New Unitree G1 Robot Just Got 6 HUGE Upgrades… Think, Move, Fight, and Learn

[00:00:04]Unitry may have just shown the widest range of skills any humanoid robot has displayed so far, but this is not about one viral clip or one lucky demo. The G1 is being pushed in six very different directions at once, and each one tests a different limit. It listens to spoken commands and turns them into full body motion. It dances in perfect sync with dozens of other robots. It copies a human operator in real time, throws kung fu kicks, works on factory floors, and lands back flips that most humanoids could never survive. That means the G1 is not just walking better, but understanding better, coordinating better, and slowly moving toward real work. A year ago, humanoids were judged on whether they could walk. Now, the G1 is forcing the whole industry to prove what comes next. So, let's break down the six abilities that make this small robot so important. The first ability is not just voice control. It is real-time language to motion generation. In Unit's demo, the G1 does not need a joystick, a remote, or a programmer typing exact movement commands. It listens to spoken instructions, replies with lines like, "Please wait a moment. I'm thinking," and then converts the request into full body action. That is why this demo matters. The robot is not only following simple orders like warming up, jumping in place, or holding a plank. It also handles multi-step movement such as turning around, placing its hands on its hips, raising both arms straight out, walking forward, and returning to the starting point. Then the difficulty increases with half squat walking where the G1 has to control balance, leg movement, posture, direction, and timing together. The same ability also lets it perform more expressive actions like kneeling on one knee to propose and saying, "Wish you happiness or doing a Gangnam style routine followed by a bow." Unitry calls this voice-driven real-time arbitrary action generation.

[00:01:49]And the key shift is simple. Language is no longer only text on a screen. With the G1, language becomes physical movement. Is voice control the real breakthrough here, or is the Gangdom style dance more impressive? Share your thoughts in the comments. One G1 responding to voice commands is impressive, but Unit's larger test was making many G1 robots move like one connected team. That is what happened on America's Got Talent where eight Unitary G1 robots performed beside Chinese dancer Wu Yu Fay, also known as Flying Bug in a synchronized routine that earned a standing ovation and unanimous approval from the judges. This mattered because the robots were not hidden inside a lab or shown only to engineers.

[00:02:28]They were performing on NBC in front of a prime time American audience on one of the biggest entertainment stages in the world. Unitry then scaled that same coordination much further at the 2026 Spring Festival. gala where dozens of G1 robots performed with kung fu students in front of hundreds of millions of viewers. The performance mixed martial arts with robot coordination and was described as the world's first fully autonomous humanoid cluster performance.

[00:02:54]The important part is that this was not just a group of robots replaying the same routine. A high concurrency cluster control platform kept them synchronized with very low delay, while onboard 3D LAR helped each robot scan the stage and hold its exact position. During formation changes, the robots moved at 3 to 4 m/s without crashing into each other. Each unit could also monitor its own movement and rebalance if a landing went slightly wrong. In 2025, Unitry showed simpler Yangko dancing. By 2026, the G1 had moved into full dynamic choreography, proving it could scale from one robot into a coordinated crowd.

[00:03:33]But what if you want to control every joint yourself? From America's Got Talent to China's Spring Festival Gala, robots like the Unitry G1 are changing what humanoids can do together.

[00:03:45]Subscribe to the AI Nexus for clear breakdowns of the robotics news that actually matters. Now, the third ability flips the G1 into something very different, an embodied avatar that can copy a human operator in real time.

[00:03:57]Instead of the robot deciding every movement alone, a person moves their own body, and the G1 maps those movements onto its joints, posture, and actions.

[00:04:07]The football demo shows why this ability matters. G1 robots controlled by human operators move through a coordinated game, reacting with human timing, while the robot body handles the physical motion. This is important because full autonomy is still the long-term goal.

[00:04:21]But telly operation makes humanoids useful before they can do everything alone. A trained human can provide judgment, quick decisions, and task understanding, while the robot gives that person a physical presence in places that may be unsafe, distant, or difficult to reach. The hardware makes this possible. Depth cameras and LAR support spatial awareness. The robot's joint freedom helps it mirror human motion, and the dextrous hands on the EDU version allow finer control. But the deeper value is training. Teley operation is not only remote control. It also creates demonstration data for future autonomous models. The human pilots the robot. The robot records the movement and that data can help the G1 learn better actions later. So what happens when that training is pushed towards something much more athletic.

[00:05:05]Now unitry pushes the G1 into martial arts where the point is not fighting but testing how much balance, speed, and dexterity this small humanoid body can handle. In the Kung Fu Kid 5 6.0 O demo.

[00:05:17]The G1 throws a high left leg kick, extends its arms to stay balanced, then spins, punches, and drops into low sweeping movements. Unitry says the video has no speedups, so the motion is being shown at real pace. Then the test becomes even harder as the robot performs rapid somersaults, aerial twists, and back flips before landing cleanly. The engineering value is in the recovery because the G1 uses human-like stances and arm movement to stay stable after motions that would make most humanoid robots fall. But the bigger layer is the hand control. At the spring festival gala, the G1 did not only move its body, it gripped and switched real martial arts props like staffs and nunchaku while performing with human fighters that demands force control, timing, grip changes, and full body balance at the same time. The drunken fist showcase made it even harder because the movement is fluid and unpredictable. So when people ask why it does kung fu instead of laundry, the answer is simple. Every kick, flip, sweep, and prop switch is unitry stress testing the G1 at the edge of motion.

[00:06:18]And that leads straight to the factory floor. What is more impressive to you?

[00:06:22]The G1 being controlled like an avatar or the G1 doing back flips and kung fu on its own. Share your thoughts. Factory work is where all these G1 abilities start to connect into something useful.

[00:06:33]Because a real workplace needs more than one impressive movement. It needs a robot that can understand instructions, move through changing spaces, repeat tasks, and recover when small things go wrong. That is why Unit's factory direction matters. The same G1 that can respond to voice commands could be redirected by a worker during a task.

[00:06:50]The same teley operation system could let a human guide it through a tricky step. And the same embodied AI approach could help it repeat that work without needing every motion controlled by a person. Unitry is building onboard models meant to help the G1 navigate complex environments and adapt instead of only replaying a fixed routine. This is important for logistics, inspection, assembly, and other dull or risky jobs where humanoids could eventually help.

[00:07:14]But the honest part is that real factory adoption is still harder than the demos make it look. A clean demo floor is not the same as a messy warehouse. Still, the direction is clear. Unitry wants to ship 20,000 units in 2026, up from 5,500 last year. and factory use is the ability that could turn the G1 from a viral robot into a working robot. But for the final ability, Unitri goes back to pure motion and pushes it even further. Dynamic mobility is where Unitry shows how far the G1's body can be pushed when balance, power, and coordination all work together. At the Temple of Heaven in Beijing, dozens of G1 robots performed in formation near the Hall of Prayer for good harvests using synchronized punches, kicks, and high difficulty flips to test group motion in a real public setting. The numbers show why this matters. The G1 can perform somersaults reaching 3 m in height, run at speeds up to 4 m/s, complete continuous single leg flips, pull off two-step wall assisted back flips, and even hit an airflare spin with 7 and 1/2 rotations. These are not just show moves. They test explosive power, joint torque, landing control, real-time balance recovery, and the same cluster system that keeps many robots moving together without breaking formation. For a compact 35 kg humanoid, that level of motion shows how much stress the frame can survive. So, what do all six abilities add up to? What is more important for humanoid robots, factory usefulness or crazy athletic movement like flips and wall assisted back flips? Tell me in the comments. The real question is no longer whether the Unitary G1 can move. The real question is how many different jobs one small humanoid body can be trained to handle.

[00:08:54]Voice control, synchronized dancing, teley operation, martial arts, factory movement, and high difficulty acrobatics are not random tricks. They are different tests of the same bigger idea.

[00:09:05]A humanoid robot should be able to understand commands, move with other robots, follow human control, recover its balance, and slowly move toward real work. The honest part still matters.

[00:09:16]Most of this is happening in control demos, and the everyday household robot is not here yet, but the jump is clear.

[00:09:23]A year ago, humanoids were judged by walking. Now, the G1 is forcing the whole industry to prove what comes after that. While everyone is still waiting for humanoid robots to become useful, something strange is already happening.

[00:09:34]The race has quietly moved out of clean demo rooms and into factories, warehouses, and production lines. Figure 03 is no longer just standing in a lab.

[00:09:43]It is sorting packages for hundreds of hours. Xpang Iron is not just walking like a human on stage. It is already being tested inside EV production lines.

[00:09:52]Boston Dynamics Atlas is not doing another backflip. It is carrying heavy, awkward objects like a real factory worker. UB Techch Walker S2 can even swap its own battery without human help.

[00:10:02]Apollo is being trained with Google Deep Mind and Agility Digit is already clocking in at warehouses. But here's the real question. Which of these robots is actually ready for work and which ones are still just impressive demos?

[00:10:14]Today we are breaking down the humanoid robots moving from hype to real labor.

[00:10:18]First up is Figure 03. And this one matters because it is not just another humanoid robot doing a flashy demo.

[00:10:24]Figure is trying to prove something much bigger that humanoid robots can actually become reliable workers. The first proof came from BMW's Spartanberg plant where Figur's earlier robot spent 11 months doing real factory work. It loaded more than 90,000 sheet metal parts with over 99% accuracy which showed that humanoids can handle repetitive jobs where endurance, safety, and consistency matter more than viral tricks. But Figur's recent live stream made the story even more interesting. In a man versus machine sorting challenge, intern Amea Gerard went head-to-head against figure 03 robot for 10 straight hours.

[00:10:59]Aay barely won sorting 12,924 packages while figure 03 sorted 12,732.

[00:11:08]But the difference was what happened after a finished with blistered hands and serious forearm strain while Bob simply kept going. And this was only one part of a live stream that crossed 200 hours where a rotating fleet of Figure03 robots handled more than 249,000 packages with autonomous battery swaps and robot handoffs. The real engine behind this is Helix O2 Figures AI system that helps the robot see objects, understand what to do, move its body, and manipulate packages without telly operation. And now those same skills are starting to move beyond warehouses. Two figure 03 robots recently tidied a bedroom and made a bed together in under two minutes, showing that factory style reliability could eventually transfer into human spaces. That is why Brett Adcock's comment feels so important.

[00:11:55]This may be the last time a human wins.

[00:11:56]If figure 03 can sort packages for more than 200 hours while humans need breaks, is this the moment humanoid robots become real factory workers? Comment below. But this robot race is not only being shaped in American labs or premium automaker plants. In China, Xpang is showing a very different path with its humanoid robot iron. Xpang is mainly known as an electric vehicle company.

[00:12:18]But with Iron, it is trying to turn its own car making ecosystem into a training ground for humanoid robots. And the first time Iron grabbed global attention, it was not because of factory work. It was because of how human it looked on stage. When X Pong introduced Iron, its catwalk movement looked so smooth and natural that many people thought there was a real person inside the robot suit. So CEO Hey Xiaopang literally cut open the robot's leg on stage to prove it was not a human wearing a costume. But that viral moment was only the surface. The real story is what Iron is being built for. Its bionic bone muscle skin structure and flexible spine are designed to make the robot move more naturally around people. While more than 60 joints and 22 degrees of freedom in each hand help it perform factory style actions like reaching, gripping, guiding parts, and handling objects. Then comes the brain behind the body. Xpang gave Iron three in-house touring AI chips with 2,250 to ops of compute and its VLA 2.0 model helps the robot connect vision, language, movement, and decision-making in the real world. That matters because Iron is not just walking for cameras anymore. Hundreds of units are reportedly already working inside Xpong's Guanjo EV production lines, helping with material handling, assembly, and process guidance. He Xiao Pong says large-scale production should begin by the end of 2026. And his bigger prediction is even bolder. One day, robots could become a bigger business for Xpong than cars. Then comes Boston Dynamics Atlas. And this time, the story is not about another backflip or viral parkour move. Boston Dynamics is now showing what Atlas may actually be trained for. Heavy, awkward factory work. The clearest example came from its recent mini fridge demo. In that footage, Atlas lifted a loaded mini fridge, carried it, turned with it, and placed it down while constantly adjusting its balance in real time. That matters because factory work is not always clean or perfectly arranged. A robot may have to deal with heavy parts, awkward shapes, tight spaces, and objects that are not placed exactly where they should be. That is where its new three-fingered grippers with tactile sensing become important because they help the robot feel and control what it is holding instead of just grabbing blindly. Atlas also has 56° of freedom and 360° rotating joints which allow it to bend, reach, and move through factory spaces in ways a human body cannot. But the bigger story is where Atlas is going next. Boston Dynamics is now working with Hyundai Motor Group and Atlas units are being sent to Hyundai's robotics metaplant application center and Google DeepMind for training. DeepMind's large behavior models are expected to help Atlas learn factory tasks faster and become more useful in real production environments. The rollout is planned to begin at Hyundai's Georgia Metaplant in 2028, followed by Kia's Georgia plant in 2029 across major manufacturing processes. So Boston Dynamics is no longer just proving that Atlas can move beautifully. It is preparing Atlas to become a real production worker. Then the race moves back to China with UB Walker S2. And this robot is focused on one of the most boring problems in robotics, but also one of the most important uptime. A humanoid can look amazing on a factory floor. But if its battery dies and a human worker has to stop everything just to recharge it, then the robot is not truly ready for real industrial work. That is exactly the problem UB Tech is trying to solve with Walker S2. The robot uses a dual battery system and can swap its own battery without human help. When it detects low power, it can walk to a swap station, remove the drained battery pack, install a fresh one, and return to work in about 3 minutes. That may not sound as exciting as a backflip, but for factories, it is a huge deal. It means Walker S2 can keep supporting inspection, logistics, assembly assistance, and material movement with far less downtime. Its 52 degrees of freedom help it bend, reach, and handle full body factory tasks, while UBTEK's co-agent system helps the robot with task planning, tool use, anomaly handling, and fleet coordination. And this is not just a future promise. UB has already delivered industrial humanoids to major partners like BYD, Gile Auto, FAW Volkswagen, Audi FAW, Foxcon, and SF Express. So, when the 1,000th Walker rolled off the production line, the battery swap demo became more than a cool feature. It showed China's larger strategy, building humanoids that can operate continuously inside real factories. Then comes Aptronic Apollo, and this robot sits in a very interesting middle ground. On one side, Apollo is already being tested for real factory work. On the other side, it is also becoming part of Google Deep Mind's bigger push into physical AI. The factory side is already serious.

[00:17:06]Mercedes-Benz is testing Apollo at its Berlin digital factory and its Kexat plant for tasks like moving parts, feeding assembly lines, and supporting quality inspection. Jabil is also using Apollo inside its own manufacturing network, which matters because Jabil is not just testing the robot. It can also help build and test the systems needed to scale more Apollo units. And Apollo's hardware is clearly designed around practical industrial use. It can carry up to 55 pounds, 25 kg, uses force torque sensing to move safely around people, has LED signals on its face and chest to communicate status, and uses hot swappable batteries so it can stay useful across long shifts. But the bigger AI story comes from Google Deep Mind. Gemini Robotics works like the action model, helping the robot turn visual information and human instructions into movement. Gemini Robotics-er works more like the reasoning layer, helping the robot plan more complex physical tasks. In Google's demos, Apollo can follow natural language instructions and handle sorting style tasks. And that is exactly what factories eventually need, a robot that can see the job, understand the goal, and act without every single motion being manually programmed. And then there is Agility Digit. Compared to some of the others, Digit may look like the least dramatic robot here, but it may also be one of the most proven. It does not need a viral catwalk, a human versus robot challenge, or a flashy stage demo to make its point. Digit has already been working in warehouses since 2024.

[00:18:35]Its job sounds simple. Move totes, feed lines, carry items, and handle repetitive logistics work. But in real operations, those simple tasks are exactly where companies lose time, money, and human energy every day. Digit uses neck lighter, chest cameras, and force torque sensors in its feet to move through busy warehouse spaces, understand where it is stepping, and stay balanced while handling objects.

[00:18:58]Agility also makes the business side easier with a robots as a service model, where companies can rent Digit at about $30 per hour fully loaded instead of buying the whole system upfront. And the realworld traction is already visible.

[00:19:13]Toyota Motor Manufacturing Canada expanded from 3 to 10 digit units after a year-long pilot. Schaffler is using Digit in Germany to feed 25 lb 11 kg totes into washing machines. Marcato Libre has signed on for Texas fulfillment centers and GXO helped prove the commercial warehouse use case early.

[00:19:32]That is why Digit's boring success matters so much. While some humanoids are still trying to prove they can work, Digit is already clocking in. Would you rather bet on Apollo's AI powered future with Google Deep Mind or Digit's simple but proven warehouse work? Share your pick in the comments. So, here's the bigger shift. The humanoid race is no longer about one perfect demo or one viral clip. It is about which robots can survive real work, repeat boring tasks, reduce downtime, and earn a real place on the factory floor. America is pushing deeper AI systems, premium factory pilots, and warehouse deployment. China is pushing speed, scale, uptime, and vertical integration. Optimus is entering the part of the journey where the factory matters as much as the robot itself. Tesla is preparing pilot production at Fremont, not just to build a few machines, but to test the process that could eventually support a much larger ramp. That shift also comes with a major farewell. The Model S and Model X lines, which once represented Tesla's rise from startup to serious automaker, are now being moved aside for Optimus Manufacturing. And the buildout does not stop in California. Tesla's automation teams in Germany, Toronto, and the Midwest are working on the specialized machines that could make large-scale humanoid robot production possible. The first real signal is coming from Fremont, where new footage appears to show Tesla preparing the early production setup for Optimus. A Basinor report says the clip looks like the pilot line for Tesla's humanoid robot, which means this is not the final high-volume factory running at full speed yet. A pilot line is where Tesla proves the process, fixes assembly problems, tests tooling, and learns how to build Optimus more efficiently before pushing towards serious scale. Tesla's own Q12026 shareholder update supports the cautious version of this update because it lists California Optimus and Texas Optimus under construction while also warning that installed capacity is different from actual production. So this is not proof that Tesla is already building 1 million Optimus robots a year. It shows the ramp is taking shape with Fremont being prepared for a line designed around that future target. But Fremont is not just another factory space Tesla found. These are the same lines that helped make the Model S and Model X two cars that changed how people looked at electric vehicles. The Model S proved an EV could be fast, premium, and genuinely desirable. While the Model X pushed Tesla into luxury SUV territory with falcon wing doors, a massive glass windshield, and family practicality. Now Tesla is giving both programs a farewell, not because they failed, but because the company's center of gravity has moved toward autonomy, robo taxi, and Optimus. This moment works like a handoff. Tesla is turning one of its most historic EV production areas into the launch pad for its humanoid robot future. If Tesla reaches its goal of producing 1 million Optimus robots a year, how much do you think that could change the world? Comment below. Fremont may be where Optimus gets assembled, but the production backbone reaches far beyond California. Tesla VP Lars Moravi showed real footage from automation groups working in Germany, Toronto, and the Midwest. And these teams are building the advanced production systems that make robot manufacturing possible.

[00:22:39]In Germany, Tesla Automation, formerly Groman, brings deep experience in highly automated lines for batteries, electronics, and complex assembly. In Toronto, the former HighAR team adds high-speed precision work that can support actuator, joint, and electronics assembly. The Midwest adds another layer of manufacturing know-how for specialized production equipment.

[00:22:59]Optimus is not a simple product. Every joint, sensor, cable path, and hand component needs extreme accuracy if Tesla wants to build humanoid robots at serious scale. The next layer is the computing power inside Optimus. And this is where Tesla's AI5 chip becomes a major part of the plan. Tesla has completed the tape out for AI5, which means the chip design has reached its final stage before samples are produced.

[00:23:22]The expected jump is huge with roughly eight times the compute of AI4 and about nine times the memory. Musk says one AI5 can perform like an Nvidia H100 while a dual setup could reach Nvidia Blackwell class performance at lower cost and power. Samples are expected in late 2026 with volume production around mid 2027.

[00:23:42]AI4 may already be strong enough for cars, but Optimus needs more local intelligence so the robot can see, balance, move, and react on its own without depending on Wi-Fi. The physical design is also moving toward a more productionready robot. At a keynote in Zurich, Tesla's Optimus program lead Constantinos Lasceris showed a slide with the Gen 3 silhouette and described it as the first mass manufacturable Optimus. That phrase points directly to Tesla's shift from research versions like Bumblebee and Gen 2 toward a robot that can actually be built at scale. The silhouette appears more human with thicker forearms, smoother lines, and more refined hands that look closer to a real working humanoid. Engineers have described the design as a human in a superhero suit. Tesla still has not given the full reveal because Musk says competitors study every frame carefully, but the direction is clear. Gen 3 is being shaped around usefulness, safety, reliability, and mass production. The real breakthrough in Gen 3 may come from the hand because that is the part that decides how useful Optimus can be in the physical world. Tesla's new hand has 22 degrees of freedom with 50 actuators spread across both forearms and hands, bringing the design closer to human level movement and control. Walking is only one part of a humanoid robot's job.

[00:24:55]To work in a factory, Optimus has to pick up delicate objects, handle small parts, tighten bolts, and adjust items without crushing or dropping them. That is why Tesla is putting so much focus on dexterity. A smarter brain helps the robot understand the task, but the hand is what lets Optimus actually do the work. If a robot could pick up delicate objects, tighten bolts, and work beside humans, would you trust it in a real factory? Comment below. Musk's bigger vision for Optimus goes far beyond one production line or one factory job. He has described Optimus as a possible vonoman machine, meaning a robot that could eventually help build more robots and expand production over time. That idea is extremely ambitious, but Tesla is starting with a more practical path.

[00:25:37]Optimus is expected to begin inside Tesla's own factories, taking on repetitive or dangerous tasks before any wider outside sale at an estimated price of $20,000 to $30,000. Musk has also claimed Optimus could one day generate more than $10 trillion and become bigger than Tesla's car business. That future is still a long way off, but Gen 3 is the first step toward testing whether the vision can become real. For that kind of scale, Tesla would need more than robot factories. It would need a massive and reliable supply of AI chips.

[00:26:08]That is where the reported Terraab plan in Texas becomes important. Tesla is said to be working with SpaceX on its own chip foundry with an investment starting around $55 billion and possibly rising much higher over time. The goal is to reduce dependence on outside suppliers and control more of the silicon stack behind FSD, Dojo, Optimus, Roboaxi, and Starlink. If Tesla wants millions of robots and autonomous vehicles running alongside huge AI systems, chip supply becomes one of the biggest limits. Terrafab would be Tesla's attempt to build that foundation from the ground up. Above the chip supply layer, Tesla is also building the intelligence layer that could decide how Optimus receives and understands instructions. Gro 5 is now being trained with versions reportedly targeting six and even 10 trillion parameters, while XAI's Colossus system in Memphis is built around more than 200,000 GPUs and is moving toward Nvidia Blackwell hardware. Musk has described Grock as an orchestration AI which means it could handle broad instructions while Optimus uses onboard compute for movement, balance, vision and real-time decisions.

[00:27:12]So the stack starts to connect clearly.

[00:27:15]FSD gives Optimus vision, Grock gives it higher level understanding. AI5 gives it local brain power and Gen 3 gives it the body and hands to act in the real world.

[00:27:24]That is why the production line, the chip plan, and the hand upgrade are all part of the same Optimus push. What sounds more powerful to you? AI5 inside the robot, Grock guiding the robot, or the new Gen 3 hands doing the real work.

[00:27:37]Tell me below. So, the real Optimus update is not just one robot walking into a factory. It is Tesla connecting production lines, custom chips, AI models, robot hands, and global automation teams into one bigger system.

[00:27:50]The boldest claims still need time because pilot lines are not the same as full-scale output, but the direction is clear. Tesla wants Optimus to move from a lab project into a real factory worker. And once one company starts building humanoid robots this way, the bigger race is no longer about demos. It becomes about who can scale useful robots first. While we wait for Optimus Gen 3, Figure AI shocked everyone this month. Figure just proved who's really winning the humanoid race. Figure AI built one AI brain, Helix O2, that controls the whole robot, fully autonomous. It revealed a seventh generation hand that can feel a touch as light as a paperclip. It unveiled Vulcan, a system that keeps Figure03 standing even after its parts fail. And it built a new factory that now cranks out one robot every hour. Figure04 is already designed and heading to the floor. While most of the robot world was busy waiting on Tesla, Figure quietly stacked five wins back to back. Let's go through all five, and by the end, you'll see exactly why this company is breaking away from the pack. First up, the brain.

[00:28:51]Figure calls it Helix O2, one AI model that controls the entire robot at the same time. The legs, the torso, the head, the arms, and every single finger, all run by one system instead of a stack of separate programs bolted together.

[00:29:05]Figure threw out more than a 100,000 lines of handwritten code, and replace the whole thing with a single neural network. The payoff is full autonomy.

[00:29:12]Figure 03 can now walk to a dishwasher, unload it, cross a full kitchen, stack the dishes into cabinets, and then load the dishwasher again. 61 decisions in a row, 4 minutes straight, with no human touching anything. Normally, that is exactly where robots fail. One small mistake early on collapses the whole task. Helix O2 keeps track of the job and adjusts in real time instead of freezing. And here's the other half.

[00:29:35]Helix O2 can feel. Touch sensors in the fingertips pick up a force as light as 3 g. That is the weight of a single paperclip. Would you trust a robot to handle your good glasses with a grip that gentle? Tell me in the comments.

[00:29:48]But the real proof came on a live stream. On May 14th, Figure put three Figure03 robots named Bob, Frank, and Gary on a live feed sorting packages in a San Jose warehouse. The plan was 8 hours, just enough to prove one full shift. 8 hours passed. The stream kept running. 24 hours, 72 hours. Then day 7, day 8, day 9, still going. In the first 72 hours alone, the three robots sorted around 88,000 packages with no logged failure. Gary alone sorted around 10,000. When Gary's battery ran low, Gary walked to the charging station alone, and another figure 03 took over the line. No human stepped in. Figure even ran a man versus machine contest, and a human worker barely won by less than a tenth of a second per package.

[00:30:34]You cannot fake any of that with editing. A short clip can hide the bad runs. A 9-day live stream shows every slip the moment it happens, and the slips barely came. Second move, the hand. And this is the one figure is most proud of. The big idea here, dexterity is the single hardest problem in all of robotics. Plenty of companies can make a robot walk. Almost none can make a robot's hands work like yours. So, while everyone else chased better walking, Figure spent three quiet years on hands.

[00:31:01]The new one is a seventh generation hand on a third generation body. Read that again. The hand is four generations ahead of the robot wearing it. Each finger moves on its own. The thumb rotates across the palm and can reach the tip of every finger. the same trick that makes your hand so capable. There are more than 20 degrees of freedom packed in palm cameras so the robot can see what its own fingers are doing even when the head can't and fingertip sensors that feel that same 3 g. For comparison, Tesla's Optimus hand sits around 11° of freedom today. Figure is past 20. If a robot can feel pressure lighter than a paperclip, would you let it handle glass in your kitchen? Drop your answer below. Third move, reliability. This one is called Vulcan.

[00:31:41]A robot that does not die when one part breaks. For years, the rule was brutal.

[00:31:46]A single bad joint in a humanoid's lower body meant the whole machine lost balance and collapsed. Task over. A human had to walk in and drag it off the floor. Vulcan ends that rule. With Vulcan running, figure 03 can lose up to three joints or actuators and still keep standing, keep walking, and keep working. Figure proved it live. An engineer triggered a fake knee failure.

[00:32:06]While figure 03 was sorting packages, the robot did not fall and did not freeze. Figure 03 shifted its balance, changed how it moved and limped.

[00:32:15]Actually limped like an injured person straight to the maintenance area on its own. A robot beside you loses a knee and just keeps working like nothing happened. Cool or creepy? Let me know in the comments. And here's why this matters so much. Figure wants lights out factories where robots run all night with nobody watching. That only works if a robot can handle its own problems instead of stopping the entire line every time something goes wrong. While Figure03 is outsorting real packages, Figure has already finished designing the next robot. On May 13th, CEO Brett Adcock announced that the F.04, the fourth generation figure robot, has hit design lock. In plain terms, the design is finished and frozen. The parts are being ordered and the robot is moving into pre-production. Adcock says F.04 04 is the biggest jump between generations the company has ever made with the engineering pushed to a whole new level.

[00:33:06]Now think about the pace here. Figure 01, then O2, then 03, and now a locked in O4. While figure 03 itself is barely out the door. Most companies are still trying to perfect their first humanoid.

[00:33:18]Figure is already building its fourth, fifth, and final move, the factory.

[00:33:22]Because a brilliant robot you cannot build does not change anything. At its bot Q plant in California, Figure went from making 1 figure 03 a day to 1 figure 03 every single hour in under 120 days. That is a 24 times jump in 4 months and more than 350 figure 03 robots have already shipped out the door. The numbers behind it are wild.

[00:33:43]Custom factory software running across more than 150 workstations. Over 50 inspection points along the line. End of line first pass yield already above 80% and climbing every week. The battery line hitting 99.3%, over 9,000 actuators built, and every single robot gets pushed through more than 80 tests before signoff, including burn-in sessions of thousands of squats, shoulder presses, and even jogging just to catch weak parts before they ever leave. This is the difference between a science project and a real company.

[00:34:14]Anyone can show one robot in a video.

[00:34:17]Building one an hour is a completely different sport. So, put all five together. The brain with Helix O2, the hands four generations ahead, the reliability of Vulcan, the next robot already locked in, and a factory pumping out one humanoid an hour. That is not one lucky breakthrough. That is the whole stack all moving at once. And that is exactly why Figure is pulling ahead of everyone right now. Which of these five updates impressed you the most?

[00:34:41]Tell me in the comments. After a 200 hour YouTube live stream where Figure03 sorted packages in real time, Figure is now preparing its humanoid robots for massive retail warehouse deployment.

[00:34:52]Then Limx Luna walked onto a runway and proved humanoids can also become public-f facing performers. This is the clearest sign yet that humanoids are leaving the demo phase and entering realworld deployment. But Boston Dynamics had a different message as Atlas started learning football and Unitry just showed a robot that can understand, clean, and adapt inside a messy human space. Figure AI just made a massive move that could reshape the future of American retail. On May 26th, 2026, the company signed a commercial agreement with Catalyst Brands to deploy humanoid robots across its distribution network. The first rollout will begin at Catalyst's Reno, Nevada distribution logistics center. And this is a big step because Catalyst is only Figur's second major commercial customer after BMW. So why is this deal such a big deal?

[00:35:38]Catalyst Brands is not a small retail company testing robots for fun. It was formed in early 2025 after Spark Group and JC Penney merged together. Today Catalyst controls some of the most recognizable American brands including JC Penney, Aerop Pastel, Brooks Brothers, Eddie Bower, Lucky Brand, and Nautica. Altogether, the company operates around 1,800 stores and has nearly 60,000 employees across the US and Canada with major backing from Brookfield Corporation and Simon Property Group. But the most important part is what Figures robots will actually do. Its next generation humanoids, likely the Figure03 model, are expected to work inside the Joey Pouch sorting system, where they will help with the difficult sorting and packing tasks that normally put heavy physical pressure on warehouse workers.

[00:36:24]Catalyst says this will allow human associates to move toward higher value work instead of spending long hours on the most exhausting warehouse jobs. And the timing makes this even bigger.

[00:36:34]E-commerce has created serious pressure on retail warehouses and companies across the country are struggling with labor shortages. Figure says its robot can already sort packages at near human speeds and its demos have shown more than 200 hours of autonomous operation.

[00:36:48]Now look at the money behind figure. The company raised $675 million in series B funding in 2024 at a $2.6 billion valuation. Then in 2025, a $1 billion series C pushed its valuation close to $39 billion. Its investors include Nvidia, Microsoft, Intel Capital, and Jeff Bezos. Figure even moved away from OpenAI to build its own Helix vision language action model. And behind all of this, Brookfield's connection to both Figure and Catalyst may be the key. If the Reno deployment succeeds, Catalyst could eventually expand Figure Humanoids across its massive retail network, creating a new blueprint for how major retail companies use humanoid robots.