pour_pi05_gt

pour_pi05_gt

PyTorch pi05 policy fine-tuned on the local UR5 real-robot LeRobot-format dataset with the GT stage-2 foreground cross-view distillation setup:

• dataset: ur5_place_and_pour_nuts_camera_shifts
• training config name: pi05_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage2_hard
• experiment name: pi05_ur5_pour_cross_attn_fg_zeroinit_gt_hard_stage2_tandon_2gpu_b16
• source Slurm job: 9649620
• base camera view: observation.images.context_left_rgb
• wrist camera view: observation.images.wrist_right_rgb
• base model init: /scratch/yz11445/pi05_base
• stage-1 init weights: /scratch/yz11445/tmp/openpi_cam/checkpoints/pi05_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage1_hard/pi05_ur5_pour_cross_attn_fg_zeroinit_gt_hard_stage1_tandon_2gpu_b16/5000
• currently included checkpoint steps: 20000
• reserved checkpoint steps for later release updates: 25000, 30000

This run is still in progress. At packaging time, the stage-2 training had passed 20000 steps and had not yet produced 25000 or 30000.

Included Files

• config.json: base Pi0.5 model config copied from the initialization checkpoint
• model_architecture_config.json: fine-tuned architecture settings used by this run
• training_config_summary.json: training/data/run summary for this release
• assets/ur5_place_and_pour_nuts_camera_shifts/norm_stats.json: normalization statistics
• checkpoints/20000: symlink to the canonical stage-2 checkpoint directory
• checkpoints/25000/PENDING.txt: placeholder for the future 25000 checkpoint
• checkpoints/30000/PENDING.txt: placeholder for the future 30000 checkpoint

Inference

Serve the included checkpoint with:

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi05_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage2_hard \
  --policy.dir=/path/to/pour_pi05_gt/checkpoints/20000

After future updates, replace 20000 with 25000 or 30000.

Notes

• The policy loader uses the code-defined training config pi05_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage2_hard from src/openpi/training/config.py.
• This GT stage-2 variant uses PRoPE ray encoding, foreground cross-view fusion, and hard-confidence auxiliary point supervision from the grid-224 GT target cache.
• checkpoints/20000 is a symlink rather than a copied checkpoint payload, per this release's packaging choice.
• Normalization stats are also present inside the checkpoint asset tree; a top-level copy is included for convenience.