Part 7 of a series on Physical AI, tactile sensing, and the road to dexterous machines.
If you want to see the hardest unsolved problem in robotic touch, you don't need a lab. You can watch it happen on a trade-show floor.
Spend a few days at a robotics conference and the pattern repeats. A dexterous hand rolls a ball between its fingers, flips a cube, picks up something delicate, and the crowd is rightly impressed. But linger, and the cracks show. A compliant fingertip that looked flawless on day one is visibly worn by day three, its soft skin abraded by the same repeated demonstration. A hand starts firing phantom contact signals while touching nothing at all — because its own motors have warmed it beyond what its baseline can track. A force sensor that aced its datasheet barely registers a gentle touch in person.
None of these are failures of intelligence. The planner and model may perform exactly as intended. What broke was something far less glamorous and far harder to fix: the sensor's ability to report physical contact accurately under real-world conditions, over time. That gap between a tactile sensor that works in a demo and one that remains trustworthy through thousands of cycles is where the field stands today. And it's mostly unsolved.
Reliability is the bottleneck, not data
There's a comforting story in robotics that says the path to dexterous machines runs through more data and bigger models. Collect enough grasps, train a large enough policy, and touch will follow. The researchers closest to the problem are increasingly skeptical of that story. The bottleneck they keep naming isn't data volume; it's reliability and high-frequency feedback. The logic is blunt: you can't build a useful product on a system that starts dropping objects consistently after a few days of operation.
There's an important wrinkle here worth being precise about, because the industry can look like it's contradicting itself. Walk the same floor and you'll see a dozen vendors hawking sensor-laden data-collection gloves, apparently insisting the field is starved for data. Both things are true, because they're about different data. Visual data is abundant: the internet is a near-infinite training set for vision. Tactile data barely exists, because you cannot scrape the feeling of a wet glass from a video. So the field is simultaneously drowning in pixels and parched for force. But neither problem is the one standing between a good demo and a deployable hand. That one is reliability — and it has a short list of culprits: hysteresis, drift, and wear.
Problem one: hysteresis — the skin that remembers
Hysteresis is perhaps the most pervasive challenge. It describes a sensor's tendency to lag its own input. Press a soft tactile fingertip and it reports a force. Release it, and a good sensor should snap instantly back to zero. A real one often does not. The elastomer holds a fading memory of the last grasp, like a thumbprint slowly filling back in on memory foam; the reported force lingers after the object is long gone. During fast manipulation, that stale signal may tell the controller the hand is still loaded when it is empty, disrupting the real-time feedback loop. Materials scientists often identify hysteresis as a core challenge in soft tactile sensing. It affects many approaches built on compliant skins, including capacitive arrays whose force maps may return slowly to baseline.
What makes hysteresis interesting rather than merely discouraging is that it's a materials problem, not a physics dead end. That's the encouraging shape of this whole category: the hard problems aren't waiting on a breakthrough in transduction. They're waiting on better materials, better compensation, and the discipline to pair the two.
Problem two: baseline drift — zero won't hold still, especially when it's hot
Earlier in this series, we discussed baseline drift in touchscreens: the gradual shift in a sensor's no-touch reference as conditions change. A sensing system must continually re-learn what no touch feels like. On a phone, the drivers are temperature and humidity. On a robot hand, the problem is far worse, because the hand is full of motors that heat the entire assembly as it works, shifting the electrical properties of the skin, dragging the baseline around with it.
The failure mode is exactly what you'd predict, and it shows up live: a hand that has simply gotten warm begins reporting contact that isn't there, triggering grips and reactions to phantom objects. The only defense is the same one that made high-volume touch sensing work — an adaptive baseline that tracks the drift without swallowing genuine slow signals, supported by differential measurement and an on-board temperature sensor with compensation tuned against it. This is where owning the full signal chain stops being a talking point and becomes the difference between a sensor that holds a stable reference through a working shift and one that hallucinates by lunchtime. Notably, the approaches that fare best here are the ones whose thermal behavior can be characterized and compensated in firmware, rather than fought mechanically.
Problem three: the surface wears out
Then there's the simple, brutal fact that a robotic fingertip lives a violent life. It's pressed, dragged, and abraded thousands of times a day against everything from soft fabric to sharp metal. Soft materials that feel beautifully sensitive when new degrade in ways that are easy to overlook on a spec sheet: the surface abrades, the elastomer develops stiction and sticks instead of releasing cleanly, reporting a sustained pressure that isn't there. The careful mechanical recovery described in "From Glass to Skin" decays grasp by grasp. Compliant tips are, for now, closer to a consumable than a permanent component. Designing coatings and elastomers that survive millions of cycles without losing their feel is foundational, unglamorous, and genuinely unsolved.
The unifying idea: every sensor has a kryptonite
Step back from the individual failures and a clean organizing principle emerges, one that the most clear-eyed people in the field have started saying out loud: no single sensing modality wins outright, because every one of them can be broken by a specific aggressor.
Magnetic sensing is fast and captures shear beautifully until a stray magnetic field walks past. A magnetically sensed fingertip can be thrown into reporting phantom forces by something as ordinary as the small magnet inside a pair of wireless earbuds. Camera-based optical sensors produce gorgeous, high-resolution contact images and the richest training data of any approach. However, they're bulky, run slower, generate a firehose of data, and have soft windows that wear over time. Piezoelectric materials hear the high-frequency shudder of an incipient slip better than almost anything — but they physically cannot hold a reading on a static force, so they can't tell you how hard you're squeezing a stationary object. Each technology, in other words, is defined as much by the interference it tolerates as by the spec it targets.
That doesn’t mean there’s a single winner. But it does sort the weaknesses into two very different piles. Some are fundamental to the physics of the approach — a piezoelectric element will never hold static force; a magnetometer will always live in a world full of magnets; and a camera will always need focal depth and produce a flood of data. Others are materials and engineering problems that yield to effort over time.
It's worth being completely candid about where capacitive sensing sits in this picture, because the honest answer is what makes the rest credible. Capacitive's classic weakness is bandwidth: slip's telltale micro-vibrations live up in the hundreds of hertz, and a typical integrated capacitive skin, sampling slowly, simply cannot hear them. On its own, a capacitive array is best understood as a superb detector of grip state — contact, force distribution, shear, the migrating center of pressure that warns of an impending loss of stability — rather than a true detector of the first millisecond of incipient slip. That is a real limitation, and pretending otherwise would be dishonest.
But notice which pile it falls into. It isn't a dead end of physics; it's an architectural problem with a known, cheap fix — the very two-channel design this series has been describing since Part 2. Pair the thin, dense, low-cost capacitive array (the static and spatial channel) with a fast dynamic transducer for the vibration transient, and the hybrid does what neither half can alone. The capacitive backbone keeps its decisive advantages — it's thin enough to fit a real finger, manufacturable at volume, low in power, and amenable to temperature compensation in firmware — while the fast channel supplies the high-frequency hearing. A modality whose remaining weaknesses are the tractable kind, and that brings that profile of advantages, has a far cleaner road to the mass market than one whose weakness is baked into its physics.
Apply that to the three problems above, and the case for a capacitive backbone becomes clear. On drift, it has a long track record: adaptive baselines, differential measurement, and firmware temperature compensation are the same techniques refined across decades of high-volume touch sensing. On hysteresis, capacitive arrays are not immune — the elastomer still remembers — but because the sensor's behavior can be characterized precisely, that lag can be modeled and compensated in firmware rather than accepted as error. On wear, the sensing electrodes sit beneath the skin rather than at the surface that takes the abuse, so the outer layers can be engineered for durability independently of the electronics. None of this makes the problems disappear. It makes them the tractable kind — and in a field where reliability is the bottleneck, that is the difference between a demo and a product.
Why the boring problems are the moat
This should change how you read every dexterous-hand demo from now on. None of these unsolved problems, including hysteresis, drift, wear, bandwidth fusion, and the long tail of aggressors, is solved by a cleverer transducer or a bigger model. They are solved through materials science and signal processing that holds a stable reference while the world shifts underneath it. They require silicon, firmware, sensor, and algorithm co-designed to fight noise across every layer, along with fusion logic that makes a fast channel and a slow channel into one coherent sense. They also require the hard-won, decades-deep instinct for how real-world sensing fails that you can only earn by shipping at scale.
That's the quiet thesis underneath this entire series. The flashy part of robotic touch — the resolution, the demo, the reconstructed contact image — is not where the difficulty lives. The difficulty lives in the boring, durable, system-level reliability that determines whether a sensor is still telling the truth on day five hundred. It doesn't fit on a spec sheet. It's also exactly the kind of problem that takes a very long time to get good at — which is precisely why it's worth so much.
In short, this field is earlier than its best demos suggest. The hardest problems are real, and no one has fully cracked them. Far from being a reason for pessimism, that’s why the opportunity is as large as it is. It also leads to the question we turn to next: if reliable touch at scale is this hard, who is positioned to deliver it, and how big is the prize for whoever does?
Next in the series: The Trillion-Touch Opportunity — the market for machines that can feel, the scale of the dexterity wave that's coming, and why the company that can ship reliable touch at volume stands to define an era.