The robot reads the subtle changes in tactile sensation transmitted to the fingers gripping an object, then reapplies force just before it slips. This was a scene actually reproduced on the main stage of Swiss Robotics Day 2026, held on the 13th (local time) in Zurich, Switzerland. At this event, where roughly 170 robots—including humanoids, robotic arms, and quadruped robots—gathered, what participating companies rushed to present was not walking but “manipulation.” As humanoid walking and balance performance have improved rapidly in recent years, hand dexterity has emerged as the next hurdle blocking deployment on industrial floors.
The most striking element was the precision of tactile sensors. Swiss startup Inveel unveiled a robot skin 500 micrometers (㎛) thick. A single thin sensor detects not only pressure and contact but also the approach of an object before it touches. It recognizes objects up to 15 centimeters away and distinguishes forces down to the 0.0005N (newton) level. This means the robot can read, as numerical data, minute forces that are difficult for humans to feel.
Wang Xingxing, founder and CEO of China’s Unitree, presented a new robot hand roughly the size of a human hand with more than 20 degrees of freedom. Degrees of freedom refer to the number of directions in which joints can move independently; the more joints move in each finger, the better suited the hand is for precise manipulation such as gripping objects of different shapes or changing posture. CEO Wang explained that the drive structure between the fingers and the hand is designed to absorb large impacts.
Professor Aude Billard, head of the EPFL Robot Center, cited dexterity as the key obstacle to actual deployment. Her research team has spent the past five years studying where people place their hands and how much pressure they apply when learning precision tasks such as watchmaking. She explained that understanding the very process by which humans acquire fine manual skills and transferring it to robots is more important than simply collecting large amounts of motion data.
Factories Without Public Data: The Real Test
It was Hexagon Robotics that delivered the most realistic warning at the venue. Enno de Lange, Director of Physical AI and Robotics, displayed photos of bananas, kitchens, and Swiss mountain roads in turn, then said, “Our robots will never see things like these.” This is because what the humanoid “AEON” handles in the factories where it is deployed is not bananas—which are overflowing with photos on the internet—but metal parts with gleaming surfaces or small industrial components with almost no distinguishing features. In many cases, the processes themselves are confidential, so there is no public training data at all.
The BMW project case illustrates a concrete benchmark. The time required for AEON to take a battery cell out of a tray and place it into assembly equipment is about 30 seconds per cycle. It must repeat the same motion at a set speed all day long without dropping the part. Director de Lange noted, “Demos are easy. You can prepare for a long time and just try again if you fail,” adding, “Many people underestimate the gap between showing something once or a few times and showing it reliably hundreds of thousands of times.”
How to Teach Experience
Another axis of manipulation capability is data and reinforcement learning. Rory Sexton, a director at Agile Robots, estimated that a person accumulates about 24,000 hours of experience perceiving space and controlling the body over a lifetime. The idea is that people can open doors and find cups to make tea even in a house they are visiting for the first time because this experience has already accumulated. “This is not a hardware problem, nor a computer problem,” he said. “It’s a problem of experience—that is, data.” The company runs a “robot school” with humans as teachers and robots as students, combining force and image data obtained through teleoperation, videos of human work, simulation, and internet materials for training.
Swiss startup Flexion took a different direction. David Höller, Chief Technology Officer (CTO) and co-founder, explained that rather than focusing on pre-training on millions of hours of human behavior data, the company emphasizes post-training using simulation and reinforcement learning. The software consists of three layers. At the top, a vision-language model (VLM) interprets the scene and commands to decide what to do; the next stage calculates movement paths for the limbs; and finally, it controls the motors to convert this into whole-body motion. The company aims to apply this entire software stack not only to specific humanoids but also to other forms such as wheeled robots and robotic arms.
Component Competition: From Sensing to Actuation
Even after learning is complete, sensors, joints, and drive systems must support the conversion of that learning into actual movement. The exhibition hall featured a large number of companies that build robot bodies.
CompanyKey Products/TechnologiesCore SpecificationsInveelRobot skin (tactile sensor)500㎛ thick, detects approach up to 15㎝, detects 0.0005N forceUnitreeNew robot handHuman-hand size, more than 20 degrees of freedomBota SystemsForce/torque sensors, TACTO platformIntegrates force sensors, tactile feedback, cameras, and robot statusSoncebozActuators for robot handsUnder 50g in weight, up to 1.5Nm torquemaxonIntegrated robot joints6-axis demo with integrated motor, encoder, reducer, brake, and controllerDuaticHigh-power compact actuators, lightweight robotic armsHumanoid DXTR carries logistics boxes up to 25㎏TACTILIA RoboticsRobot hands SDH, SVH, SIH3-finger, 5-finger, and tendon-driven types, combined with grasp-learning technology
Bota Systems presented the “TACTO” platform, which collects force sensors, tactile feedback, camera images, and robot status all at once and uses the resulting physical interaction data for AI training. Sonceboz unveiled an actuator for robot hands that weighs under 50 grams while producing up to 1.5Nm of torque. The key is to generate sufficient force in narrow spaces like fingers while reducing weight and heat. The company is integrating actuators, reducers, thermal management devices, and control electronics into finger-level modules.
maxon exhibited a 6-axis robot demo composed of robot joints that bundle a motor, encoder, reducer, brake, and controller into one unit. Duatic, which originated from research at ETH Zurich, develops high-power compact actuators and lightweight robotic arms, and introduced the humanoid “DXTR,” which lifts and moves logistics boxes weighing up to 25 kilograms. TACTILIA Robotics exhibited the three-finger “SDH,” the five-finger “SVH,” and the tendon-driven “SIH,” and supplies not just hardware but also grasp-learning technology as a bundled system.
South Korean Manufacturers Seek Component Collaboration Points
Officials from South Korean manufacturing also attended the event. An official from Daesung Hi-Tech, an ultra-precision machining and precision components company, cited the research foundation from which original technologies emerge as the difference between Switzerland and South Korea. Switzerland’s strength, the official said, is an environment in which universities, research institutes, and innovation centers are clustered closely together and government support is linked with industry investment, allowing R&D to continue. “Many future-oriented technologies emerge, but what matters is how they are implemented and who implements them,” the official said. “Robotics is ultimately an industry, and manufacturing technology is one of the elements that turns it into reality.” The view is that since South Korea has strengths in manufacturing technology, it can expand cooperation with Switzerland in this field.
An official from Hanjin Industrial, a manufacturer of wind power and ship engine components, focused on actuators and drives. “Actuators can be called the heart of robots and account for a large share of the cost,” the official said. “The more high-value-added a component is, the more difficult it is to make, and finding a company that can meet both requirements and price is another matter.”
The implications of this event are clear. The competitive axis for humanoids is shifting from “the ability to walk” to “the ability to precisely grasp and repeat.” This requires, at the same time, components such as tactile sensors that distinguish 0.0005N, hands with more than 20 degrees of freedom, and actuators that deliver 1.5Nm at under 50 grams—as well as learning software that can withstand hundreds of thousands of repetitions in environments with no public data. The assessment emerging is that whoever possesses hardware, software, and data together is increasingly likely to seize leadership in the industrial robot market.