n0-twam-neosim-delta

N0-TWAM post-training — NeoSim 12 tasks (4 single-arm + 8 dual-arm) / pi05_delta (horizon delta)

Multi-task post-trained N0-TWAM checkpoint (final release). Trained on marker-less rgb GelSight tactile streams; each task uses its own normalization statistics.

Serving note. Multi-task checkpoints must be served with the per-task norm / camera+tactile keys / action channels of the task being evaluated — the pooled envelope in train_meta.json is an unreachable fallback and would de-normalize actions at the wrong scale. Use the multitask_server config (TWAM_SERVE_POOL / TWAM_SERVE_TASK env — see DEPLOY.md); prompts verbatim from the table below.

Contents: transformer/ (config.json + safetensors, local_tactile tensors = 18), train_meta.json (training snapshot), and norm/ (per-task normalization stats + conversion scripts).

Per-task normalization stats

The exact per-task q01/q99 tables used in training are released under norm/, together with the raw-quantile reports and the HDF5 → LeRobot conversion scripts. The element-wise envelope of these tables reproduces the pool-level norm_stat in train_meta.json bit-exactly.

Task prompts

Send the training prompt verbatim at serve/eval time (also provided as norm/PROMPTS.json):

Citation

If you use this model in your research, please cite the N0-TWAM paper:

@misc{neoteaiteam2026n0twamscalingtactilenativeworldaction,
      title={$N_0$-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation},
      author={NeoteAI Team and Fudan TEAI Team},
      year={2026},
      eprint={2607.23783},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2607.23783},
}