A 𝒩₀-VTLA single-arm policy for the UniVTAC simulator, post-trained from the 𝒩₀-VTLA pretrained base. This card documents the insert_hole setup.

model.safetensors                          the full checkpoint (~8.25 GB)
assets/n0_insert_hole_norm/norm_stats.json
config.json                                architecture summary

The evaluation adapter speaks ZMQ + msgpack, so use scripts/serve_zmq.py from the code repository:

Serving reads only model.safetensors and assets/<asset-id>/norm_stats.json; no dataset is needed. A correct load prints tactile=True with 2 views and no missing or unexpected state-dict keys. If it reports either, the config does not match the checkpoint.

Then run the evaluation from a UniVTAC checkout with a deploy YAML pointing at port 5557. Set exec_horizon to 50, the model's action horizon: executing fewer steps clips the tail of each chunk, which is where the grasp-closing motion lives.

Unlike the pretrained base, which predicts end-effector deltas in a canonical 32-dim rot6d container, this policy predicts joint actions: 7 joints plus gripper. The joint dims are element-wise deltas against the current state; the gripper column is absolute. Do not feed it end-effector data or reuse an end-effector normalization asset.

These weights are derived from Google's PaliGemma/Gemma parameters and are made available under the Gemma Terms of Use and the Gemma Prohibited Use Policy, not under the CC BY-SA 4.0 licence that covers the source code. This is inherited from the base model.