Robots Need to Feel Friction to Truly Understand the Physical World

Key points

• Contactile, an Australian tactile-sensor company, proposes a new AI model class called VμA (vision-mu-action) that feeds robots real-time friction data alongside camera vision.
• Current robot AI models lack a key physical value: the coefficient of static friction (mu), which determines whether a gripped object slips or holds.
• Contactile's PapillArray sensor computes friction and slip state at the sensor itself, with no processing delay.
• The gap between simulation training and real-world performance is blamed largely on the absence of friction physics in training environments.

A robot arm reaches into a box, grips a wet bottle, and drops it. The camera saw the bottle; the joints felt the load. Neither told the robot how slippery the surface was, and that single missing fact is why robot grippers still fail on tasks a toddler handles without thinking.

Contactile, an Australian sensor company, argues in a new technical piece first reported by The Robot Report that this gap runs deeper than anyone has fixed. The problem sits inside the AI models robots use to reason about the world.

What is a world model, and why does it need friction?

A world model is the robot's internal picture of physical reality: software that predicts what happens next so the robot can plan. Think of it as the robot asking itself, "If I squeeze harder, will this egg crack?" A good world model means the robot doesn't have to memorise every task. It can reason through new ones.

But a world model is only as good as the information it starts with. Today's models condition mostly on two inputs: images from cameras and joint positions from the robot's own motors. For moving through open space, that's often fine. Contact-rich manipulation is a different matter.

The missing number is the coefficient of static friction, written as mu. It decides whether a grasped object stays grasped, and it shifts depending on material, surface condition and temperature. A camera can't see it. Motor currents can't calculate it accurately. A pressure map tells you something is being gripped, not whether the grip is about to fail.

Without mu, a robot's world model reasons from an incomplete picture. It can learn patterns from training data but can't reliably transfer those lessons to a new object or a changed surface, because the variable that governs grip success simply isn't there. AI2Day covered the wider world-model race on 13 July 2026 in "Beyond Chatbots: World Models Are Teaching AI to Simulate Reality"; friction conditioning is exactly the gap that piece flagged as unsolved.

The sensor that measures what cameras cannot

Contactile's PapillArray Tactile Compute Module is a fingertip-sized sensor array that sits on the robot's gripper. Each tiny sensing point, called a taxel, measures force in all three directions. The sensor then calculates mu directly at the sensor itself, with no upstream processing delay.

Its outputs include the full 3D force at every taxel, global force and torque across the contact patch, the live friction value, the slip state at each point, and the minimum grip force needed to prevent a drop. These are grounded in contact physics, not learned guesses.

Contactile proposes naming models that use this full signal stack VμA: vision-mu-action. Earlier model generations got their own names when a new input became important enough. Vision-language-action (VLA) models added natural language. Vision-force-action (VFA) models added force. VμA adds friction physics that were absent from every prior approach, and Contactile says the PapillArray is currently the only sensor that computes all of these values at the sensor edge.

On the training side, VμA datasets gain causal physical signals rather than statistical proxies. During deployment, a robot with a PapillArray knows the friction value of what it's holding right now and can adapt its grip the moment conditions change.

For industries where robots handle wet produce or variable-surface packaging, getting grip physics right isn't a research curiosity. It's the difference between a working production line and one that needs constant human correction.

Should you worry about the hype?

Contactile is the company proposing VμA, and the PapillArray is its own product, so the case deserves scrutiny. The physics argument is sound: mu is a real quantity that current sensors genuinely don't deliver. What's missing is independent evidence that VμA-conditioned models outperform VFA models on standardised benchmarks. Watch for third-party trials before treating the model class as settled.