Downloads · 30 days
18
36% of all-time downloads
Elvinky/pi05-so101-stationery-v12-human1to1-fp32-step8553
pi05-so101-stationery-v12-human1to1-fp32-step8553 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.
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.
Downloads · 30 days
18
36% of all-time downloads
All-time downloads
50
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
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.
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.
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