yangzhixing/oat_rfsq_pair_so101_tokenizer

OAT paired-RFSQ action tokenizer for SO-101

This repository contains the complete action-tokenizer checkpoint used to initialize yangzhixing/oat_rfsq_pair_so101_policy_tok50k.

• Tokenizer checkpoint: checkpoint-00050000 (50,000 steps)
• Dataset: maxlium/so101-box-to-plate
• Policy type: oat_rfsq_pair
• Action horizon: 32
• Latent horizon: 8
• FSQ levels: [8, 5, 5, 5]

Files

• action_tokenizer.safetensors: tokenizer weights
• tokenizer_config.json: architecture and training configuration

Use with the matching LeRobot source

The OAT integration is available in the following source revision:

git clone --branch feat/oat-rfsq-unified-integration https://github.com/Yangzhixing123/lerobot.git
cd lerobot
git checkout 82b706cc

Download the tokenizer and pass its local directory when training an OAT policy:

hf download yangzhixing/oat_rfsq_pair_so101_tokenizer \
  --local-dir ./oat_rfsq_pair_so101_tokenizer

lerobot-train \
  --dataset.repo_id=maxlium/so101-box-to-plate \
  --policy.type=oat_rfsq_pair \
  --policy.action_tokenizer_path=./oat_rfsq_pair_so101_tokenizer

The released policy checkpoint already embeds the frozen tokenizer weights, so this separate tokenizer repository is mainly needed to reproduce training or initialize a new policy.

License notice

The FSQ/RFSQ implementation used to create this checkpoint includes code adapted from EPFL and Apple Inc. under the EPFL-Apple Sample Code License (Non-Commercial). Review the corresponding license terms in the linked LeRobot source before redistribution or commercial use.