This Robotic Hand Stole the Show at ICRA 2026

[00:00:00]This is not CGI. This is a robotic hand that weighs less than a bag of sugar, packed with 20 independently moving joints. It was built by a tiny startup out of Shenzhen that just 5 years ago did nothing but manufacture basic electric motors. But today, this hand, along with one secret ingredient, is about to completely change the trajectory of the robotics industry. And while Elon Musk keeps Tesla's latest design strictly behind closed doors, and Boston Dynamics focuses on ultra-conservative engineering to guarantee reliability, this team in Shenzhen has built a solution that breaks all the traditional rules. A piece of hardware so disruptive that AI teams from Google Deep Mind and Physical Intelligence are flying all the way to ICRA 2026 in Vienna just to get a look at it. Today we're looking at how a small-time Chinese motor company solved a problem that for 21 years has stumped Stanford, Carnegie Melon, DARPA, Google DeepMind, and Shadow Robotics, a company with a robotic hand that costs a staggering $100,000.

[00:01:00]It's a problem tech companies have carefully hidden for years behind flashy Boston Dynamics montages and Nvidia revenue charts. Who exactly is Wuji Technology? And why are top American labs like Liber AI choosing these hands to push their AI models into the physical world, completely bypassing options from Tesla, Figure, or Sanctuary AI? How does a $5,000 robotic hand do things a $100,000 hand can't? And finally, the most uncomfortable question of all. How did we end up in a world where artificial intelligence can write complex code, paint masterpieces, and negotiate deals, but still can't physically pull a $5 bill out of your wallet? Let's be honest, over the last 2 years, you and I have watched dozens of jaw-dropping robot videos. But most of them hid one very important thing, actual handwork. And even when hands were shown on screen, it was strictly in controlled lab conditions. We have absolutely no idea how many takes it took to get that shot. Yes, there is a reason Elon Musk only published Tesla's hand patents after they were already outdated. And there is a reason he is carefully hiding the work his team is doing on the hand right now. On one hand, it's just marketing, creating the illusion that they're making something so unique it's too secret to show. On the other hand, it might mean that all the issues with the hand are still nowhere near solved. Optimus Gen 3 is claimed to be nearly ready with 22 degrees of freedom, but while everyone has heard about it, nobody has actually seen it. It's a presentation slide, nothing more. As of the recording of this video, no one has seen a live demonstration of Optimus Gen 3 grabbing an object and holding it without dropping it. Not a single public appearance with a real object. And Musk isn't the only one struggling with the master key to generalurpose robots. Take Boston Dynamics. Their new Atlas is fully ready for mass production and factory work. But look at its hands.

[00:02:50]They are not designed for fine, precise manipulation, only heavy labor. Because that is the engineering compromise the developers had to make. And yet this is the best thing that has ever happened to this robot. Because for most of its history, it didn't have hands at all.

[00:03:04]Just endcaps. Boston Dynamics, the best engineering team on the planet, which spent $200 million of DARPA money and has been working on humanoids for 30 years, spent 10 years meticulously solving the balance problem while officially giving zero thought to the hands. When DARPA first required participants to handle tool-based tasks at the 2015 Robotics Challenge, opening doors, turning valves, using drills, Boston Dynamics slapped a primitive third-party three-pronged gripper onto Atlas at the absolute last minute. But you can't launch a commercial humanoid without hands. So, the production version of Atlas does have them. They deliberately chose mechanical simplicity because on a real assembly line, a broken finger means a stalled multi-million dollar factory floor. It's brilliant industrial engineering, but it leaves a massive gap. It completely sacrifices the highfidelity dexterity needed for the next generation of AI. Of course, there is also the figure 03 robot, the best Western result to date.

[00:04:02]16 degrees of freedom, cameras in the palms, sensors that can feel a weight of just 3 g, about 0.1 o. The press is ecstatic. Figure itself is valued at nearly a billion dollars. And yet during their own reveal stream, their engineering lead dropped that uncomfortable truth. The hand was the hardest part of the entire robot. It wasn't balance, AI, sensors, or power that turned out to be the toughest part of the robot. It was the hands. It's worth adding that in real world conditions at the BMW plant, their hands lasted for 10 months of work. After that, the robots had to be completely scrapped due to wear and tear. In other words, the promises that you would buy a robot and have it grind for you for years without vacations or sick days simply fell flat. Then there is the massive pool of robots from China, WRC, IRS, IC, GIT X, the biggest robot expose of the last 2 years have showcased dozens of startups. The ultrammobile and agile engine AIT 800 robots, unusual prototypes from PND Botics, early demo robots from robot era. Most of them don't have any hands at all, just end caps. Those that technically do have hands use the Inspire RH56. Remember that name. It is the de facto standard of the Chinese humanoid segment. The RH56 has six degrees of freedom, only six. That is the level of an industrial gripper from 1985, where each finger is essentially a separate pneumatic actuator with no independent control.

[00:05:32]The flagship humanoid from industry giants still relies on basic 6DF grippers. In the precision demanding world of 2026, a 6DF gripper isn't a hand. It's a digital mitten that locks the robot out of 90% of human environments. Even though the company started developing its own hand, the Dex 5, with a claimed 20° of freedom, we haven't heard a single thing about it for a year. And in their latest videos, the most we see are basic grippers. So, that's the whole picture as of May 2026.

[00:06:03]Actually, scratch that. There's one more thing that absolutely has to be mentioned. Let's explain it with an example. October 2025, Norway's 1X, one of the most hyped humanoid projects in the world, the flagship of the Open AI startup fund with over $100 million in funding, rolls out pre-orders for its home humanoid, NEO. The price tag is $20,000. The marketing pitch calls it the world's first consumer humanoid.

[00:06:29]Reserve yours now. Delivery in 2026. But if you dig into the FAQ, one tiny detail appears in fine print. For chores, Neo doesn't yet know. Owners can schedule a 1x tea operator to guide it. Let's break down what that actually means. When you buy this $20,000 machine to clean your kitchen, you're not just getting an autonomous AI. You're booking a session with a human operator sitting on the other side of the world physically steering the robot through a VR headset.

[00:06:57]And as industry insiders point out, relying on thousands of remote humans to guide home robots simply isn't a scalable AI solution. It's just shifting the labor bottleneck from your kitchen to a remote office. To achieve true autonomy, the AI model doesn't need a human permanently pulling the strings from afar. It needs highfidelity data directly aligned with human level dexterity. And that is exactly where the industry is hitting a massive wall. And this brings up the most uncomfortable question of this entire video. If hands are the bottleneck and the industry has known this for 20 years, why has no one made a normal hand that you can just buy and plug into an experiment? Well, actually, someone did. Well, almost. The Shadow Hand E-series is the only robotic hand where the number of degrees of freedom gets close to what you could call real dexterity. 24 degrees of freedom, almost matching the human hands 27. It is the gold standard of the entire industry used by DeepMind, Open AAI, Stanford, MIT, and Carnegie Melon.

[00:07:58]An engineering marvel. The price tag anywhere from 100,000 to $150,000 a piece. In 21 years, this hand hasn't moved the embodied AI industry forward by a single inch. Not because it's bad, but because you can't just buy 10,000 of them and put one in every lab, every factory, and every home. It is a handbuilt laboratory tool, not a mass product. One broken joint and the robot sits in a corner for two weeks waiting for a technician to fly in from London.

[00:08:26]And this is exactly where the story begins that none of the big media outlets will tell you today because it's about motors and motors are boring. Who is going to watch a video about motors?

[00:08:37]Well, maybe you will. And let's start it like any real story from the beginning.

[00:08:41]March 2019, Shenzen. A guy named Pan Yun J registers a company called Wuji Technology. They create unique pancake- shaped drives under the brand name Pan Motor. These are brushless motors, tiny ones, literally the size of a matchbox and weighing about as much as a tub of cottage cheese, just 200 g, about 7 oz.

[00:09:02]But this little thing packs so much punch that if you attach a lever to it, it can lift a bucket of water all by itself. Engineers call this field oriented control. But to put it simply, they managed to cram the power of an industrial machine into the size of a toy. What usually weighs a full kilogram, 2.2 lb, and takes up a ton of space on standard robots. They shrank by a factor of five without losing a single bit of strength. It was engineering magic right on the edge of what's possible. Who buys these motors? Servo manufacturers, exoskeleton builders, a couple of labs playing around with the cheap and mass-produced approach. The business is stable and quiet with no massive funding rounds up until 2023.

[00:09:43]And then this happens in 2022. Elon Musk wheels Tesla Optimus onto the stage and the entire world suddenly decides that humanoids are the new oil. In 2023, Boston Dynamics reveals the electric atlas. In 2024, Sergey Lavine and his team raise $400 million at physical intelligence to build foundation models for robotics. Stanford's Aloha proves that bimmanual teley operation works for just $32,000.

[00:10:12]And the whole industry realizes we have the brains, massive foundation models trained like chat GPT, but for movements. And we have the bodies, dozens of humanoids for every possible use case. But between the brain and the body, there is a gaping void, the hand.

[00:10:28]And Pyunha, a dual major in computer science and chemistry who has never written a highbrow scientific paper on dynamic balance theory in his life, sits down and thinks something like, "Wait a second, what is the most expensive component in a robotic hand? It's the actuator, the drive, a tiny motor with a gearbox and a controller. It makes up 30 to 50% of the total cost. And we just happen to know how to build the most compact, highdensity motors in the world. So, why don't we build a hand where the motors aren't dangling in the forearm on tendons, but sit directly inside the fingers. This is the ultimate architectural shift. Nearly every other hand on the market, Shadow, Tesla, Figure, Sanctuary relies on tendon-driven systems. The motors sit in the forearm and cables run from those motors into the fingers, mimicking human muscles and tendons. It feels organic and biomimetic, but it is a total nightmare. Cables stretch. Cables create backlash. Cables introduce nonlinearity that is incredibly difficult to model accurately in a simulator. And an ML model trained in simulation eventually has to run on real hardware. If the physical hardware doesn't match the simulation, the model breaks completely.

[00:11:36]The Wuji team went completely against the grain. They ditched the cables entirely and packed micro motors with rigid gears directly inside the felanges of each finger. This is what's called a direct rigid drive system. Whatever the control signal tells the motor to do, the finger does instantly. It is linear, predictable, and has absolutely zero backlash. On top of that, the design turned out to be so rugged that it easily survives an accidental drop from a desk. On September 17th, 2025, Wujite drops a video on Billy Billy. This isn't just another robotic hand. It features 20° of freedom, weighs under 600 g, about 1.3 lb, and delivers 15 newtons, about 3.4 4 lb of force at the fingertip. It boasts 300,000 cycles of factory testing and over a million internal test cycles. The price anywhere from 5,500 to $16,000 depending on the configuration. That is 18 times cheaper than the Shadow Hand.

[00:12:32]And it's not because it's some budget model. It is in the exact same research grade category matching those same 20 DOF. A month later, the entire industry caught wind of Wuji when leading AI model companies, next-gen humanoid builders, and top tier industrial integrators started rolling out their latest high dexterity demonstrations.

[00:12:52]Guess whose hand was quietly powering their hardware? Not Shadow, not Tesla, not Inspire, Wuji. This is the first time in history that an American AI lab has integrated a Chinese robotic hand into its flagship project. And that was just the first version, Wuji hand. At ICRA 2026 in Vienna, Wuji introduced its second generation, Wuji Hand 2, an upgraded platform designed for more advanced dextrous manipulation, data collection, and embodied AI research.

[00:13:21]The Wuji Hand 2 isn't just a V1 with minor upgrades. It's a leap into a completely different league. And here is what's truly mind-blowing about it.

[00:13:29]First, it is engineered to be fundamentally humanlike, right down to the millimeter. The palm dimensions are exactly 185x 80 mm, about 7.3x3.1 in, the precise size of an average adult man's hand. Not roughly similar, one to one. And despite packing all its motors directly inside the fingers, the entire mechanism weighs a mere 580 g, about 1.3 lb. That's lighter than a standard tablet, yet it replicates human anatomy with stunning fidelity. Why does this matter? So the robot doesn't have to redesign the world around itself. Any doorork knob, light switch, or wrench built for a human fits this hand perfectly. Inside there are 20 controlled joints. Humans have 27, but some of those are passive. So comparing a hand with 20 actively controlled joints to the hand of a surgeon performing a micro operation is completely accurate. This allows the hand to pull off incredible balancing acts. For instance, the fingers can gently hold a raw egg without cracking it and in the exact same second forcefully pop a pill out of a rigid blister pack. But the most important breakthrough is hidden in the control electronics. The engineers are incredibly proud of their 1,000 hertz spec. What does that mean in plain English? When you pick up a full glass of wine from a table, your brain isn't thinking about it once a second. You adjust your fingers continuously, catching the balance on autopilot. The Wuji hand does the exact same thing, but at the speed of a supercomput. It checks 1,000 times a second exactly how much force each finger is applying to the object. That operates faster than human reflexes. That is exactly why it possesses a critical feature, compliance, or what engineers call back drivability. Most robots on the market are as rigid as mannequins. If you slam into a robot like that or drop a heavy book onto its arm, the gears inside will strip and snap. The Wuji hand, however, knows how to give and relax at the exact moment of impact. It cushions the blow, saving both itself and any human standing nearby from getting hurt. And finally, here is the feature that has AI developers absolutely ecstatic. Right out of the box, this hand is natively compatible with all the major virtual physics simulators, native ROS 2, Sim tore in Mujoko, and Nvidia Isaac Sim.

[00:15:43]Remember in the Matrix when Neo has martial arts uploaded straight into his brain in a matter of seconds? This is the exact same thing. Programmers don't need to spend months calibrating physical hardware. Instead, they train the neural network inside a computer game where time can be sped up thousands of times over. The robot virtually picks up a cup a million times until it figures out the technique. And then that fully formed skill is uploaded into the physical Wuji hand in a single second.

[00:16:10]Because the mechanical build of the fingers is so rigid and precise, the robot can transition what it learned in the digital world directly into physical reality. But training in a simulation is one thing. Facing the real world is another. That is why Wuji chose IC 2026 in Vienna, not for a polished pre-recorded promo video, but for live realtime demonstrations in front of the global robotics community. And the response was clear. researchers, engineers, and industry teams were paying attention. But truth be told, the Wuji Hand 2 isn't even the centerpiece of their ecosystem. The real star is this thing right here, a data collection glove. It's a wearable glove packed with sensors on every single joint, mapping out 20 DOF. Those are the exact same 20° of freedom found on the Wuji Hand 2. And this is where the real magic begins. Let me explain something to you that you've never heard in standard robot reviews.

[00:17:05]The most expensive, rarest, and most limiting part of AI for robotics isn't computing power. It's not motors. It's not sensors. It's data. To train Chat GPT, you need terabytes of text from the internet. The internet already exists.

[00:17:21]It's been accumulating for 30 years. You just take it and train. To train a robot, you need data on how hands and bodies move in the real world. That data does not exist on the internet. It has to be collected from scratch. And there are only two ways to collect it. Method one is teleoperation. A person puts on a special suit, grabs two controllers, and every single move they make is recorded, and then played back by the robot. The industry standard for this is Stanford's mobile Aloha. It costs $32,000.

[00:17:51]That is four to five times more expensive than just recording a video of a human doing a task. And get this, to collect just one data set on the level of Google's RT1, you need 13 robots, 17 months of non-stop work, and 130,000 recorded trajectories. Andre Karpathy, former head of AI at Tesla and co-founder of Open AI, spells it out directly. The internet scale pre-training paradigm simply does not scale in robotics. Google Deep Minds Deisabis agrees. The primary bottleneck in robotics isn't hardware, it's intelligence. Because physical context cannot be described with words. And Sergey Lavine, founder of the premier robo startup Physical Intelligence, sums it up. We are on the verge of a self-improving flywheel for generalpurpose robots, but it requires data. If you think this is all just theoretical talk, let's look at a historical example. 2019, Open AI, the exact same Open AI that would roll out Chat GPT just 3 years later and upend the global economy, publishes its flagship robotics project called Dactyl.

[00:18:54]The video shows a five-fingered robotic hand solving a Rubik's cube. To make that happen, the engineers ran a physics simulation. Inside this virtual world, the robot turned the cube non-stop, racking up a grand total of 13,000 years of continuous experience. Just wrap your head around that. a computer intellect trained for longer than human civilization has even existed. And what was the outcome of this technological breakthrough of the decade? In the real world, the robot could only solve the cube about half the time. The success rate fluctuated wildly between 20 and 60%. Depending on how heavily it was scrambled. In other words, after burning through a colossal amount of computing power and 13,000 years of virtual practice, scientists ended up with a hand that failed every other test. So, what did OpenAI do? A year after that publication, in 2020, they quietly disbanded their robotics team. In July 2021, they made it official. Robotics is shut down. We are pivoting entirely to generative AI. It wasn't until 2024, after 4 years of dead silence, that Open AI quietly returned to robotics. They reconstituted the team and started hiring top tier engineers. Why now?

[00:20:05]Because an idea emerged on how to finally bypass data scarcity. And that idea is foundation models paired with data collected from wearable sensors on actual humans. The exact same idea that Wuji is showcasing live in Vienna. So let's look back at that glove. What if data collection wasn't about complex teleoperation with multiple operators confined to a specialized lab room? What if a regular factory worker simply slipped on a glove with 20 sensors and went about their normal shift 8 hours a day every day multiplied by several thousand workers? By the end of the shift, you have data. Clean, realworld, highly diverse data on how a human hand performs thousands of actual tasks in a live environment. No sterile lab conditions. No operators wearing tracking goggles. Just pure work. And here is the kicker. At any given microcond, the glove records exactly 20 data points, the precise angle of every single joint. The Wuji Hand 2 features the exact same kinematics in the exact same joints. So when the neural network processes these 20 numbers, it has no idea whether it's looking at a human wearing a glove or a robot executing a task. In the eyes of the AI, there is absolutely no difference between a human hand in that glove and the Wuji dextrous hand. This is what's known as embodiment agnostic learning. It is the absolute hottest academic concept of 2025 and 2026. Projects like Stanford's universal manipulation interface, DEXCAP, and embodied R1 have all been trying to bridge the embodiment gap, the massive mismatch between the physical body used to collect data and the physical body that eventually has to execute it. Wuji bridges this gap entirely at the hardware level. Same hand, same kinematics, same exact signal. And that is the strategic master stroke. Wuji isn't trying to compete with the AI giants building the master brains for these robots. Instead, they are acting as the ultimate roadbuilder for foundation models. While everyone else is stuck trying to sell isolated hardware, Wuji bridges the embodiment gap entirely at the hardware level.

[00:22:07]Because their data collection glove and the Wuji Hand 2 share identical kinematics, every single twitch, bend, and micro adjustment of a human hand translates seamlessly into the robotic equivalent. It's an open invitation to every AI lab on the planet. You build the intelligence and we'll provide the perfect physical highway to bring it to life. It's a data flywheel that starts spinning on its own the moment the industry adopts this unified standard.

[00:22:31]Let's be totally honest. What are the things robots flat out cannot do right now? Tying shoelaces. In late 2025, China proudly announced a breakthrough, but under the hood, the robot had an 80% success rate inside a highly controlled lab on one specific pair of sneakers. It won't tie yours. pulling a bill out of a wallet. This problem remains publicly unsolved by anyone. The paper is too thin. The wallet is too flexible. Robots simply lack that tactile finger sensitivity. Popping a pill out of a blister pack. Another total failure. It requires an incredibly precise calculation of force. Securely holding the plastic container with one part of the hand while neatly punching through the foil with another. Peeling a banana, cooking a dynamic meal, or conducting delicate laboratory experiments. Virtual robots can perform these tasks flawlessly inside computer simulators.

[00:23:22]But the moment you bring those models into the real world, standard hardware fails because it lacks the tactile sensitivity to handle soft, unpredictable objects without crushing them. And here's the ultimate insight.

[00:23:34]None of these tasks are failing because of the brain or the algorithms. They are hitting a wall because of the physics of the hand itself. A robot needs four critical things. the ability to spring and flex, human-like anatomy, ultra lightweight, and insane power density in every single finger. Without that foundation, any manipulation task is physically impossible, no matter what software you run. Which means the geniuses at Google Deep Mind or physical intelligence can build a 100 brilliant neural networks. But as long as AI lacks proper hands, it will remain completely trapped inside computer simulations. But right now, it feels like history is shifting and you can witness it firsthand. Case in point, ICR 2026 running from June 1st to June 5th at the Mesaveen Exhibition Congress Center.

[00:24:20]More than 3,000 engineers, researchers, and founders from all over the globe are gathering here. Absolutely everyone will be in the room. The engineering teams from Tesla, Figure, and Sanctuary will be walking the floor, even if they don't have official booths. Demis Hassabis won't be there in person, but half of his R&D team definitely will. And on Tuesday, June 2nd, the day of the first major keynote and peak press coverage at booth 121, the Wuji Hand 2 stepped up for a live unedited demonstration of realworld manipulation tasks. No studio edits, no pre-rendered graphics, just pure realtime performance in front of the global robotics industry's toughest critics. at ICRA. The demo was simple to understand but difficult to execute.

[00:25:05]Wuji Hand 2 rotated a Rubik's cube in real time, coordinating multiple fingers to maintain contact, adjust force, and keep the object under control. The stakes couldn't have been higher. Going into Vienna, the question wasn't just whether the hand worked, but whether it could redefine the market. Over those five intense days, this hand was either going to emerge as the breakout star that finally unlocks embodied AI or remain just another premium piece of research hardware for a select few labs.

[00:25:34]We bet on the first one. And looking at the waves it just made across the industry, that bet is already paying off. At the beginning of this video, we talked about how an artificial intelligence that writes complex code still can't pull a bill out of your wallet. In 30 years, we've successfully modeled billions of parameters for large language models. We've learned how to render photorealistic video from a single text prompt. We've built robots that can stick back flips and dance to Michael Jackson. Yet, we still haven't figured out at a product scale how to collect enough data on the basic movements of a human hand. This is the exact physical bottleneck of AI, a phrase Wuji anchors in its core messaging. And unlike most corporate marketing buzzwords, this one is entirely literal. Morgan Stanley estimates the humanoid market will reach five trillion dollars by 2050 with a billion androids deployed worldwide. But for a single one of those robots to actually work inside your home, a hospital, or a manufacturing plant, someone has to solve the exact problem that Shadow Robotics got stuck on back in 2005. Want concrete proof of just how massive this problem is in cold, hard cash? Look at Foxcon, Apple's primary manufacturing partner. The actual factory floor where your iPhone is put together. Back in 2012, Foxcon CEO Terry Go promised Apple, and I quote, "We will have 1 million robots on our assembly lines in 2 years." By 2019, Foxcon had deployed only 100,000, 10 times less than promised. Why did a million robots shrink to a fraction of that? A report from Apple Insider states it verbatim.

[00:27:06]Apple uses screws so tiny that robots had no way to measure the force used to drill them in. Human workers can feel the resistance from their hands. The screws inside an iPhone are so incredibly small that automated machinery cannot feel the tightening force, but a human can right through their fingertips. In 2012, Apple itself, running parallel to Foxcon, opened a secret robotics lab just 6 mi from Apple Park. Their goal was to build their own in-house assembly automation. In 2018, that lab was shut down. Apple brought the humans back. The wealthiest company in human history, boasting a $3 trillion market cap, could not replace human fingers on the iPhone assembly line. 13 years of attempts, billions of dollars poured in. The end result, back to humans. If you still think the hand problem will just resolve itself with a they'll fix it when they need to attitude, remember that number. $3 trillion couldn't fix it. Now, I'm not claiming Wuji is guaranteed to solve it.

[00:28:04]Maybe it will be Tesla if their Gen 3 ever drops. Maybe it'll be someone else we haven't even heard of yet. But what I am saying is that after June 2nd, 2026, the list of players actually capable of solving this is going to narrow down to just four or five names. And one of them is a team out of Shenzhen that 6 years ago was just building motors for robots.

[00:28:26]And you know what? This feels a lot like what happened with Hyundai and Atlas.

[00:28:30]Back then, nobody believed a car manufacturer could solve a problem that had stumped Google, SoftBank, and DARPA.

[00:28:36]And back then, what caught us off guard was the actuator insight. The fact that Hyundai already had that exact component running on their assembly lines. Wuji had that exact same advantage. 6 years of building tiny motors, 6 years living and breathing FOC controllers, and now they have a hand. The most expensive component in a robot is still the actuator. And the player who already has them rolling off the production line is going to keep winning. Whoever that ends up being, Hyundai, Wuji, or a third player will see at 2027. If you want to understand the physical bottleneck of embodied AI, look at what Wuji showed at ICR 2026 in Vienna. Wuji Hand 2, a second generation dextrous hand, a data collection glove, and a live Rubik's Cube manipulation demo that brought the hand problem into focus. In the meantime, drop your thoughts in the comments. After seeing what Wuji Hand 2 demonstrated at ICR, what do you think will matter most in the humanoid race by 2030? Better AI models, better bodies, or better hands?