DexTacWAM Gives Robots a Sense of Touch

🔥 What happened Researchers from UC Berkeley and collaborators unveiled DexTacWAM, a world-action model that fuses video prediction with per-fingertip tactile sensing. On a 22-DoF bimanual platform it wins all six contact-rich manipulation tasks, averaging 70.6 versus 38.0 for the strongest baseline. 💡 Why it matters The real lever isn't tactile conditioning — it's modeling contact dynamics as part of the predicted world state: strip that out and the four-task mean collapses from 74.7 to 26.6. Just four hours of adaptation with ~100 demos per task extends a pretrained video model to touch, losing only 0.5 dB of visual prediction quality. ⚡ Our take Vision-only robotics just became a legacy architecture. Anyone still building dexterous manipulation without tactile sensing in 2026 is burning compute on a problem skin solved millions of years ago.