Unitree’s New $16,000 Robot Just Proved Robots Can Become Shockingly Human…

Added: 2026-05-05

[00:00:04]The Unitree G1 has just demonstrated a level of control and versatility that shows how far this technology has already advanced. The Unitree G1 recently moved from rolling on wheels to gliding on rollerblades to skating across a full ice rink, all using the same hardware and control system in real recorded footage. And that was only the beginning. More recently, Unitree's H1 reached speeds close to 22.4 mph, 10 m/s, on a real athletics track. Around the same time, Unitree humanoid robots, including both the H1 and G1, appeared in a full 13.1 mile, 21 km, half marathon in Beijing alongside human runners. Taken together, these moments show how quickly humanoid robots are improving and how fast the gap between robotic and human movement is beginning to shrink. Let's start with the video that grabbed so much attention online.

[00:01:09]The Unitree G1 is built as a compact and agile humanoid platform designed for balance, movement, and adaptation in real environments. Instead of focusing only on strength, Unitree engineered this robot to handle unstable surfaces and changing conditions through a flexible control system. In early 2026, the company released footage showing the G1 performing something unusual, not by moving faster, but by showing an unexpected level of balance. The robot rolled forward on motorized wheels [music] mounted at its ankles, staying upright without any external support. At one point, it lifted one leg completely off the ground and continued moving on a single wheel using upper body adjustments to stay balanced [music] in a way that closely resembled human movement. Because the footage was recorded outdoors on real asphalt, Unitree clarified that it was not sped up or artificially generated, anticipating skepticism. The challenge then increased. The G1 switched to inline rollerblade frames, similar to those used by humans, which reduced sideways stability and made balance much harder. In this setup, every movement required precise control because even a small shift in weight could lead to instability. The robot responded with constant posture corrections using subtle arm motion and rapid adjustments to stay centered while moving forward.

[00:02:45]In one sequence, it glided on a single rollerblade with one leg raised, maintaining a steady path without visible wobble. That level of control is difficult even for experienced human skaters. Do you think humanoid robots will ever outperform Olympic athletes or is there still a limit they cannot cross? Tell me what you think in the comments. The most demanding transition came when the robot moved onto ice skates. A thin [music] steel blade offers almost no lateral support, which means balance depends on extremely precise coordination across the entire lower body. The G1 adapted by lowering its center of gravity and adopting a speed skating posture with bent knees and a forward lean. As it moved across the ice, its upper body shifted with each stride, mirroring the motion human skaters use to maintain momentum and stability. What made this sequence important was that the robot itself did not change. Only the foot attachments changed while the same system handled three completely different balance conditions. This adaptability shows what Unitree is really building. The G1 does not rely on separate fixed [music] programs for each surface. Instead, it uses a generalized balance system that processes motion, position, and acceleration data in real time and calculates how to respond. Additional sensors [music] help it detect terrain changes and adjust before losing balance. That means the robot is not repeating a fixed routine, but actively adapting to its environment, which is essential for real-world use. Real-world robot movement is advancing fast and we explain what it actually means.

[00:04:38]Subscribe to the AI Nexus and don't miss what's coming next. This demonstration also points to a larger goal focused on speed and real-world mobility. That is where the focus shifts to Unitree's H1, which is designed to push how fast a full-size humanoid can move. In testing footage released around the time of the Beijing half marathon, the H1 accelerated down an athletics track and reached 22.6 mph, 10.1 m/s, >> [music] >> with timing equipment confirming the result. That places it close to the average speed Usain Bolt maintained during his 100 meter world record, around 23.3 mph, 10.44 m/s. Company statements also confirmed similar peak speeds in sprint tests, [music] placing the robot near a level that once seemed out of reach. What makes this more significant is how [music] quickly progress has happened. Earlier versions of the same platform had already set records, but were still far from human-level speed. In March 2024, the H1 set a Guinness World Record at 7.4 mph, 3.3 m/s. At the time, that was a major milestone, but reaching over 22 mph in sprint testing is not a small improvement. It represents a major leap in locomotion capability. Once that level of speed and balance became reliable, the next step moved beyond controlled demonstrations. It moved into real-world competition. That happened at the Beijing E-Town Half Marathon, where humanoid robots and human [music] runners took part on the same route while staying in separate lanes. The course covered 13.1 miles, 21 km, and included varied terrain, sharp turns, and elevation changes of about 328 feet.

[00:06:38]These conditions required sustained stability rather than short bursts of performance. The robot cut-off time was 3 hours and 40 minutes, meaning only systems capable of endurance and consistent navigation could finish. Many teams chose full autonomy, requiring robots to handle the entire course without human control. If a robot is already running close to Usain Bolt's speed, what happens when it goes beyond that? Tell me your thoughts in the comments. The Unitree H1 stood out as one of the most widely used platforms in the event. Multiple teams used it in the autonomous category and official race coverage documented several H1 units on the course. At least two teams were photographed crossing the finish line, confirming that the platform completed the full 13.1 mile, 21 km, course autonomously. This required the robot to manage real turns, inclines, and terrain without guidance, showing that it can perform beyond short demonstrations.

[00:07:47]Before the main race, the H1 had already shown strong performance during qualifying. Viewed together, these results show a clear pattern. The G1 demonstrated that Unitree's control system can adapt across different physical surfaces without needing to be redesigned each time. The H1 extended that progress into speed and endurance, showing that the same engineering approach can scale to full-size humanoids in real environments. These are not isolated demonstrations, but part of a larger progression toward more capable robotic movement. This progress is also supported by Unitree's growth as a company. In 2025, it shipped more than 5,500 G1 units, reporting 335% revenue growth year over year. The company is targeting 20,000 humanoid shipments in 2026. Around the same time, it open-sourced its vision language action model, allowing robots like the G1 to perform tasks through natural language commands. This shows a focus on both hardware and software development.

[00:09:01]Does the idea of robots improving this fast excite you or does it make you uncomfortable? Share your thoughts in the comments. These developments reflect a much larger shift in humanoid robotics, especially in China, where production, public demonstrations, and real-world testing are moving forward at a rapid pace. The Beijing Half Marathon was not just a race. It was a clear snapshot of what these systems can already achieve outside controlled environments. What stands out is the progression itself. A robot balancing on a single wheel, adapting across rollerblades and ice, reaching near-human sprint speeds, and then completing a full outdoor half marathon autonomously, all of this is now happening in real time with results that can be measured and verified. And if this is where humanoid mobility stands today, then what comes next is not a distant possibility. It is already beginning to take shape because this is not just about one robot or one company.

[00:10:08]It is part of a much bigger shift that is unfolding across the entire industry.

[00:10:14]Humanoid robots are entering sports and some are already matching human-level athletic performance.

[00:10:20]>> [music] >> From 90% accurate tennis rallies to skateboarding, combat moves, and kung fu at Shaolin Temple. This is happening right now. Researchers from Tsinghua University, Peking University, and robotics company Galbot developed a system called Latent, which stands for learning athletic humanoid tennis skills from imperfect human motion data. This system was tested on the Unitree G1 robot, a compact bipedal platform that stands around 4 ft tall. What makes this work stand out is that the training data was not clean or perfect. The recordings were noisy, incomplete, and often inconsistent. Yet, the system was able to learn the core patterns of human athletic movement and map them accurately onto the robot's joints. On a real tennis court, the Unitree G1 achieved a forehand return success rate of 90.9% while backhand returns reached 77% accuracy. The robot was not limited to single shots as it maintained continuous multi-shot rallies with human players while tracking balls moving at speeds above 15 m/s. When you consider how little reaction time that allows, the level of control and timing becomes extremely impressive. If a robot can hit 90% accurate tennis shots, can it beat a pro player today? Comment below your opinion. Now, let's move into combat where Engine AI is pushing humanoid robots into organized fighting systems.

[00:11:53]The company has already announced plans for a robot boxing tournament, showing that this technology is moving beyond demonstrations and into structured competition. At CES 2026, Engine AI introduced the T800 and revealed something no other company had shown publicly as the robot physically kicked the CEO to the ground during a live test. [music] Founder Zhao Tianyang recorded the entire moment in a plain studio without special lighting or post-production effects, making the demonstration clear and direct. The T800 runs on a fully integrated joint system that delivers up to 450 Nm of peak [music] torque and 14,000 W of instantaneous joint power, allowing it to perform high-impact movements with strong control. The robot has been shown executing flying kicks and breaking through doors while maintaining balance, force accuracy, and precise coordination. Engine AI has confirmed that all footage is real with no CGI or speed changes, making this one of the clearest demonstrations of controlled robotic combat.

[00:13:04]>> [music] >> Following that, attention moves towards speed and natural motion with KAIST in South Korea. The humanoid version 0.7 is fully developed in-house, including its motors, gear systems, and control electronics, giving the team full control over performance. The robot can generate up to 320 of torque at the knee joint and reach running speeds of nearly 13 km/h while staying balanced and stable. Instead of relying on simple movement patterns, the system learns from human motion capture data, allowing it to reproduce actions with smooth timing and coordination.

[00:13:43]This is clearly visible when the robot performs a fluid moonwalk, showing how well it understands rhythm and balance.

[00:13:51]The focus here is not just speed, but building movement that closely reflects natural human motion. After that, the same Unitree G1 demonstrates [music] a completely different skill through skateboarding. Researchers from China Telecom's AI Institute, along with Shanghai Jiao Tong University and partner teams, created the Husky framework, a physics-aware whole-body control system. Training runs inside a MuJoCo simulator with more than 4,000 parallel environments, allowing millions of motion trials to be completed in about 20 hours. Through this process, the robot learns to generate momentum, control direction using body balance, and step on or off the board across different surfaces. The system works without external support or human guidance, making it a fully learned behavior. This marks the first real-world skateboarding demonstration achieved entirely through reinforcement learning. Back in China, the Unitree H2 demonstrates precision under extreme motion in a much more demanding way. The robot stands around 180 cm tall and is built with 31 joints, giving it a wide range of movement and control across the entire body. In one demonstration, the robot performs a full 360° aerial spinning kick, striking a suspended watermelon in midair before landing cleanly on both feet without losing balance. This is not designed as a performance clip, but as a controlled stress test that pushes the robot's sensors, actuators, and real-time control systems to their limits. During such fast and high-impact movements, even a small timing error can cause instability. Yet, the robot maintains strong balance and coordination throughout the motion. This level of accuracy under dynamic conditions shows how stable and responsive humanoid systems are becoming in real-world scenarios. A clear example of power-focused design comes from Fybot [music] with the M1 humanoid. This robot is built to deliver high levels of physical output while maintaining control during complex movements.

[00:16:07]Standing about 5.8 ft tall and weighing under 60 kg, it can generate more than 10 kW of power in short bursts. Its joints can produce up to 530 Nm of torque, allowing the robot to perform demanding actions with strong force and stability. The M1 has demonstrated a full standing backflip with a controlled landing, which requires precise timing and coordination across the entire body.

[00:16:36]With 32° of freedom and swappable batteries that support over 2 hours of operation, this system is designed for real physical work, not just controlled demonstrations. Skateboarding, backflips, and aerial kicks, are robots becoming more athletic than humans?

[00:16:54]Comment below. The REI Institute, led by Marc Raibert, is developing a new approach to athletic movement using a robot called RoadRunner. This system combines both legs and wheels, allowing it to adjust how it moves depending on the terrain it encounters. Weighing around 15 kg, the robot can walk, roll forward smoothly, stand up from the ground, and even balance on a single wheel when needed. What makes this system important is that all of these movements are performed without task-specific [music] training, meaning the robot can execute them directly without learning each action separately. This highlights how general movement intelligence is improving, bringing robots closer to flexible and adaptive real-world mobility. The focus returns to sports with a basketball system built on the Human X framework from Hong Kong University of Science and Shanghai AI Laboratory.

[00:17:51]Using this system, the Unitree G1 performs actions like dribbling, shooting, and turning with control similar to a trained player. The most important part is how the robot senses the ball during movement. In a special operating mode, it does not rely on cameras or external tracking systems.

[00:18:10]Instead, it uses internal feedback from its joints to detect motion and force, similar to how a human feels the ball while dribbling. This method achieved over 80% success in complex moves such as pump fake turnarounds, showing how perception is evolving beyond traditional vision systems. Agibot takes this even further with the A3 humanoid by applying full-body coordination to martial arts. The robot was taken to the Shaolin Temple where it performed kung fu sequences alongside [music] trained practitioners. It is powered by the Genie Operator 1 AI model, which allows the robot to execute movements without needing prior training for each specific action. This approach, known as zero-sample generalization, enables the system to adapt quickly to new motion patterns. The robot had already demonstrated advanced movements like the Webster flip, which requires precise control during airborne motion. This shows how humanoid robots are beginning to learn coordinated, expressive movement that closely reflects human physical behavior. Would you train with a robot at Shaolin Temple if it could match human-level kung fu? Comment honestly. Supporting all of these advancements is a foundational system designed for long-term development. Tian Gong 3.0, developed by Beijing's humanoid, is a full-size humanoid platform built for high-dynamic movement and full-body control combined with tactile interaction. It operates on the Wise Kai Wu AI system, which connects perception, decision-making, and physical execution into a continuous loop. The robot can navigate obstacles, perform fast sequences of motion, and operate with very high precision. A key aspect of this platform is that it has been open-sourced, including its mechanical design, control systems, and training tools. This allows researchers and companies to build on top of it, turning humanoid robotics into a scalable ecosystem for real-world athletic AI.

[00:20:19]This level of progress shows one thing clearly. Humanoid robots are no longer limited to controlled environments. They are learning real-world movement at an incredible pace, from sports to combat to full-body coordination. This is just the beginning of what physical AI can achieve. And if you think this is impressive, the next development takes it even further. A humanoid desktop robot just held a full conversation with its own creator, and the interaction looked shockingly human. A head forms new desktop humanoid robot just sat face-to-face with its own creator, CEO Yuhang Hu. The robot made eye contact, blinked, [music] smiled, and responded to questions in real time, giving one of the clearest glimpses yet of where social robotics is heading. Watch this.

[00:21:18]After watching that exchange, what stands out immediately is how natural the interaction actually feels. CEO Yuhang Hu sits directly across from the robot and begins speaking to it the same way he would talk to another person. He asks the robot simple but personal questions about its personality, about how it sees itself, and about how it interacts with people. The robot responds calmly and clearly, holding eye contact while speaking. As the words come out, the lips move in perfect sync with the voice, just like a human speaking in conversation. The robot keeps its eyes focused on Hu's face the entire time. And when Hu shifts slightly in his seat, the robot's gaze follows him smoothly. The vision system is clearly tracking him throughout the interaction, making the moment feel less like talking to a machine and more like sitting across from another person. What makes the exchange even more surprising is how expressive the robot becomes during the conversation. When Hu asks light and playful questions, the robot's face relaxes and a small smile begins to form. The eyes soften and the expression becomes warmer. When Hu asks the robot to show different emotions, the face changes instantly. Confusion appears as the eyebrows lower slightly and the eyes narrow. When he asks for surprise, the robot's eyes widen and the [music] expression shifts quickly. Happiness, sadness, curiosity, and even anxiety all appear as different facial reactions, each one looking distinct and surprisingly natural. At one point, Hu makes a playful remark and the robot answers back with a light joke of its own. As it delivers that response, the corners of its mouth slowly lift and the eyes narrow slightly, forming a natural smile that rises and fades with the rhythm of the sentence. Instead of looking mechanical or exaggerated, the expression appears timed and controlled in a way that feels remarkably human.

[00:22:57]What was the most surprising moment in the clip you just watched? The conversation, the expressions, or the eye contact? Comment below and tell me.

[00:23:05]Another detail that makes the interaction feel surprisingly real is the way the robot pauses before answering. When Hu finishes asking a question, the robot does not respond instantly like a typical chatbot.

[00:23:16]Instead, it's gaze lowers slightly and the eyelids blink once, almost like a person briefly thinking about what to say next. Then it lifts its head and gives the answer. That short pause may seem small, but it adds a natural rhythm to the conversation and makes the exchange feel far more human. So, what is actually happening behind the scenes to make this possible? Inside the A head form robot [music] are two main systems working together at the same time. The first system is the artificial intelligence that handles the conversation itself. This AI listens to what Hu says, understands the question, and then generates a new response in real time. Instead of repeating fixed lines or pre-written answers, the robot is able to respond differently depending on what is asked during the conversation. The second system controls the robot's face and physical expressions. Hidden beneath the synthetic skin are many tiny brushless motors that move different parts of the face. These motors control the eyebrows, eyelids, lips, and cheeks at the same time. Because these movements are very small and precise, the robot can create subtle expressions instead of large robotic movements. Those tiny changes are what make the face look believable when viewed up close. These two systems are connected through software that links speech and emotion together. When the AI generates a response, the system also analyzes the tone of that response.

[00:24:33][music] If the robot is speaking playfully, the face may show a small smile. If the response sounds thoughtful or uncertain, the expression changes accordingly. This is why the robot's voice and facial expressions stay synchronized throughout the conversation. Hey guys, every week we break down the most fascinating AI and robotics breakthroughs. From humanoid conversations to machines that can show real expressions. Subscribe to AI Nexus so you never miss what's coming next.

[00:24:58]The mouth movements are especially detailed. As the robot speaks, different lip shapes appear for different sounds, allowing the mouth motion to match the words almost perfectly. At the same time, the neck motors add another layer of realism. The robot tilts its head slightly toward the speaker, nods during the conversation, and shifts its position subtly. These are small behaviors that humans naturally use during communication, but in this robot, every movement is produced by motors and software working together. A head form was founded in Shanghai in 2024 by robotics researcher Yuhang Hu, who earned his PhD from Columbia University.

[00:25:32]The company focuses on building expressive humanoid faces designed for natural human interaction. Instead of factory work, A head form's robots are built for conversation, sitting at eye level on a desk to create a more personal and comfortable interaction with people. This idea may sound simple, but it could change how people interact with technology. Studies in human-robot interaction show that people respond very differently when a machine looks at them, reacts to their emotions, and shows expressions. Talking to a voice assistant on a phone still feels like using software, but when a robot sits across from you, makes eye contact, blinks, and smiles while speaking, the experience becomes much more personal.

[00:26:07]That is exactly the direction A head form is exploring. Instead of building robots that lift heavy objects or work in factories, the company is building robots designed for human interaction, machines that can sit on a desk, hold conversations, and become part of everyday life. And A head form is not the only company working toward this idea. Around the world, many robotics teams are now building robots designed mainly for social interaction. These machines are not meant to replace factory workers or warehouse labor.

[00:26:34]Their purpose is conversation, companionship, and human connection.

[00:26:39]Companies from China, Japan, Europe, and the United States are all developing new humanoid robots that focus on communication and emotional interaction, rather than physical strength. One example is EVA.i, a social robot created by IO Technology.

[00:26:56]EVA.i is designed with electronic skin that can detect touch and pressure, allowing the robot to react when someone interacts with it. The system even includes temperature control technology that keeps the surface of the robot warm, close to normal human body temperature. Early versions of EVA.i are expected to sell for around $6 to $7,000, targeting people interested in AI companionship. Which future do you think is more exciting? Robots that talk and show expressions like A head form's robot, or robots that feel human-like like EVA.i? Tell me in the comments below. Another project comes from famous Japanese roboticist Hiroshi Ishiguro.

[00:27:34]His team recently introduced a humanoid robot called Avida at Mobile World Congress 2026. Avida uses realistic silicone skin and precise servo motors to create lifelike facial expressions.

[00:27:45]The robot is also expected to gain walking ability through a partnership with the robotics company Unitree. And in Shanghai, another humanoid robot called Moya has already shown how far social robotics has progressed. Built by the company Droid Up, Moya stands about 1.65 m tall and can walk with a movement pattern that is more than 90% similar to human walking. The robot also uses facial micro-expressions to make its interactions feel more natural during conversations. All of this shows something important about the direction robotics is taking. The industry is now moving in two different paths. One path focuses on robots designed for physical work. Machines like Tesla's Optimus, Figure 03, and NEO Gamma are being built to perform tasks in factories, warehouses, and eventually inside homes.

[00:28:29]The other path focuses on robots designed for social interaction. Robots like the A head form desktop robot, EVA.i, Avida, and Moya are not designed to lift heavy objects. Their purpose is communication, talking with people, listening, reacting, [music] and creating a sense of presence during interaction. Do you think the future will be dominated by robots that work for us, or robots that interact with us?

[00:28:52]Share your thoughts in the comments.

[00:28:54]Work robots are being developed to handle physical labor and improve productivity. Social robots, on the other hand, may help address problems like loneliness, emotional support, and easier communication with technology.

[00:29:05]What A head form demonstrated with its desktop robot offers an early preview of that future. A machine that can hold eye contact, read emotional cues, and respond in a way that feels more natural and human. In the coming years, robots may not only help us complete tasks, some of them may sit across from us, talk with [music] us, and slowly become part of everyday life. Social robots are developing faster than ever, and what we just saw might be an early glimpse of a much bigger future.