English 
Grasping a glass would seem like a straightforward task. A robot could detect it, compute its location, and maneuver its arm in its direction. However, the real challenge comes after the fingers of the robot come in contact with the object. What amount of pressure must it exert? What if it starts slipping from the fingers of the robot? Would it crush the glass while applying excessive force?
The above questions are those that humans take care of in most cases on a subconscious level. Robots have to take it upon their sensors, computing and system design to overcome such challenges. This is where robot skin and tactile AI come to rescue. The concept entails equipping robots with an ability to understand physical interaction and respond to it. As more robots find their application in warehouses, hospitals, and industries, this skill could prove very helpful.
What Is Robot Skin and Tactile AI?
Tactile AI in robotics uses touch-sensing hardware and artificial intelligence to help robots interpret physical contact and respond accordingly. The hardware collects information when a robot touches something. The AI processes that information and helps the machine decide what to do next.
Robot skin, also called electronic skin or e-skin, is the physical part of the system. It uses flexible materials with sensors built into them. Depending on the design, these sensors can detect pressure, temperature, bending and certain surface details. Some systems can cover a larger part of a robot’s body instead of limiting touch detection to its fingertips.
The software side has a different job. Imagine a robot holding an egg. Its sensors may register pressure, but that reading alone does not tell the machine whether its grip is safe. It needs to interpret the pressure, understand the task and adjust its movement if necessary.
That is where tactile AI becomes useful.
Robots in traditional robotics systems perform admirably well in situations where the environment and conditions are fairly predictable. Place a robot in an unpredictable environment, and the difficulty of the task rises. Robots cannot use just the programmed motions in order to cope with such situations.
The robot skin, together with tactile AI, can be used to solve this problem. The combination introduces an additional perception capability to the field of Physical AI, which involves interaction between a machine and the physical world. The main task is to make the sensors and the AI work in unison in order to elicit the proper response.
How Does a Machine Feel?
Robots don’t have feelings the way humans have feelings. Robots perceive physical changes through measurement and use these measurements to interpret their surroundings. There are a number of approaches by which different types of sensors accomplish this task.
Piezoresistive sensors use measurement of electrical resistance change under the influence of the deformation of a particular material, whereas capacitive sensors rely on detection of the change in electrical capacitance due to the change in distance between two conductive surfaces. Both sensors may be used to estimate the amount of force applied to the object by a robot.
Optical tactile sensors operate in a completely different way. To illustrate, let us consider the case of GelSight tactile sensors. This gives the robot information that an ordinary camera may struggle to capture, particularly when the contact point is hidden beneath its fingers.
Collecting the data is only half the job, though. The robot must process the readings and respond quickly enough to make a difference. If an object begins slipping, a delayed response could mean losing it altogether.
Think of this as a machine’s version of a feedback system. The sensor detects a change, the software interprets it and the robot adjusts its grip or movement. The process sounds straightforward, but getting it to work reliably across different materials, objects and environments takes considerable engineering.
Why Vision Isn’t Enough for Physical AI
Cameras tell robots a lot about the world. They can identify objects, estimate distances and track movement. But they have a blind spot. Once a robot grips an object, its fingers may cover the very area where contact is happening. Lighting and reflections can create further problems.
Touch provides information that vision cannot always supply. A tactile sensor may detect pressure changes or sideways forces that suggest an object is slipping. The robot can then tighten its grip slightly or reposition its fingers. Without this feedback, it risks dropping the object or applying too much force.
Google’s Gemini Robotics ER 2 provides an example of how AI can interpret visual information during a physical task. Google achieves an accuracy of 91.3% and a mean absolute distance of 0.96 seconds in moment-finding evaluation. The algorithm is able to track progress in video and make changes if anything happens wrong along the way.
The above numbers refer to visual reasoning and not tactile perception. Nonetheless, it highlights the issue that is relevant to robotics at large. The machine should be able to detect any changes in the situation and act accordingly. Tactile AI provides another source of data, which comes from physical interaction.
Real-World Applications Transforming Industries
Such tasks seem like obvious applications for the technology. Robots can be used for moving goods, sorting packages and manipulating items. However, the items cannot always be identical. Some are breakable, some have odd forms, some may change their shape during manipulation. Touch would be able to provide robots with more information about how to grip the item.
The scale of robotic handling is already significant. The International Federation of Robotics reported 117,500 professional service robots sold for transportation and logistics in 2025, an increase of 21%. This does not mean all these robots use tactile sensors. It does show why better physical handling could matter as robotic systems take on more logistics work.
There are other problems in healthcare. In prosthetics, researches are developing methods of restoring sensory information to the user of an artificial limb. It means providing the user with some natural sense of contact and grip rather than relying on vision only. Force feedback can also prove useful for surgical robots when dealing with fragile tissues. However, detecting pressure will not provide surgeons with all information about the health state of the tissue.
The last problem is connected with humanoid robots. These machines are supposed to operate in human environment and interact with people. It means that a robot should be able to react appropriately in case of physical contact with the surroundings.
The second edition of ISO 13482, developed by ISO, entered the final-draft approval stage on September 15, 2026. The document concerns safety of service robots with physical interaction with people. It remains under development. It does not require robots to have artificial skin, but it highlights the importance of managing risks when machines operate around people.
IFR also reported around 7,000 full-size humanoid robots sold globally for commercial and professional applications in 2025. Many applications remain specialized and often require human teleoperation. That tells us something important. Building a robot that looks capable is one thing. Getting it to perform reliably without constant human help is another.
Challenges and the Future of Tactile AI
ロボット skin still has some practical problems to solve. Flexible materials must survive repeated contact, stretching, friction and accidental impacts. Sensors can lose accuracy with wear, while covering larger surfaces creates additional wiring and maintenance demands. Replacing damaged skin could also become expensive.
Then there is the amount of data involved. A sensor-equipped robot will be able to produce a constant flow of data from all its sensors. The robot’s systems should be able to process important signals fast without any delay in movement. Additional sensors won’t bring much benefit if the robot will not be able to interpret information from them.
The second question is connected with training. It is possible to train robots to perform tasks using simulations but the reality is not always the same as the simulation. Materials can have different surface and texture and the friction conditions can change. Thus, a model that has been successfully trained in a certain environment will have difficulties in a new situation. That is how the sim-to-real problem arises in tactile AI as well.
The solution may be related to the concept of multimodal AI. The robot can locate objects via vision, understand the job through verbal instructions and perceive touch sensations regarding a grip. Vision-language-action can be used here.
However, to combine these two systems, it requires more than just adding sensors to a robust architecture. The robot needs to make sense of contradictory inputs, react promptly and protect itself from any dangerous actions if the predictions have proven to be inaccurate. This is the challenging part.
結論
こちらもお読みください： フジソフト、Anthropicと提携し、AWS上で「Claude」を提供する
Robot skin and tactile AI could broaden what a machine is able to do, especially if it has to handle fragile things or unpredictable interactions. But the crucial bit isn’t whether a robot can feel the pressure. It’s whether it can interpret that pressure – what it means – and act accordingly. The tech still struggles with ruggedness, processing, and how it works in the real world.
That’s more significant than the astonishing demos, because robots are going to have to interact with regular people and the stuff they use, and making mistakes there will have bigger repercussions.
Giving machines a sense of touch will not make them human. It may, however, make them less clumsy, more dependable and better suited to the physical world they are being built to enter.
日本語