Downloads · 30 days
12
27% of all-time downloads
Elvinky/pi05-piperx-16h-seed1000-fp32
pi05-piperx-16h-seed1000-fp32 is a robotics model from Elvinky. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot.
Final 67,500-step checkpoint, trained from official lerobot/pi05base on 692 complete episodes / 1,728,000 frames at 30 FPS = exactly 16 hours. The data consists of all original 558 demonstrations plus 134 seed-selecte…
Downloads · 30 days
12
27% of all-time downloads
All-time downloads
45
Public
Parameters
4.1B
16.6 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors16.6 GB · 100%
From the Hugging Face model README
Final 67,500-step checkpoint, trained from official lerobot/pi05_base on
692 complete episodes / 1,728,000 frames at 30 FPS = exactly 16 hours.
The data consists of all original 558 demonstrations plus 134 seed-selected
4090A manual-teleoperation episodes. No 4090D data, no intervention replay,
no value-selected fragments, and no episode truncation were used in this run.
8 RTX PRO 6000 GPUs, batch 8 per GPU (global 64), 2.5 epochs, seed1000.
LeRobot 0.6.1, PyTorch 2.11.0+cu128, torchvision 0.26.0+cu128,
Transformers 5.5.4, CUDA runtime12.8, cuDNN9.19.
Official source commit: 7e241bd630a3719a56157a497ce5d08f244784f1.
Full-parameter FP32 weights, AMP off, gradient checkpointing on, compile off.
FP32 weights do not imply TF32 matrix multiplication was disabled.
AdamW lr2.5e-5 to2.5e-6, warmup2250, cosine67500, betas(0.9,0.95),
eps1e-8, weight_decay0.01, gradient clip1.0.
Same official training/model/loss/optimizer source as the 7h experiment;
no custom padding-loss mask or 1:1 sampling.
PyAV timestamp tolerance is1ms because float32 timestamps exceed100us at some frames.
Model weights, config, both processor pipelines and their normalization tensors,
and tokenizer files are at the repository root. No optimizer or RNG state is uploaded.
Use both saved processor pipelines; camera input names are base_0_rgb,
left_wrist_0_rgb, and right_wrist_0_rgb under observation.images.
Task: Insert the copper screw into the black sleeve.
The exported tokenizer processor points to this HF repository, not a server-local path.
For an offline local snapshot, override tokenizer_processor.tokenizer_name with
the local snapshot directory. Weights and normalization tensors are unchanged.
Checkpoint was strictly reloaded and tested using a real recorded observation: finite action output shape1x50x14. This is a loading/inference sanity test, not a robot success-rate evaluation.
See reproducibility/:
selection_16h_seed1000.json: exact ordered source-episode selection and pinned revisions.a_episode_rank_seed1000.json: complete seeded A ranking and lengths.source_episode_map.json: source-to-training episode mapping.normalization_stats.json: exact fresh statistics used for the full16h dataset.dataset_metadata/: actual training metadata (including canonical camera names).training_command.json, environment.json: actual parameters, code fingerprints and environment.verify_selection.py: independently verifies deterministic selection, counts, mapping and stats.historical_build_v3_view.py: archived construction implementation for audit; its paths
are historical and must be adapted on another machine. It is not a portable one-command downloader.Run python reproducibility/verify_selection.py after downloading the records.
The selection manifest's historical status string predates dataset construction;
completed training and validation are recorded separately in final_validation.json.
Sources:
MINT-SJTU/RW-RL-Dataset, revision ddfd1144bf073b68b180bc5263ddac49c46f68d7,
subdirectory piperx_insert_copper_screw (all558 episodes,1,255,729 frames).Elvinky/piperx-copper-screw-manual-teleop-20260904-0906, revision
d97b1f535def5f5055209ba0db57285371a150d5 (selected A episodes,472,271 frames).Selection uses SHA256 seeded ranking followed by exact whole-episode subset sum; it is not the D1/D3/D7 prefix-selection algorithm. See the frozen manifest. Use the supplied full16h stats, not the full558-only normalization. This repository records the experiment and exact data selection; it does not include source videos, or promise bitwise-identical retraining across hardware.