aaroncaozj/pi05_BlockPAP

π0.5 BlockPAP Checkpoints

All π0.5 checkpoints for the BlockPAP task, organized as <training method>/<camera setup>/<checkpoint>.

👉 Normalize Statistics Files Stored in aaroncaozj/pi05_norm_stats_collection. Evaluation should use the same norm_stats file as training.

Structure

pi05_BlockPAP/
├── co-sft/
│   ├── front/
│   │   ├── global_step_5000/
│   │   ├── global_step_15000/
│   │   └── stride2_global_step_15000/
│   └── front-back/
│       ├── global_step_10000/
│       └── global_step_15000/
└── nft-rl/
    └── front/
        └── global_step_30/

Each checkpoint folder contains:

• actor/model_state_dict/full_weights.pt — consolidated model weights (7.9GB), for inference.
• actor/dcp_checkpoint/ — PyTorch distributed checkpoint (4 shards + .metadata, 22GB), for resuming training. Only available for co-sft/front.

Checkpoints and Training Settings

Method	Camera	Checkpoint Folder	norm_stats	Training Stride	DCP
co-sft	front	/co-sft/front/global_step_5000	BlockPAP-v0_Mix/norm_stats_unaligned.json	1	✅
co-sft	front	/co-sft/front/global_step_15000	BlockPAP-v0_Mix/norm_stats_unaligned.json	1	✅
co-sft	front	/co-sft/front/stride2_global_step_15000	BlockPAP-v0_Mix/norm_stats.json	2	✅
co-sft	front-back	/co-sft/front-back/global_step_10000	see norm_stats_collection	—	—
co-sft	front-back	/co-sft/front-back/global_step_15000	see norm_stats_collection	—	—
nft-rl	front	/nft-rl/front/global_step_30	see norm_stats_collection	—	—

Download

Download a single checkpoint instead of the whole repo:

hf download aaroncaozj/pi05_BlockPAP \
  --include "co-sft/front/global_step_15000/actor/model_state_dict/*" \
  --local-dir ./pi05_BlockPAP

Notes

• front: trained using only front camera input; front-back: trained using front and back camera inputs.
• Using a stride when reading the dataset and constructing action labels with a larger timestep gap leads to more significant movement per action and faster task completion.
• Using different norm_stats.json results in different data distributions for the model to handle, so evaluation should use the same file as training.
• This repo merges the former repos aaroncaozj/pi05_co-sft_blockpap_front, aaroncaozj/pi05_co-sft_blockpap_front-back and aaroncaozj/pi05_nft-rl_blockpap_front.