pour_pi0_gt

pour_pi0_gt

PyTorch pi0 policy fine-tuned on the local UR5 real-robot LeRobot-format pour short-prompt dataset with the GT stage-2 foreground cross-view distillation setup.

• dataset: ur5_place_and_pour_nuts_camera_shifts_shortprompt
• training config name: pi0_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage2_hard
• experiment name: pi0_ur5_pour_shortprompt_cross_attn_fg_zeroinit_gt_hard_stage2_tandon_2gpu_b16
• source Slurm job: 9671920
• base camera view: observation.images.context_left_rgb
• wrist camera view: observation.images.wrist_right_rgb
• base model init: /scratch/yz11445/pi0_base
• stage-1 init weights: /scratch/yz11445/tmp/openpi_cam/checkpoints/pi0_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage1_hard/pi0_ur5_pour_shortprompt_cross_attn_fg_zeroinit_gt_hard_stage1_tandon_2gpu_b16/5000
• currently included checkpoint steps: 20000, 25000, 30000

This run has completed through 30000 steps.

Included Files

• config.json: base Pi0 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/<step>/: hard-linked checkpoint snapshots with model.safetensors, metadata.pt, and copied assets

Inference

Serve the included checkpoint with:

uv run scripts/serve_policy.py policy:checkpoint \
  --policy.config=pi0_ur5_pour_pytorch_cross_attn_fg_distill_gt_stage2_hard \
  --policy.dir=/path/to/pour_pi0_gt/checkpoints/30000

Notes

• The policy loader uses the code-defined training config pi0_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.
• The checkpoint assets use ur5_place_and_pour_nuts_camera_shifts as the asset id even though the training dataset path is the shortprompt variant.
• Release checkpoints intentionally exclude optimizer.pt.
• Normalization stats are also present inside the checkpoint asset tree; a top-level hard-linked copy is included for convenience.