The New Standard for Humanoid Robots | Best Tech at ICRA 2026

[00:00:00]The International Conference on Robotics and Automation IC 2026 just took place in Austria and surprisingly American companies were almost completely absent this time. Instead, Chinese engineers brought so much hardware that China's leadership in this field is becoming more and more obvious. Today, we'll look at and break down the most revolutionary technologies. From Nvidia neural networks that understand the world almost like chat GPT to robots that can twist balloon animals, all-weather allterrain humanoids, Vietnamese androids, and shape-shifting robot dogs.

[00:00:34]Overall, there were many systems for training robots at IC, and most of them were brought by Chinese companies. For example, Gento Robotics set up a full-scale lab at their booth and showed robots in action non-stop for all five days of the expo. And this was the first public presentation of their entire tech suite. From the robots to the software they run on, virtual simulation systems, and their proprietary remote control system called Apex. Visitors were entertained by two robot models, Gento Luna and Gento Sky. They danced with people, hugged visitors, shook hands, and quickly and accurately mirrored an operator's movements. The company has a new product, the Marvin robot, which they didn't bring to Vienna, but officially presented at the conference.

[00:01:14]While Luna and Sky are basic robots for general tasks and demonstrations, Marvin is built for ultrarecise, delicate work in the real world. Simply put, if a regular robot might accidentally crush an egg or a glass, Marvin feels objects almost like a human. The robot has a seven axis mobile wrist, meaning its joints bend in a more complex, freer way than a human hand, approaching an object from virtually any angle. It also has extreme sensitivity, powered by micro electronic pressure sensors. On top of that, the robot has a 360 degree view.

[00:01:46]Gento Robotics was founded by a significant figure in the industry. Dr. Hanwin Kong spent a long time leading development at a secret Huawei lab, focusing specifically on how robots see and navigate through space. At the conference, he presented a paper on how the company combined camera vision and Marvin's tactile sensors so the robot could pick up an unfamiliar, fragile, or soft object from a table on the first try without breaking it.

[00:02:12]A robot at the Agilink booth showed something similar. It was twisting balloon animals slowly but delicately.

[00:02:18]Not a single balloon was harmed. And it was the best demonstration of dexterity and grip force control at IC this year.

[00:02:25]Imagine a balloon is light, slippery, constantly changing shape, and the pressure inside keeps changing. If the robot squeezes it even a tiny bit too hard with its fingers, pop, the balloon bursts. If it doesn't squeeze hard enough, it slips out and the whole structure falls apart. A human does this intuitively. We feel the tension in the rubber and instantly adjust our grip force. But for a robot, until now, this was almost impossible. It's not enough for it to just move its fingers along coordinates. It needs to feel the contact. To achieve this result, engineers split the robot's intelligence into one part responsible for movement and another that controls contact with the object. The first was trained on a magician's movements and then trained further like this. Every time the balloon started to slip from the robot's grip, an operator took over control, and the neural network then analyzed those exact moments. That's where the second AI model came in. It learned to understand the physics of contact right at the moment of touch, where the friction is, how the balloon has deformed, and how much pressure to apply to stay in that narrow zone between the balloon slipped out and the balloon exploded. Also, right there in Vienna, Agilink rolled out its latest tech product, the Omnihand 3 UltraM mechanical hand. It's exactly the size of an adult human hand and has 20 active degrees of freedom. Basically, every joint in every mechanical finger bends independently, almost completely replicating the motor skills of our hand. But that's not all. The hand has two more key features. First is direct drive. This means there are no complex belts or gears inside. The motors are placed right in the joints, so the fingers respond to commands instantly.

[00:03:57]Second is the fantastic skin. The entire palm and fingers are covered with tactile sensors. There are micro cameras in the fingertips that see the skin deforming from the inside and more than 300 pressure points are scattered across the palm. A hand like this is capable of feeling even a feather. This is exactly what you need for a home robot or a robot in micro electronics assembly. And in China, solutions like this are popping up like mushrooms after rain.

[00:04:25]By the way, Agilink is part of the Agabot group of companies which hosted a large-scale embodied AI competition right at the conference, the Agabot World Challenge. The idea sounds simple.

[00:04:36]Combine the brain and body of a robot into one system. So, the robot not only sees the world, but understands it, makes decisions, and immediately turns them into actions. To do this, the competition was divided into two tracks.

[00:04:49]The first is reasoning to action. Here, robots were tested on whether they could transfer skills from a virtual simulator into real life. The second is world modeling. Here they evaluated how well the AI understands physics and whether it can predict what will happen to an object in the next second. In the qualifying stage of the challenge, 526 teams from 27 countries took part. For embodied AI, this is already on the level of the largest international competitions in artificial intelligence.

[00:05:18]Agibbot is now considered one of the fastest growing humanoid manufacturers in the world. The company became the global leader in Android shipments in 2025 and is actively developing its own Agibot world ecosystem which includes robots, simulators, artificial intelligence models, and one of the industry's largest data sets for training robots.

[00:05:39]But it wasn't just the Chinese who came to conquer IC. The South Korean startup to Solo brought its humanoid, but its real product is a robot hand. Right now, the global industry is running into a bottleneck with hands. Without them, it's simply impossible to integrate robots into real life. So, what makes this solution different? According to the developers, it primarily solves two common problems: affordability and low weight. The basic DG5 FM hand is the size of an adult human hand. It has 20 degrees of freedom and each joint is controlled independently by its own servo drive. But if a developer doesn't need that kind of complexity, the hand can be software limited to 15° of freedom. The fewer ways a robot can bend its joints, the fewer options the neural network needs to calculate and the faster it understands how to perform a simple task. Another distinctive feature is the weight. This mechanical hand weighs just 880 g, about 1.9 lb. Most counterparts weigh well over 1 kilogram, 2.2 lbs. And that difference is critical. The lighter the hand, the less battery power the robot's shoulder wastes just lifting it. Now, for the price tag, this hand costs around $23,000, which is a very competitive price for the industry. Plus, it has what's called back drivability, meaning the fingers gently spring back under external pressure. If the robot misses the table and bumps into it with its hand, the joints won't break. They simply absorb the impact.

[00:07:04]At the expo, Nvidia showed technologies without which modern robots simply wouldn't be able to understand commands or navigate through space. The key announcement was a new generation of AI for robots. Before, androids had to be trained separately to pick up a mug, open a door, or stack boxes. Now, Nvidia has shown a model that understands a task instantly through text, images, and the real environment around it. It sees an object, understands the command, and performs the action just like Chat GPT, but for the physical world. Another big announcement was the Grasp Gen X system.

[00:07:36]With it, robots automatically calculate how to pick up objects they're seeing for the first time. A new box shape, an unfamiliar part, a random object on a table. The system calculates how to grab it. It sounds simple, but for machines, it's a huge problem. And if that problem is solved, they will get one step closer to autonomy, a very big step. Nvidia published another interesting announcement after the expo. The company announced the construction of an AI factory in partnership with LG Group. As a reminder, at CES 2026, LG showed its home robot assistant. But for a home robot to avoid getting stuck in the very first doorway, and to tell a cat from a slipper, its neural networks need to be fed millions of hours of video and simulations. To process that volume, the two giants are creating an artificial intelligence factory. In the literal sense of the word, it will be a massive computing infrastructure where future robots, autonomous vehicles, and industrial systems will be trained. And that's not all. At the same time, Nvidia is creating the first reference humanoid robot for developers around the world.

[00:08:38]It's based on a Unitry humanoid and inside is a full suite of Nvidia technologies from onboard chips to virtual simulators and AI training software. In fact, the company is trying to do for robots what Android once did for smartphones. Not create one perfect robot, but create a standard that thousands of different models can be built on. And if this plan works, in a few years, Nvidia could become to robotics what it is to artificial intelligence today. A company whose technology almost no one can do without.

[00:09:06]And by the way, many robots at the traditional ICR 2026 parade run on Nvidia. This year, the conference brought together 7,000 leading engineers from companies, labs, and startups around the world. The main trend being debated in the halls was the end of the era of blind hardware. Engineers officially admitted it. Classical motion programming is dead, giving way to endtoend neural network training, where AI directly connects the camera image to the motors and programmers no longer need to write code for every single step. And the real stars of the event were cloud-based universal platforms.

[00:09:39]Not just Nvidia, but also for example, Oll Robotics, which lets you train robots in basic skills and behavioral physics right in the browser using ordinary text prompts to neural networks.

[00:09:51]But while we've seen robots from South Korea before, a humanoid from Vietnam was a real discovery, especially at such a major event. Vin Robotics is the robotics arm of Vietnam's largest conglomerate, Vin Group, and they brought their pride and joy to Austria, their third generation industrial humanoid, the VRH3. Moreover, the Vietnamese team emphasizes that all the mechanics, electronics, and software code were developed by their engineers from scratch. Unlike service robots for the home, the VR3 was initially built to work in factories, plants, and logistics warehouses. The robot is equipped with 31 highresolution servo drives. The drives are powerful. The robot is designed to dynamically carry loads weighing from 6 to 8 kg, 13 to 18 lb.

[00:10:37]Inside the chassis, there are two computers at once. One is responsible for maintaining balance and walking, while the second handles vision and object recognition. The robot's movements are trained using reinforcement learning algorithms.

[00:10:49]Essentially, this is a carrot and stick method where the neural network searches for successful movements on its own in pursuit of a virtual reward. As of today, the VRH3 can walk autonomously across different types of terrain.

[00:11:02]However, its arm movements are still controlled by an operator. And there's an interesting detail here. Usually, teleoperation requires an entire studio with external tracking cameras. The Vietnamese team removed all the clutter around. The operator just needs to put on a VR headset and a motion capture suit. The robot received the signal through the EtherCAT industrial communication protocol, which transmits data with zero latency and instantly mirrored the human's actions while navigating with its own cameras. This is a ready to go solution for working in hazardous areas, for example, in chemical plants. However, this robot is not suitable for delicate operations, unlike the next one.

[00:11:41]Paxini Tech is a company from China, but it has Japanese roots. It was founded in 2021 by alumni of the legendary Sugano Laboratory at Japan's Waseda University.

[00:11:51]This is exactly where Weebot 1, the world's first humanoid robot, was once created. After returning to Shenzen, Paxini's founders decided to focus on one thing, giving robots a sense of touch. At IC 2026 in Vienna, company representative Gial Lee presented a paper titled building a bridge between AI and the physical world. And for Paxini, this bridge consists of a unique artificial skin. They developed the PX6AX series of tactile sensors. At their core is clever physics, the so-called six-dimensional hall sensor matrix, a technology that reads the slightest changes in the magnetic field inside a finger upon contact. What does this mean in practice? Most robots only understand grip force, basically pressing hard or pressing softly. Paxini sensors capture micron level deformationations in every direction.

[00:12:42]The robot instantly understands the texture of an object, smooth or rough, its elasticity, soft or hard, and most importantly, friction and slippage. The robot feels if an object starts slipping from its fingers and squeezes tighter even before the item falls. Their flagship Dex H13 robotic hand features more than 1,100 of these micro receptors which output an insane 7,000 tactile signals. Essentially, this is an artificial nervous system that is almost caught up to human skin in terms of sensitivity.

[00:13:14]Want another impressive number? 9,000 km about 5,600 mi. That is exactly the distance from which the operator of the Chinese startup Astrobot was controlling a robot at the expo. The S1 Android is famous for arms that move at a speed of 10 meters/s, about 33 feet per second.

[00:13:32]At that speed, a human hand would just be a blur on camera. Yet, the robot maintains accuracy down to fractions of a millimeter. At the booth, it demonstrated wonders of smoothness, pouring drinks, organizing things on shelves, and even practicing calligraphy with accuracy down to 1/10enth of a millimeter, about 0.004 in. And all of this was controlled directly from the company's office in Shenzen using a VR headset and a pair of controllers. To make this possible, the company developed special signal latency compensation algorithms. The robot's neural network literally predicts the operator's micro movements a fraction of a second ahead to smooth out the inevitable internet lag over such a massive distance. The operator could feel the resistance of objects, carefully pick up fragile items, and hand them to expo visitors in Vienna.

[00:14:17]Astrobot's second feature is the seamless switching system. In practice, it looks like this. The robot performs a task completely autonomously. For example, sorting objects scattered across a table. It sees them with its cameras and decides how to pick them up.

[00:14:30]If an unfamiliar object appears on the table and the robot freezes, the system switches control to the operator seamlessly without stopping or rebooting. The operator remotely makes literally one or two movements, showing the robot how to handle the difficult part. The robot instantly takes the initiative back and keeps working. By the way, a similar control handoff logic currently exists in the American home robot NEO by 1x. But at the expo in Vienna itself, American companies were almost non-existent. So in this arena, the Chinese were mostly competing with themselves.

[00:15:02]And this applies not only to humanoid robots, but also to robot dogs. The most interesting one at the expo was the D1 from Direct Drive Technology. Its truly unique feature is its absolute modularity. This robot wasn't born a quadriped. It's assembled from separate, completely independent bipeedal modules.

[00:15:19]Each pair of legs with its own wheel drive is a separate robot capable of balancing and moving. But the magic happens when these modules dock with each other. They can instantly combine into a four-legged or even six-legged configuration, completely changing their physical capabilities right on the fly.

[00:15:35]This approach solves the problem of narrow specialization. The D1 robot reconfigures itself for the task. For patrolling, it can split into independent bipeedal modules. But if a heavy load needs to be transported, the modules click into a single rigid chain on four or six legs. On top of that, the architecture allows additional modules to be integrated between these mobile blocks. This is a real gold mine for engineers. And of course, the robot has the company's main killer feature, the five bar leg system. It works on the principle of a complex car suspension, allowing the robot to literally swallow any bumps without slowing down.

[00:16:12]Competing with direct drive for visitors attention were robots from Deep Robotics. But you've seen them many times. So here it's better to tell you about the company's absolute latest product, the DR02 humanoid robot. They didn't bring the new robot to such a serious conference. Perhaps to avoid damaging its reputation at the start with an accidental glitch. After all, this android is claimed to be the world's first all-weather humanoid. You could also add all terrain. The robot received an IP66 protection rating. In simple terms, it's not afraid of rain, dust, mud, or high-press water jets. It was initially designed to work outdoors and at industrial sites, and it can operate in temperatures ranging from -20° C to + 55° C, -4° F to 131° F. But that's not what brought it real popularity. It was its insane off-road race and also the fact that the robot tries to combine the ability to run with the capability to manipulate with its arms, though that part is shown less convincingly. Still, the company claims the robot standing about 175 cm, about 5'9 in tall, can carry up to 20 kg, 44 lb of cargo, and is equipped with a powerful 275 TOPS computer, meaning it performs 275 trillion operations per second. So, the robot has time to analyze potholes and rocks while running. The humanoid has a modular design. A damaged arm or leg can be quickly replaced right on site without sending the robot back to the factory, though it's a long way from Boston Dynamics Atlas robot.

[00:17:48]The Chinese compensate for imperfections in mechanics with advances in robot brains. For instance, Spirit AI was promoting its Spirit version 1.6 system at the expo. Here's the core idea.

[00:17:59]Today, most robots need to be trained separately to grab every single new object. But Spirit AI showed a model capable of finding a way to grasp objects it has never seen before. A new box, a new tool, a new part. The robot simply calculates how to pick it up. The company is betting on three data sources at once for training. The real world, computer simulations, and internet videos. The idea is simple. If humans learn from their own experience, books, and watching other people, then robots should also learn from multiple sources at the same time. And this isn't just another startup with yet another AI system. In early June 2026, the Spirit version 1.6 model took first place on the global robotics benchmark, Robo Arena, outperforming Nvidia's latest Cosmos 3 model. And that wasn't its first victory, which helped the company join the ranks of the fastest growing players in the Chinese embodied AI market and attract nearly 2 billion yuan in investment. That's around $300 million.

[00:18:56]Yes, in China, money is flowing into robotics like a river. But these numerous robots need realworld applications. And here, developers are scrambling however they can. Take PND Botics for example. At ICR, the company's robots mostly danced and moved objects around. But at a recent industrial expo in China, the developers presented their robots in an unexpected role. As factory workers, the shift in focus makes sense. Today, many companies in China are coming to the exact same conclusion. Factories are ready to pay for robots, while the homeroid market still remains somewhere in the future.

[00:19:30]PND Botics belongs to that rare group of companies that develop practically everything: drives, controllers, control systems, batteries, and software. They call themselves a full stack developer, meaning they don't depend on third party factories and do everything themselves from batteries to software. And that means they can build robots cheaper and fix bugs faster than their competitors.

[00:19:50]Just like Limx Dynamics, which brought its new ultra flexible Luna robot to ICR 2026, this is a full-sized 160 cm tall, about 5'3" in humanoid with 27° of freedom in a new generation of the SYS Zeromotion control system. The robot learns to move from ordinary video clips. You show it a dance, a gesture, or a sequence of actions, and the system can reproduce them without manually programming each trajectory. But the most interesting part is that Limx is building more than just a robot. At the beginning of the year, the company introduced Limx Kosa, its own agentic OS for humanoids. To put it very simply, this is an attempt to create an operating system that will connect the robot's perception, planning, and body control into one hole. And in April, Limx open- sourced Flux VLA, its own platform for vision, language, action models. These are the systems that allow a robot not just to see the world, but to understand commands in human language and turn them into actions. And it looks like investors believe in this approach.

[00:20:51]In February, the company raised $200 million in investment, making it one of the largest funding rounds among robotics startups this year. Unlike them, Flexion Robotics does not develop its own robots. It takes other companies hardware and teaches them practical things that can turn simple metal into a useful worker. At the expo, these guys showed the corrier of the future. The robot walked up the stairs, picked up a package, and in a purely human way, tucked it under its arm. And all of this was done in fully autonomous mode dozens of times in a row without a single glitch. The key to the demonstration was absolute autonomy, which rests on two pillars, reinforcement learning and a vision language model. The first technology is responsible for balance and allows the android to move stably.

[00:21:36]The second is its brain, which evaluates the environment recognizes the box and understands what to do with it. The trick with the unusual way of carrying the box was a success. All the engineers took notice since robots usually carry cargo directly in front of them, carefully balancing their center of gravity. For an AI, squeezing a box under its arm is a much harder task from a biomechanics standpoint. But it's too early to celebrate. The test took place in almost perfect lab conditions. Will this guy be able to repeat his trick in a real dark hallway where an abandoned kid scooter is lying on the stairs?

[00:22:08]That's a big question. Nevertheless, Flexon Robotics took a beautiful step forward, and the day when such humanoids actually start working in delivery services has gotten just a little bit closer. Unlike them, Booster Robotics isn't aiming for such tasks and is quietly taking over a completely different, underrated, but high volume market. Their booster K1 humanoid is an affordable mass-produced platform created specifically for education and research. The price for the basic geek version starts at $6,000. The educational modification costs $15,000 and the top tier professional version is priced at around $19,000.

[00:22:47]For the anthropomorphic robot market, where prices usually skyrocket past hundreds of thousands of dollars, this is very modest. Standing 120 cm, about 3'11 in tall, and weighing around 30 kg, 66 lb, the K1 robot has 20° of freedom, and is capable of autonomously running at speeds of up to 1.5 m/s, about 3.4 mph. All of this mechanics is controlled by the onboard Dubau neural network from Bite Dance, the same people who created Tik Tok. This allows the robot to recognize complex voice commands and scenarios without latency and without connecting to external servers. In addition, Booster Robotics is an official partner and platform provider for RoboCup, the World Robot Soccer Championship.

[00:23:33]IC 2026 wrapped up with an Imperial evening at the Hapsburg residence, but the true summary of the expo came from the company booths, not the entertainment. American players were barely visible, while Chinese engineers proved that hardware is no longer a problem. Robots have gained sensitive skin, modular bodies, and neural network brains. And judging by what we saw in Vienna, these exact companies are reshaping the rules of what our future will look like right now. Subscribe to the channel so you don't miss these breakdowns. See you in the next video.