pi05_telearms_v4dr (partial: step 31,250 of 44,455)

This is a partial checkpoint (70% of the planned epoch). The previous upload here (step 44,455) had NaN weights: the run resumed from a corrupted checkpoint at step 33,000 after a preemption and trained on NaN from then on. Step 31,250 is the last clean checkpoint. Training is being resumed from it (with NaN guards); the final checkpoint will replace this one.

π₀.₅ (lerobot/pi05_base) full fine-tune on mliu48/telearms-yam-cube-v4-dr: 99,496 simulated bimanual YAM demonstrations (61,027 originals over 48 tasks + 38,469 colour/letter variants), 30 Hz, three cameras rendered like the real RealSense cameras (16:9, real fields of view, randomized lighting/backgrounds) and letterboxed to 224×224.

• Code: helen9975/RoboEval-telearms branch PI05_training @ d6dc57a (training/train_pi.py), 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 (original instruction + 50 rewrites, meta/task_prompt_variants.json).
• 8 × H100 (1 node), batch 128/GPU (global 1,024), planned 44,455 steps = one selected-data epoch (99,309 episodes; 10% of each task held out for validation). This checkpoint: step 31,250.
• Peak LR 1.414e-4 (2.5e-5 × sqrt(1024/32)), cosine decay to 1.414e-5, 1,000 warmup steps; relative actions with absolute grippers.
• Evaluation: training/evaluate_pi.py renders the simulator cameras 16:9 and letterboxes them, as this dataset was rendered (read from the dataset's meta/info.json).