decay to 1e-5 over 30,000 steps; 50-step action chunks; relative actions with absolute grippers.
  • Recent training loss: 0.024.
  • These files are the checkpoint's pretrained_model/.
  • Evaluation: simulation, training/evaluate_pi.py suite on the 10 trained tasks, 20 episodes per task (fixed seeds 900000001+, the same scenes for all eight experiment models), up to 5,400 steps, 25-step action chunks, success = the task succeeded at any point. 189/200 = 94.5% (95% Wilson interval 90-97%).
Task Success
Cube Patch Placement 20/20
Front Edge Cube Row 18/20
Front To Back Cube Line 20/20
L Cube Arrangement 16/20
Letter Face Alignment 17/20
Red Cube In Front Of Blue Cube 20/20
Red Left Of Blue Placement 20/20
Semantic Cube On Red Patch A Up 18/20
Sequential Push Block to Target 20/20
Three Cube Wall Row 20/20
Downloads last month
65
Safetensors
Model size
4B params
Tensor type
F32
·
BF16
·
Video Preview
loading

Dataset used to train helen9975/pi05_telearms_v5_exp_elo_1500