Elvinky/pi05-so101-stationery-v12-human1to1-fp32-step8553

PI05 SO101 stationery v12 — human 1:1 fine-tune, final step 8553

Inference-only LeRobot PyTorch release of the completed 8 x A800 run. The starting JAX v12 policy was converted to PyTorch before this fine-tune. Training completed 2026-09-10 19:25 Asia/Shanghai. Model weights are FP32.

Training provenance

LeRobot 0.6.1, PyTorch 2.11.0+cu128; 8553 optimizer steps, seed 1000. 8 GPUs x batch 8 = global batch 64, with 4 original demonstrations and 4 human-intervention samples per rank per batch. The human view contains 741 contiguous intervention=true segments, 109466 frames at 30 FPS (1:00:48.9). The original view contains 397 episodes and 1256576 frames. The target was 2.5 human-data passes, not 2.5 passes through both sources. The original v12 normalization was retained for both sources. Temporal/action-dimension padding was excluded from training loss. Full-parameter FP32 DDP; TF32 disabled, gradient checkpointing enabled, AdamW foreach=false, no AMP/ZeRO/FSDP/LoRA. LR 2.5e-6 to 2.5e-7, warmup 200, weight decay 0.01, grad clip 1.0, chunk/action horizon 50. Task: put stationery table into bag.

Inference and portability

Model/config/processor files and tokenizer assets are at the repo root. Use both saved LeRobot processor pipelines and the saved camera/action mappings. Only the training-only mask_action_padding_loss config field was removed for stock LeRobot PI05Config compatibility. Tokenizer location now references this repo. These deployment adaptations do not alter weights, normalization, or the inference algorithm. Original small configs and audit are in provenance/. For offline inference, download the whole release and override the tokenizer processor's tokenizer_name with the local downloaded directory. The saved config pretrained_path is historical provenance, not a required external base-model dependency when loading these full weights.

release_manifest.json contains file sizes and SHA256 values. No optimizer, RNG training state, raw training videos, or credentials are uploaded. This release has not yet undergone a physical-robot evaluation; training loss does not establish deployment performance or safety. Validate in a controlled setup.

Mirror: https://hf-mirror.com/Elvinky/pi05-so101-stationery-v12-human1to1-fp32-step8553