pick_tube_05_tactile_decoder_deploy

Pick Tube 05 Direct Tactile Action Decoder

Private deployment payload for the VB3 direct tactile action decoder ablation.

The runtime composes:

1. frozen lerobot/smolvla_base at revision c83c3163b8ca9b7e67c509fffd9121e66cb96205;
2. frozen KaiyueChen/pick_tube_01 PEFT adapter at revision c2bb4296cf7405ac3c0ad89e6f577fa620a660a6;
3. frozen converted tactile ResNet18;
4. formal action_tactile direct Transformer decoder.

The decoder predicts a physical-unit action chunk shaped [1, 20, 20] after the saved action postprocessor. Four current-frame tactile RGB images are required in this exact order:

• observation.images.tactile_left_0
• observation.images.tactile_right_0
• observation.images.tactile_left_1
• observation.images.tactile_right_1

Layout

• smolvla_base/: base config and full weights.
• smolvla_adapter/: PEFT adapter and saved pre/postprocessors.
• tactile_encoder/: converted encoder weight and parity manifest.
• decoder/best.pt: formal action+tactile checkpoint selected at epoch 3.
• tokenizer/: pinned SmolVLM2 tokenizer/processor assets.
• deployment_manifest.json: revisions, tensor contract, and key hashes.
• docs/NETWORK_STRUCTURE.md: complete architecture and evaluation protocol.

Training data, formal feature cache, optimizer state, smoke checkpoint, and action-only checkpoint are intentionally excluded.

This payload contains the required weights, but the VB3 real-robot client must use the tactile decoder deployment integration on the ablation branch. The offline evaluation does not establish real-robot safety or success.

Download

hf download wjstx/pick_tube_05_tactile_decoder_deploy \
  --local-dir /home/ljl/assets/pick_tube_05_tactile_decoder_deploy