NeoteAI's N₀-TWAM Adds Touch to Robot World-Action Models

Robot foundation models have been almost entirely a vision plus language story so far, and that has been a real ceiling on the tasks people actually want robots for. You can name a lot of contact-rich work (screwing something in, seating a peg in a hole, mating a connector) where being able to see the scene is not enough on its own. A new paper from the NeoteAI and Fudan TEAI teams, posted to arXiv, argues that adding touch as a first-class modality is now doable at scale.

The model, called N₀-TWAM, is described by the authors as the first tactile world-action model trained at large scale, with visuo-tactile joint training spanning six embodiments and 450 tasks. Rather than treating force as an auxiliary signal, it predicts both future vision and future contact, and grounds the tactile side in what the paper calls NeoForce, a unified force-based tactile representation. Architecturally the trick is an asymmetric Mixture-of-Transformers that pairs a full-width expert for video prediction with slim experts for downstream action and tactile prediction, so the heavy compute goes into modelling the world while the action head stays lean. There is also a mechanism the authors call tactile contact events for task staging, aimed at multi-stage manipulation.

Why this matters if you do not build robots: the manipulation tasks that unlock real deployments (assembly, insertion, cable routing, tool use) all live in the contact-rich regime that visual-only policies keep failing on. A model that has actually learned what a firm seat versus a slipping grip feels like, across multiple robot bodies, is the kind of foundation piece that could compound. The paper also indicates that code and model checkpoints will be released publicly, which is the part that would matter most to smaller labs.

The honest caveat is what the abstract itself does not give you. There are no named baselines in what has been made public, the six embodiments are not enumerated in the summary text, and there are no head-to-head benchmark numbers to sanity-check the 'first at large scale' framing, which is a claim from the authors rather than an independently verified result. Sim-to-real transfer for contact dynamics is notoriously brittle, and force-representation choices tend to be sensor-specific, so the cross-embodiment generalisation claim is exactly the part outside observers will want to prod first.

If N₀-TWAM holds up when the code drops, the interesting downstream is who picks it up: manufacturing integrators and dexterous-manipulation startups that have been stuck on visual policies now have a plausible starting checkpoint, and an open release would mean the tactile layer of robot foundation models stops being closed proprietary territory.