1/ Robots can see better than ever. They can recognize objects. Understand language. Estimate depth. Predict actions. But ask a robot: “Are you holding this glass too tightly?” And vision suddenly isn't enough. The next leap in Physical AI may depend on a sense AI has largely ignored: Touch. @axisrobotics @KaitoAI 🧵 2/ For the last decade, AI has been dominated by things machines can see and read. Text. Images. Video. And that makes sense. The internet gave us enormous datasets for those modalities. But the physical world contains information that never appears in pixels. 3/ Pressure. Friction. Resistance. Compliance. Slip. Texture. Contact force. Deformation. These aren't just properties of objects. They're properties of interaction. You often don't know them until you touch something. 4/ And that's the fundamental difference between digital intelligence and physical intelligence. A language model can understand the concept of a glass. A vision model can recognize one. A robot has to answer a much harder question: How do I physically interact with it without dropping or breaking it? 5/ Vision can tell a robot: “Your gripper is around the glass.” Touch can tell it: “The glass is starting to slip.” Vision: “The connector is near the socket.” Touch: “You're slightly misaligned and pushing against the edge.” Different information. Different kind of intelligence. 6/ This is why I think one sentence captures the role of touch extremely well: Vision helps decide what to do. Touch helps decide whether to keep doing it. Touch closes the loop. 7/ Humans use this constantly. Try plugging a cable into a port. You look at the port. Move your hand toward it. Then contact begins. And something changes. Your eyes become less important. Tiny corrections start happening through: touch + force + proprioception. 8/ You feel resistance. Adjust the angle. Push slightly. Feel alignment. Continue. You don't consciously calculate any of this. Your nervous system closes the loop continuously. Robots need something similar. 9/ And this reveals something important about manipulation: The last centimeter may be harder than the first meter. Moving a robotic arm across a table can be mostly a vision-and-planning problem. But the final millimeters become a physics problem. Contact. Force. Friction. Compliance. Timing. 10/ Consider inserting a peg into a hole. From one meter away: easy. From ten centimeters: still mostly perception. At the point of contact? A tiny positional error can completely change the task. Suddenly the robot needs to know: Did I make contact? Where? At what angle? How much force? Should I push or correct? 11/ This is why contact-rich manipulation remains so difficult. Tasks like: inserting, plugging, screwing, peeling, folding, cutting, grasping soft objects, handling cables, or manipulating tools require more than identifying what an object looks like. They require understanding what is happening physically right now. 12/ Recent research is starting to make this much clearer. SimpleTouch, released in October 2026, added a tactile expert to π₀.₅ and reported an average 77.5% success rate across six contact-rich UniVTAC tasks, versus 45.2% for one π₀.₅-based comparison method. Across four real-world tasks, it reported a 71.3% average success rate. 13/ The interesting part isn't just the number. It's what the experiment suggests. You don't necessarily need to replace the VLA. You can start with an already capable model... then give it another sensory channel: contact. The model already knows broadly what to do. Touch helps it execute the final physical interaction. 14/ Another recent system, T-Rex, approached the problem from a different angle. It collected roughly 100 hours of tactile-rich manipulation data and introduced an architecture designed to process touch at a different temporal rate from the rest of the policy. Across 12 delicate-force and deformable-object tasks, the authors report more than a 30% average success-rate improvement over their strongest baseline. 15/ And that points to a deeper issue: Touch operates at a different timescale. Language doesn't change every millisecond. The semantic identity of an object doesn't either. But contact can. A grasp can be stable now... and slipping moments later. 16/ So a Physical AI system may need multiple speeds of intelligence. A slower system asks: “What am I trying to accomplish?” A faster system asks: “Is the object slipping right now?” High-level reasoning and low-level physical reflexes are different computational problems. 17/ Humans already work this way. You don't consciously reason through every micro-adjustment when holding a fragile cup. Your sensory system reacts faster than deliberate thought. Touch triggers correction before you need to formulate: “I have detected a 3mm lateral displacement.” 18/ This suggests future robot architectures may become increasingly hierarchical. Language / reasoning “What task am I doing?” ↓ Vision / world understanding “Where are the relevant objects?” ↓ Action policy “How should I move?” ↓ Tactile reflex “Something changed. Correct now.” 19/ And touch doesn't have to live only in the fingertips. A 2026 Nature Sensors system called EmArm integrated large-area soft tactile skin across a robotic arm alongside proprioception. It demonstrated contact-aware manipulation and tactile-driven trajectory replanning, including situations where vision was occluded. 20/ That changes how we should think about robotic skin. It isn't just: “Can the robot feel an object?” It's also: “Can the robot understand when the world touches the robot?” A human bumps the arm. An obstacle appears. The robot brushes against furniture. Physical intelligence includes the entire body. 21/ This becomes even more important for humanoids. A humanoid operating around humans won't interact with the world only through its hands. Its arms, legs, torso, and body will constantly make contact with dynamic environments. A truly embodied machine may need a sense of its own physical boundary. 22/ Touch also changes how we define success. Imagine a robot receives the instruction: “Pick up the strawberry.” It picks it up. Task completed. But it squeezed so hard that the strawberry was crushed. Did the robot succeed? 23/ Most benchmark-style thinking might say: Object lifted = success. Physical intelligence should say: No. Because success isn't only: “Did you reach the target state?” It's also: “How did you interact with the world while getting there?” 24/ This is the idea behind gentle manipulation. A robot may need to complete the same task differently depending on whether it's holding: a steel component, an egg, a strawberry, human skin, or a surgical instrument. The goal may be identical. The acceptable force is not. 25/ Tabero, presented at ICML 2026, explores exactly this kind of problem by combining vision, touch, language, and closed-loop force feedback. Importantly, its evaluation considers not only whether the task succeeds, but also the quality of the physical interaction. 26/ That's an important evolution. Because eventually, intelligent robots shouldn't just understand: what to do. They should understand: how carefully to do it. A truly intelligent robot shouldn't just know whether it succeeded. It should know how hard it touched the world while succeeding. 27/ And this connects directly to the previous Embodiment Gap thesis. We talked about how the internet contains enormous amounts of human video. Humans cooking. Repairing. Assembling. Cleaning. Using tools. But video has a fundamental limitation: it records what an interaction looked like. Not what it felt like. 28/ Watch a video of someone inserting a tight connector. Visually, almost nothing may happen. But the human feels: contact, resistance, misalignment, a tiny correction, alignment, then insertion. The camera sees: hand moves forward. The nervous system experiences an entire physical conversation. 29/ This is one reason: Human video is not complete physical data. It can show us: what task happened, which objects mattered, how things moved, and what success looked like. But it usually doesn't tell us: force, pressure, friction, or tactile feedback. 30/ And that's why robot learning may ultimately require several complementary data layers. Human video gives scale and semantic knowledge. Simulation gives controllable experience. World models give imagined futures. Robot deployment gives real failures. Touch gives the missing information at physical contact. 31/ Eventually, tactile data itself may become a foundation-model problem. Today we talk constantly about: Vision Foundation Models. Language Models. VLA Models. But what happens when millions of hours of robot interaction produce enormous tactile datasets? 32/ Could a model learn general concepts like: slipping softness stiffness stable contact excessive force surface texture deformation across many objects, tasks, sensors, and embodiments? In other words: Could we eventually have a foundation model for touch? 33/ There are signs of this broader direction already. A 2026 Nature Sensors system combined fast tactile processing with higher-level language-guided reasoning, explicitly separating rapid responses from deeper semantic interpretation. It's an early example of how robotic touch could evolve from raw sensing into something closer to tactile cognition. 34/ And that distinction matters. A temperature sensor measures temperature. A force sensor measures force. But physical intelligence requires more than measurement. The system has to infer: What does this sensation mean? Is the object slipping? Is the contact dangerous? Is this resistance expected? Should I continue? Should I stop? Should I change strategy? 35/ That's the difference between a robot that has tactile sensors... and a robot that can understand touch. The same distinction existed in vision. Cameras existed long before machines could interpret scenes. Tactile hardware may be entering a similar transition. 36/ There is still a major scaling problem. Vision benefited from the internet. Tactile learning doesn't have an ImageNet-sized natural dataset sitting online. Touch requires physical interaction. Different tactile sensors also produce very different representations. And collecting high-quality contact data can be expensive. 37/ So tactile intelligence faces a familiar problem: Where does the data come from? Which brings us all the way back to the beginning of this series. Deployment may matter again. Every robot operating in the world could generate: vision, actions, force, proprioception, touch, successes, and failures. 38/ A large deployed fleet wouldn't just create visual trajectories. It could create a growing dataset of: what physical interaction feels like. And that may be far rarer — and more valuable — than another billion images. 39/ Imagine a future model trained across millions of interactions. It has touched: glass, fabric, metal, plastic, fruit, tools, cables, human hands, rigid objects, soft objects. It begins learning a general representation of physical contact. 40/ Now combine that with: Vision for understanding what exists. Language for understanding what we want. World Models for predicting what might happen. Touch for understanding what is happening physically. Action for changing the world. That's starting to look much closer to real embodied intelligence. 41/ And touch may become especially important as robots move from structured factories into messy environments. Homes. Hospitals. Restaurants. Warehouses. Construction sites. Around people. The less predictable the environment becomes... the harder it is to pre-program every physical interaction. 42/ Vision can reduce uncertainty before contact. But once contact begins, the world starts talking back. The object resists. Slides. Bends. Compresses. Moves unexpectedly. A robot needs to listen. 43/ So maybe the next phase of Physical AI isn't simply about giving robots better eyes. It's about giving them something they've largely been missing: a nervous system. 44/ AI learned the digital world through: text, images, and pixels. But the physical world is not composed only of things that can be seen. It also contains: pressure, friction, resistance, softness, contact, force, and pain. 45/ A robot that cannot feel those things may know what an object is... but never fully understand what it means to interact with it. Vision may get robots to the object. Touch may teach them what to do once they arrive.