Modelpublic

Shiki42/piperx-sortletter-0905-53ep-bi-sequential-diffusion-20260910

sourceHugging Faceupdated 16d agoView on Hugging Face

0likes24downloads

Model Card

PiperX Sort Letter — DIFFUSION / bi-sequential

Native Hugging Face LeRobot policy trained from scratch on all episodes of Shiki42/piperx-sortletter-0905-53ep-bi-sequential. Dataset revision: e75e16917f36f483870efaee23715afc7466b431.

Training: 100000 optimizer steps, batch 16, seed 1000, native optimizer and model defaults, ImageNet-initialized ResNet18. Source commit: e40b58a8dfa9e7b86918c374791599d070518d11. The later source commit only initializes missing RGB-stat dictionaries before applying the same ImageNet statistics; it does not change the algorithm or loss.

Required preprocessing

Use all three RGB views: observation.images.left_wrist, observation.images.right_wrist, observation.images.top, and 14-dimensional observation.state. Before the saved preprocessor, resize EACH full RGB frame to height 240, width 320, using torchvision Resize with antialias=True, then convert uint8 to float32 in [0,1]. Do not crop. The dataset image transform in trainconfig.json is NOT automatically applied by selectaction. Apply the same preprocessing during deployment.

Load the native diffusion policy with from_pretrained and its saved pre/postprocessors. The postprocessor restores actions to dataset units. Joint order and normalization are preserved in the configuration and processor files. Configure the hardware action mapping from the dataset before deployment.

Verification and limitations

Final checkpoint reloaded in a fresh process; five training-data samples had finite loss and produced finite 14-dimensional actions. This is not a held-out evaluation or a real-robot success-rate measurement. ACT loss and diffusion noise-prediction loss are different objectives and cannot be compared numerically. All episodes were used for training.

This repository contains final inference weights, resolved configuration and normalization processors. Optimizer state, private credentials, local logs, and datasets are excluded. SHA256SUMS records the exact published files.