Downloads · 30 days
10
26% of all-time downloads
Travor278/pi05-piperx-rand-retime-idle-mask-20k
pi05-piperx-rand-retime-idle-mask-20k is a robotics model from Travor278. 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 full fine-tune from PI0.5 base using Shiki42/piperx-sortletter-0905-53ep-rand-retime, revision 99871798005d2a6f62e984352a8841c268353d79. Eight H100 GPUs, global batch16, horizon50, FP32. Left/right arm and gripp…
Downloads · 30 days
10
26% of all-time downloads
All-time downloads
39
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 full fine-tune from PI0.5 base using Shiki42/piperx-sortletter-0905-53ep-rand-retime, revision 99871798005d2a6f62e984352a8841c268353d79. Eight H100 GPUs, global batch16, horizon50, FP32. Left/right arm and gripper losses are independently masked at each action target frame. Model padding and copied episode-tail targets are not supervised; normalization is by global valid elements. All-idle optimizer updates are skipped.
The seven original checkpoint files are preserved byte-for-byte. SOURCE_MANIFEST.json records SHA256 for every payload file except itself. Hub revision and manifest hash are recorded in the external delivery receipt. Training optimizer state remains on the platform; this repository is the final inference package.
Use the bundled lerobot-source.tar.gz (source revision is in provenance.json), matching dependencies from its pyproject/configuration and the training records. Extract into a separate source directory; put its src directory on PYTHONPATH. The deployment used LeRobot0.6.0, transformers5.3.0 and accelerate1.13.0. Do not silently substitute an unrelated upstream LeRobot version.
Call load_model.load_model('/absolute/path/to/this/repository', device='cuda') to get policy, preprocessor and postprocessor. The helper overrides the saved platform-specific tokenizer path with the bundled tokenizer and resets pretrained_path to this local folder. Original saved configs deliberately retain their training provenance; do not rely on their /inspire paths existing on the destination. Feed the original observation image keys and state into the saved preprocessor; its rename map handles base/left/right wrist names. Use a text task such as sort letters. This package does not connect to or control robot hardware.
Inference does not require idle labels. For training loss, use the supplied source and config mask_idle_actions=true with aligned dataset labels. Model and preprocessing assets must be used together.