SO101 OAT paired-RFSQ action tokenizer — step 320,000
Action tokenizer trained on maxlium/so101-box-to-plate: 6 action dimensions, horizon 32, 8 latent positions, 2 residual FSQ stages (16 full action tokens), FSQ levels [8, 5, 5, 5]. Uses dataset MIN_MAX action normalization.
action_tokenizer.safetensors: tokenizer weights.tokenizer_config.json: original training configuration.training_state.pt: optimizer, scheduler and step for continuation.
Download the repository with huggingface_hub.snapshot_download and provide the resulting local directory as policy.action_tokenizer_path when training a new compatible OAT policy. The custom LeRobot oat_rfsq_pair implementation is required. Training configuration paths refer to the original machine and may need adjustment.
Fixed training-data sample reconstruction MSE at full 16-token length: approximately 0.0003231, measured on 1,280 windows across 40 training episodes in evaluation mode. This is not held-out validation.
Companion policy: yangzhixing/oat-rfsq-pair-so101-policy-tok320k-step18000.