N_0-TWAM: Scaling Tactile-Native World-Action Model for Contact-Rich Manipulation
The model is built to predict future touch signals alongside future video and actions for contact-heavy robot manipulation.
N_0-TWAM is pretrained on visuo-tactile demonstrations across six embodiments and 450 tasks. The paper uses a unified force-based tactile representation, NeoForce, to condition action generation with contact information. It also introduces tactile contact events to stage longer manipulation tasks during execution. The authors report real and simulated evaluations showing gains from data scaling for tactile and action prediction, with code and checkpoints planned for release. HF Daily Papers' note
N_0-TWAM is pretrained on visuo-tactile demonstrations across six embodiments and 450 tasks. The paper uses a unified force-based tactile representation, NeoForce, to condition action generation with contact information. It also introduces tactile contact events to stage longer manipulation tasks during execution. The authors report real and simulated evaluations showing gains from data scaling for tactile and action prediction, with code and checkpoints planned for release. HF Daily Papers' note
score 6