Robots can't feel friction. A new model class says that's the missing sense.

A robot can register how hard it is pressing and how much its fingertip is deforming. It can read joint torques and the pose of its own hand in space. What it cannot do is directly measure the friction coefficient at the patch of skin that is actually touching the object: the tangential force that decides whether a grasp holds, a slide slips, or a tool bites into the surface.

That gap is the claim behind The Robot Report's analysis of a proposed model class called VμA, which makes friction a first-class input to the world model rather than an inferred proxy. The piece argues that current robot "world models," the internal simulators a robot uses to predict what will happen when it touches, pushes, or slides something, are conditioned on a sensor vocabulary that does not include the slip/friction state.

Most tactile sensors today encode contact as deformation images, pressure maps, or three-axis force. They report that contact is happening and how hard, not the slip/friction state at the interface. Robots then approximate friction from motor currents, a proxy that holds for slow, steady grasps on familiar materials and falls apart on deformable objects, unfamiliar surfaces, or any task where the contact patch is changing in real time.

Three preprints from the last two years converge on the same gap from different angles. Visuo-Tactile World Models (arXiv 2602.06001) shows that touch lets robots distinguish stiffness, friction, and roughness that vision alone cannot resolve, and that conditioning the world model on tactile signals improves contact-rich rollouts. OmniVTA (arXiv 2603.19201) reaches a similar conclusion: tactile signals should model contact dynamics explicitly, not be absorbed as passive observations. And a Sim2Real study of static friction in legged locomotion (arXiv 2503.01255) shows that friction is routinely excluded from domain randomization, and that adding it back materially improves real-world transfer.

The falsifier for the whole argument is sharp. If friction could be cleanly recovered from proprioception and deformation images, the way depth can be inferred from stereo, then "make friction first-class" reduces to an engineering exercise in better feature extraction. The Sim2Real literature does not currently show that it can. Static friction resists that kind of recovery, and the contact-rich tasks the field cares about (assembly, deformable handling, tool use on unfamiliar surfaces) sit exactly where the proxy breaks.

This is also a money problem, not just a sensors problem. Bessemer Venture Partners' Atlas on world models frames world models as the route past robotics' data wall, with more than $3 billion projected across robot data modalities including tactile and force. The conditioning question, what physical variables the world model is allowed to read, sits upstream of how much data any of that spend actually buys.

For a non-beat reader, the practical test is simple. When the next world-model paper arrives, ask which physical variables the architecture treats as first-class inputs and which it asks the model to infer. If the answer is "vision plus proprioception plus tactile pressure" with friction derived downstream, it is doing the old conditioning work in a new wrapper. If friction, or a directly sensed slip state, sits at the input layer, the field has moved.