Researchers at Hugging Face have unveiled N0-TWAM, a tactile-native world-action model designed for contact-rich manipulation. Unlike conventional models that treat touch as an auxiliary input, N0-TWAM predicts future contact and future vision under the same objective and at the same causal step, then derives actions from that jointly predicted future. This approach, the team claims, makes it the first tactile world-action model trained at large scale.
The model was pre-trained using visuo-tactile joint training on tactile-rich demonstrations spanning six robot embodiments and 450 tasks. A key component is NeoForce, a unified force-based tactile representation that provides a physically grounded contact signal to condition action generation. To handle long-horizon and multi-stage manipulation, N0-TWAM introduces tactile contact events for task staging, advancing through them during execution.
For real-time efficiency, the architecture employs an asymmetric Mixture-of-Transformers design: a full-width expert handles video prediction, while slim experts manage downstream action and tactile prediction. Evaluations on both real and simulated benchmarks demonstrate the model's capabilities across a range of contact-rich tasks and highlight the benefits of data scaling for precise tactile and action prediction.
The codebase and pretrained checkpoints are publicly available on GitHub, and the project page offers further details. The authors invite feedback, especially from researchers working on tactile sensing, vision-language-action models, or world models.