satoshi3/attune-l3eW

Attune Vector Level 3 policy (vector_v1.1)

A fine-tune of domFree/attune-j4 at revision 4a2651b08cb70332bfe9b86ef9b2efaff1825eb8, for the Robotensor Attune Vector lane (Level 3).

• Weights: model.safetensors, sha256 e626f3682889995f05399df7e9fee0ffb6e169b5c15ef92d2545ef42fc1ebcf0.
  • It has the same 753-tensor vector_v1.1 layout as the parent.
• Unchanged from the parent (bit-identical):
  • the CLIP vision encoders;
  • the action/proprioception normalizers;
  • the proprioception and prompt-pool layers.
• Trained: the transformer decoder, the diffusion U-Net, and the position/sink embeddings.
• Method: a weight-space ensemble (WiSE-FT) of the parent and a fine-tune of it: theta = parent + 0.5 x (fine-tune - parent).
• The fine-tune:
  • 20,000 steps, batch 64, learning rate 1e-5 (cosine), EMA 0.999, epsilon-prediction diffusion loss.
  • Imitation of expert demonstrations on all sixteen tasks, with the four weakest tasks up-weighted (beat_block_hammer, place_mouse_pad, place_dual_shoes, move_stapler_pad).
  • Expert-takeover recovery episodes (the parent acts, a scripted expert finishes).
  • A function-space anchor to the frozen parent on the other twelve tasks (weight 2).
• Data: Level 3 episodes generated with the RoboTwin-Vector harness (franka-panda, public seed range below 2,000,000). Each target is prompted with an expert demonstration from a different scene of the same task.

The parent repository declares no license; this card makes no claim beyond what the parent's terms allow.