franka_fold_towel

LingBot-VA Franka Fold-Towel artifacts

This repository contains the Fold-Towel SFT actor and its paired language-conditioned QTran-L critic.

Layout

sft/checkpoint_step_21000/
  config.json
  va_franka_fold_towel_sft_self_contained_cfg.py
  transformer/
  resume_manifest.json
  training_state.pt
  rank_*.pt
critic/checkpoint_00030000/
  config.json
  manifest.json
  training_state.pt
  dataset_config_compat.json
  federated_catalog.json
  language_sidecar_manifest.json
  episode_success_overlay.json
critic/provenance/
  raw_training_config.json
  raw_manifest.json
  nebula_runtime_config.json

The SFT checkpoint contains the fine-tuned transformer. A deployment still needs the matching LingBot-VA base VAE, tokenizer, and text encoder. The actor contract is a 30-dimensional action, frame_chunk_size=2, and the Fold-Towel sequential composite image (384x480 RGB, left/right wrists on top and head below). The task instruction is: fold the towel in half twice.

The critic is a language-conditioned twin_language_query_transformer_v1 IQL critic with window_size=4, infer_latent_chunk_size=2, and feature taps [0,4,8,12,16,20,24,28]. Its uploaded config uses relative metadata paths; the original Nebula training config is retained under critic/provenance/. The critic checkpoint must be paired with the SFT transformer contract.