Elvinky/pi05-piperx-4h-seed1000-fp32

PI05 PiperX copper-screw insertion — 4h, seed 1000

Final inference-only checkpoint at step 19275. Seven model/config/processor files are at the repository root for LeRobot inference. No optimizer state or raw dataset.

Dataset

190 complete episodes, 431749 frames at 30 FPS, 3.99767593 recording hours. Source: MINT-SJTU/RW-RL-Dataset, subtree piperx_insert_copper_screw, revision ddfd1144bf073b68b180bc5263ddac49c46f68d7. Sort original episode indices by SHA256 of UTF-8 piperx-scaling-v1|1000|INDEX, then take the longest whole-episode prefix not exceeding 432000 frames. Action/state normalization uses only this subset. The episode list is in train_config.json. A verified standalone data-reconstruction archive is included (see below).

Training

19275 optimizer steps, global batch 56 (7 GPUs x 8), target 2.5 epochs, warmup 643. Independent initialization from lerobot/pi05_base, not the 2h checkpoint. LeRobot 0.6.1; Torch 2.11.0+cu128; full FP32 weights; CUDA matmul TF32 permitted. Gradient checkpointing on; AdamW foreach off; no AMP, ZeRO or FSDP. Learning rate 2.5e-5 to 2.5e-6, weight decay 0.01, gradient clipping 1.0. Completed 2026-09-06 15:39 Asia/Shanghai.

Inference

Use both saved LeRobot processor pipelines, the saved camera rename mapping, state/action ordering, and this checkpoint's normalization. Research artifact: physical-robot success and safety are not established by training loss. Validate in a controlled environment before deployment.

HF mirror: https://hf-mirror.com/Elvinky/pi05-piperx-4h-seed1000-fp32

Exact dataset reconstruction package

Download rwrl_piperx_4h_seed1000_v1.tar.gz and SHA-256 checksum. Sampling/training seed: 1000. The readable manifest_4h.json and episodes.json match the actual training configuration exactly: 190 complete episodes, 431749 frames at 30 FPS. The archive includes the sampling algorithm, exact ordered episode IDs, pinned source revision and required video/Parquet SHA-256 hashes, frozen subset normalization, training reference arguments/environment, Chinese instructions, and reproduce.py. It contains metadata and scripts only, not raw videos, model weights, or optimizer state.

Extract the archive and run python reproduce.py check (requires pyarrow). With the original dataset, use python reproduce.py build --source /path/to/RW-RL-Dataset/piperx_insert_copper_screw --output /path/to/piperx_4h_seed1000. Otherwise use python reproduce.py download --dataset-root /path/to/RW-RL-Dataset to retrieve the required source files through https://hf-mirror.com at a pinned revision. Always pass the resulting episode list to LeRobot via --dataset.episodes. Exact data reconstruction does not guarantee bitwise-identical retrained model weights.