Sensoryczny klif: dlaczego roboty podnoszą 50 kg, ale nie czują winogrona

Here is a test. Pick up a raw egg. Do not think about it. Just do it.

Your hand approached the egg from a direction determined by your visual system. Your fingers wrapped around it with a grip pattern selected unconsciously based on its shape. As your fingertips made contact, the mechanoreceptors in your skin, roughly 17,000 of them in each hand, reported pressure, texture, slip, temperature, and curvature simultaneously. Your nervous system processed this information in real time, adjusting grip force to hold the egg securely without cracking it. The entire operation took about 1.5 seconds and required zero conscious thought.

No humanoid robot on Earth can do this reliably.

Boston Dynamics Atlas can lift 50 kilograms overhead. Tesla Optimus Gen 3 has 22 degrees of freedom in its hands. AgiBot’s A2 has 49 degrees of freedom across its entire body, the most of any commercially available humanoid. But hand an egg to any of them and the most likely outcome is either the egg slipping through ungripped fingers or the egg being crushed by fingers that cannot feel how hard they are squeezing.

The humanoid robot industry has spent five years and billions of dollars solving locomotion, balance, and AI perception. It has made extraordinary progress. Robots can walk, run, climb stairs, navigate warehouses, and identify objects with near-human accuracy. But the industry has barely begun to solve the problem that matters most for the vast majority of real-world tasks: touch.

Why touch is harder than walking

To understand why tactile sensing is the bottleneck, it helps to understand why locomotion was (relatively) easier to solve.

Walking and balancing are primarily problems of control theory and mechanical design. The physics are well understood. The number of variables is manageable. A bipedal robot has 6-12 joints in its legs, each with well-defined range of motion and load profiles. The ground is rigid. Gravity is constant. The feedback signals, primarily from IMUs and joint encoders, are clean and unambiguous. The control problem is hard, but it is the kind of hard that engineers have been solving for decades in related domains (aerospace, automotive, industrial automation).

Touch is a fundamentally different kind of problem. When a robot hand contacts an object, the number of relevant variables explodes. The object could be rigid or deformable. Smooth or textured. Dry or wet. Heavy or light. Fragile or robust. The contact point could be a fingertip, a knuckle, the palm, or the side of the hand. The grip could involve two fingers, three fingers, four fingers, or a whole-hand envelop. The required force depends not just on weight but on friction, surface compliance, center of mass, and the specific task being performed (holding, turning, squeezing, pressing, pulling).

And all of this information needs to be captured in real time by sensors that are small enough to fit on a fingertip, robust enough to survive thousands of contacts per day, sensitive enough to detect forces below 0.1 newtons, and cheap enough to put on a $20,000 robot.

What a robot hand actually feels today

The state of tactile sensing in commercial humanoid robots is, to put it plainly, primitive.

Most humanoid robots on the market have force-torque sensors in their wrists and binary contact sensors on their fingertips. A wrist force-torque sensor tells you the total force being applied by the hand as a whole. It cannot tell you where on the hand that force is applied, or whether the contact is at the fingertip or the palm, or whether the object is slipping. It is the equivalent of knowing that your hand is pushing against something, without knowing which fingers are touching what.

Binary contact sensors tell you whether a surface is being touched or not. On or off. They cannot measure how much force is being applied, or the direction of the force, or the texture of the surface. A robot with binary fingertip sensors knows its finger is touching something. It does not know if that something is a steel bolt or a ripe strawberry.

Compare this to the human hand. The human fingertip alone has four types of mechanoreceptors operating at different spatial and temporal resolutions. Merkel cells detect sustained pressure with spatial resolution below 1 millimeter. Meissner’s corpuscles detect light touch and slip at frequencies up to 50 Hz. Pacinian corpuscles detect vibration at frequencies up to 400 Hz, which is how you can feel texture by sliding your finger across a surface. Ruffini endings detect skin stretch, which tells you the shape of what you are holding and whether it is rotating in your grip. All four operate simultaneously, and the brain fuses their signals into a unified percept of what the hand is doing.

No commercial tactile sensor reproduces even one of these capabilities at the resolution and speed of the biological original. No robot hand has all four.

The five technologies competing to give robots touch

The research community and a handful of companies are pursuing five distinct approaches to tactile sensing. Each has different strengths and limitations, and the winning approach for humanoid robots is not yet clear.

1. Camera-based tactile sensors (GelSight and variants)

The most successful laboratory tactile sensor is GelSight, originally developed at MIT CSAIL. The principle is simple and clever: a soft elastomer pad is illuminated from inside by LEDs. When an object presses into the pad, the elastomer deforms. A small camera inside the sensor captures images of the deformed surface. Computer vision algorithms then reconstruct the 3D shape of the contact area, the applied force distribution, and even the texture of the object’s surface.

GelSight achieves remarkable resolution: sub-millimeter spatial resolution and force sensitivity below 0.05 newtons. It can detect fingerprint-level textures and measure shear forces that indicate slip. In laboratory settings, robots equipped with GelSight have successfully performed tasks that no other sensor technology can support, including picking up thin, flexible objects like fabric, detecting the ripeness of fruit by measuring surface compliance, and identifying objects purely by touch.

The problem is engineering. Camera-based sensors require a camera, an LED array, a flexible elastomer, and enough computational power to process video in real time, all packed into a fingertip-sized form factor. The sensors are bulky by fingertip standards (typically 2-3 cm across), fragile (the elastomer wears out after thousands of contacts), and expensive ($200-$500 per sensor in small production runs). Putting ten of them on a humanoid robot hand would add $2,000-$5,000 to the BOM, increase the hand’s weight, and require significant computational resources.

2. Resistive/piezoresistive sensor arrays

The most commercially available tactile sensors use piezoresistive materials: substances whose electrical resistance changes when pressure is applied. These sensors can be manufactured as thin, flexible films that conform to curved surfaces like fingertips. They are cheap relative to camera-based sensors, with costs as low as $10-$50 per sensor array in volume production.

The trade-off is resolution and dynamic range. Piezoresistive sensors work well for measuring normal force (how hard you are pressing straight down) but poorly for shear force (how hard you are pressing sideways, which is what tells you if an object is slipping). Their spatial resolution is limited by electrode density, typically 3-5 mm between sensing points. And they drift over time: the relationship between pressure and resistance changes as the material fatigues, meaning the sensor needs periodic recalibration.

Most commercial humanoid robots that have any tactile sensing at all use some variant of piezoresistive arrays. They provide enough feedback for coarse grasping, picking up rigid objects with known properties, but not enough for fine manipulation.

3. Capacitive sensor skins

Capacitive sensors measure changes in electrical capacitance caused by mechanical deformation. Two conductive layers separated by a flexible dielectric create a capacitor. When pressure compresses the dielectric, capacitance changes. By arranging many such capacitors in a grid, you can create a sensor skin that maps pressure across a surface.

The advantage of capacitive sensing is that it scales well. Researchers at Columbia University and several European labs have demonstrated flexible capacitive skins covering entire robot arms and hands, with hundreds or thousands of sensing points. The technology is compatible with standard printed circuit board manufacturing, which means it could potentially be produced at consumer electronics costs if demand reaches scale.

The disadvantages are sensitivity to electromagnetic interference (a robot arm full of electric motors is a noisy electromagnetic environment), limited dynamic range (they work better for light touch than heavy gripping), and temperature sensitivity.

4. Optical waveguide sensors

A newer approach uses flexible optical fibers embedded in silicone skin. Light propagates through the fibers. When the skin deforms, the fibers bend, changing the pattern of light transmission. By measuring these changes with photodetectors, the system can infer the magnitude, location, and direction of applied forces.

Optical waveguide sensors have two major advantages: they are immune to electromagnetic interference (no electrical signals in the sensing layer), and they can achieve very high spatial density because optical fibers can be packed tightly. Researchers have demonstrated prototypes with spatial resolution approaching GelSight levels but in a flatter, more conformal form factor.

The technology is still pre-commercial. Manufacturing flexible optical waveguide arrays at scale, ensuring their durability over millions of contact cycles, and integrating them with robot control systems are open engineering challenges.

5. MEMS-based tactile arrays

Micro-electromechanical systems (MEMS) technology, which revolutionized inertial measurement and enabled the IMUs in every smartphone, is being adapted for tactile sensing. MEMS tactile sensors use microscale mechanical structures that deflect under pressure, with the deflection measured by piezoresistive, capacitive, or piezoelectric transducers.

The appeal of MEMS is manufacturing maturity. Semiconductor fabs already produce billions of MEMS devices per year. If tactile sensor designs can be adapted to standard MEMS fabrication processes, the cost per sensor could drop to dollars or even cents in volume, making dense tactile arrays economically viable on a $20,000 robot.

The challenge is that MEMS sensors are rigid. They do not conform to curved surfaces easily. And they are fragile in ways that do not interact well with the impact forces a robot hand experiences. A MEMS accelerometer inside a phone is protected by packaging. A MEMS tactile sensor on a fingertip gets hit directly every time the robot grips an object.

What it costs to give a robot real touch

The economics of tactile sensing are the main reason it has been neglected. Let us walk through the numbers.

A human-equivalent tactile system for a robot hand would require approximately 17,000 sensing points per hand (matching the mechanoreceptor density of human skin), operating at update rates of 100 Hz or faster, with sub-millimeter spatial resolution, sub-newton force sensitivity, and shear/slip detection capability.

At current research-grade sensor costs of $0.40-$1.20 per square millimeter of sensing area, a full-coverage tactile skin for both hands would cost $8,000-$24,000 in sensors alone. For a $100,000 enterprise humanoid, that is a 10-25% BOM increase dedicated to a single subsystem. For a $16,000 Unitree G1, it would more than double the price of the robot.

Morgan Stanley estimated in its 2025 BOM analysis that dexterous hands already account for 5-10% of a humanoid robot’s total cost, or $2,000-$8,000 per robot. Adding research-grade tactile sensing would push hands to 15-30% of BOM, making them the most expensive subsystem on the robot, more expensive than all the actuators in both legs combined.

This is why every commercial humanoid robot on the market has chosen the same compromise: skip real tactile sensing and rely on force-torque sensors in the wrist, binary contact sensors in the fingertips, and vision-based grasp planning to compensate for what the hands cannot feel.

The approach works, barely, for structured industrial tasks where the objects are known in advance, the weights are predictable, and the grip strategy can be pre-programmed. It fails completely for unstructured tasks like loading a dishwasher, folding laundry, handling groceries, assembling small components, or any situation where the robot encounters an object it has not been specifically trained to handle.

Why simulation cannot fix this

The humanoid robot industry has used simulation to make enormous progress in locomotion and visual perception. Train a walking policy in NVIDIA Isaac Sim, transfer it to real hardware, and it works. The sim-to-real gap for locomotion has shrunk to the point where simulation is the primary training environment.

Tactile sensing is different. The sim-to-real gap for contact physics is enormous and not closing quickly.

Simulating rigid body contact (a robot foot hitting a hard floor) is well understood. Simulating deformable contact (a robot finger pressing into a soft object) is an open research problem. The physics of how skin deforms around a grape, how paper crumples under finger pressure, how a wet glass behaves differently from a dry one: these phenomena involve material properties, fluid dynamics, and surface chemistry that are computationally expensive to simulate and often poorly characterized.

NVIDIA’s Newton physics engine, co-developed with Google DeepMind and Disney Research, represents the state of the art in robotics simulation. It handles rigid body contact and articulated joint dynamics beautifully. It does not handle deformable object manipulation at the resolution needed to train tactile policies. When Figure AI or Boston Dynamics trains a grasping policy in simulation, the simulated hand grabs a rigid approximation of the object. The real hand then encounters a soft, textured, temperature-varying, sometimes sticky, sometimes slippery version of that object. The gap matters.

This is why real-world tactile data is especially valuable, and especially hard to collect. You cannot simulate your way to a good tactile policy. You need real robots touching real objects, with real sensors providing real feedback. And since the sensors themselves are immature, the data you collect is noisy, sparse, and hard to generalize from.

Who is closest to solving it

Despite the difficulty, several companies and research labs are making meaningful progress on robotic touch. The approaches diverge significantly, and it is not clear which path will reach commercial viability first.

Sanctuary AI has arguably the most serious commitment to tactile intelligence in the commercial humanoid robot industry. The company’s Phoenix robot was designed from the beginning around dexterous manipulation, and its Carbon AI system explicitly includes tactile reasoning as a core capability alongside visual and language understanding. Sanctuary claims that Phoenix’s hands have enough tactile feedback to perform tasks like retail shelf stocking without prior training on specific objects. Whether this claim holds up at scale remains to be proven, but Sanctuary is the only humanoid robot company that has made tactile capability the centerpiece of its pitch rather than an afterthought.

Tesla has taken an engineering-pragmatist approach. The Optimus Gen 3 hand has 22 degrees of freedom, which is among the highest in the industry, and includes capacitive tactile sensors on the fingertips. Tesla has not published detailed specifications on sensor resolution or sensitivity, but demos suggest the hands can perform moderately delicate tasks like picking up eggs (in controlled settings, with careful tuning). Tesla’s advantage is its willingness to iterate aggressively on hardware. The Gen 1 Optimus hand was essentially a crude gripper. The Gen 3 hand, released roughly two years later, is a recognizable mechanical analog of a human hand. If the improvement rate continues, Gen 5 hands could have meaningfully better tactile feedback.

Boston Dynamics benefits from the most sophisticated force-torque sensing in the industry. Atlas’s wrist-mounted force-torque sensors are multi-axis, high-bandwidth, and tightly integrated with the control system. This does not substitute for fingertip tactile sensing, but it allows Atlas to perform force-controlled tasks like inserting parts and handling heavy objects with precision that other robots cannot match. Boston Dynamics’ partnership with Google DeepMind could accelerate progress, as DeepMind has published research on learning tactile manipulation policies from simulation and real-world data.

The research labs remain ahead of commercial implementations. MIT’s GelSight group continues to lead in camera-based tactile sensing. Stanford’s Robotic Manipulation Lab has demonstrated learning-based approaches where robots develop tactile skills through trial-and-error interaction with objects. Columbia University’s electronic skin work is pushing toward full-body tactile coverage. And Meta’s FAIR lab has published work on self-supervised tactile representation learning, where a robot learns a general understanding of touch without being told what specific objects feel like.

The tasks locked behind the sensory cliff

The practical consequence of the tactile sensing gap is that entire categories of real-world tasks remain off-limits to humanoid robots. Understanding which tasks are blocked, and why, reveals the true economic cost of the sensory cliff.

Household tasks. Cooking, cleaning, laundry, and dishwashing all require handling diverse objects with varying fragility, texture, and compliance. This is why the $38 billion home robot market that Goldman Sachs projects for 2035 is contingent on tactile breakthroughs. A robot that can walk through your house but cannot fold your shirt or load your dishwasher is a very expensive novelty.

Electronics assembly. Printed circuit board assembly, cable routing, and connector insertion require force precision below 1 newton and positional accuracy below 0.5 millimeters. Current humanoid robot hands are roughly 10x too imprecise for these tasks. This is why electronics manufacturing remains dominated by specialized pick-and-place machines rather than general-purpose robots.

Healthcare and eldercare. Assisting patients with dressing, bathing, and mobility requires gentle, responsive touch. The consequences of gripping too hard are injury, not product damage. The liability implications make tactile reliability a hard requirement, not a nice-to-have.

Agriculture. Fruit picking, particularly for soft fruits like berries and stone fruits, requires detecting ripeness through surface compliance (how much the fruit gives when gently squeezed). Machines that can pick apples exist. Machines that can reliably pick strawberries without damaging them do not.

Retail and food service. Handling bread, pastries, fresh produce, and other deformable food items requires the kind of adaptive grip force that only tactile feedback can provide. A warehouse robot that moves sealed boxes does not need to feel. A grocery robot that stocks produce shelves does.

Add them up and the sensory cliff is blocking $31-65 billion per year in potential humanoid robot market value by 2030. That is 3-6 times larger than the current total humanoid robot market. The industry is building the legs for a multi-trillion-dollar market while ignoring the hands.

The path forward: what needs to happen

Solving the tactile sensing problem requires breakthroughs on three fronts simultaneously. No single advance will be sufficient.

Sensor hardware must get 20-40x cheaper

At current research-grade costs, full-coverage tactile sensing is economically impossible on a commercial robot. The cost needs to drop from $0.40-$1.20 per square millimeter to $0.01-$0.05 per square millimeter. That is a 20-40x reduction. For context, MEMS accelerometers achieved a similar cost reduction (from $5 per unit to $0.20 per unit) over roughly 15 years as smartphone demand drove volume production. Tactile sensors do not yet have an equivalent volume market to drive costs down, which is a chicken-and-egg problem: robots need cheap sensors, but cheap sensors need the demand that only mass-market robots can provide.

Contact physics simulation must improve dramatically

Until simulation can accurately model deformable contact, soft body dynamics, and surface friction at the resolution needed for manipulation training, tactile AI will have to be trained primarily on real-world data. That is too slow and too expensive to scale. The physics engines need to get better at modeling what happens when a rubber fingertip presses into a tomato. This is a fundamental physics and computation problem, not an engineering problem. It may require entirely new simulation approaches, not incremental improvements to existing engines.

Tactile AI models need their own foundation model

Vision has foundation models (CLIP, DINOv2). Language has foundation models (GPT, Claude). Robotics control has emerging foundation models (GR00T N1, Pi-0). Touch has nothing. There is no large-scale, pre-trained model that understands tactile signals in a general way. Every robot that learns to manipulate objects today starts from scratch. A tactile foundation model, trained on diverse tactile data from many sensors and many objects, would allow new robots to start with a baseline understanding of touch, just as vision foundation models give robots a baseline understanding of what they see.

The inconvenient truth about AI hype

There is a narrative in the humanoid robot industry that goes roughly like this: AI will solve everything. Train a big enough model on enough data, and the robot will learn to manipulate any object without needing dense tactile feedback. Vision is enough. Foundation models will figure out the rest.

This narrative is convenient for companies that have invested billions in AI and relatively little in sensing hardware. It is also, based on the evidence, wrong.

Research from MIT, Stanford, and Carnegie Mellon consistently shows that tactile feedback improves manipulation success rates by 30-60% compared to vision-only approaches, even when the vision system is state-of-the-art. The improvement is largest for novel objects (things the robot has not seen before), deformable objects (things that change shape when gripped), and occluded scenarios (when the robot’s fingers block the camera’s view of the object, which happens on literally every grasp).

Vision can tell you what an object looks like. Only touch can tell you what it feels like. And for manipulation, what it feels like, its weight, its compliance, its surface friction, whether it is slipping, is often more important than what it looks like.

The $65 billion blind spot

The humanoid robot industry is in a peculiar position. It has solved, or is close to solving, problems that were considered impossible a decade ago. Bipedal locomotion over rough terrain. Real-time visual understanding of complex environments. Natural language interaction. Autonomous navigation in dynamic spaces. These are genuine achievements.

But the industry has collectively looked away from the problem that most constrains its addressable market. The sensory cliff, the gap between what a robot can move and what it can feel, is not a theoretical limitation. It is the reason that the vast majority of tasks that humans perform with their hands remain beyond the reach of the most advanced humanoid robots in the world.

Walking was the hard problem of 2015. AI perception was the hard problem of 2020. Touch is the hard problem of 2026. And unlike walking and perception, touch does not have a clear path to solution. The sensors are too expensive. The simulation is too primitive. The AI models are nonexistent. The research community is fragmented across five competing approaches with no convergence in sight.

The company that solves robotic touch, that gives a machine the ability to feel a grape between its fingers and know exactly how hard to squeeze, will unlock more economic value than any locomotion breakthrough or AI model ever will. That company has not yet emerged. But every quarter that passes without progress is a quarter that $31-65 billion in potential market value stays locked behind the sensory cliff.

The robots can walk. The robots can see. The robots can lift 50 kilograms. But until they can feel, they cannot do the work that actually matters.