La Chine vient de construire un robot IA à deux cerveaux : un corps, deux esprits

China revealed a compact 4-ft robot with a split AI brain or dual brain architecture. Vietnam showed two new different humanoids for homes, security, factories, [music] and remote work. And NVIDIA just gave researchers a full 6-ft robot platform with over 2,000 teraflops of onboard AI power. The robot race is heating up again, so let's talk about it. So, let's start in Shanghai because Jaka Robotics just unveiled a humanoid called Pi, and it is a very different kind of robot from the massive human-sized machines we usually see in this space. Jaka Pi stands 4-ft tall or 1.22 m and weighs around 92 lb or 42 [music] kg. Its full dimensions are around 4-ft by 16.5 in by 8.7 in or 1,220 [music] mm by 420 mm by 220 mm. So, this is a compact humanoid built for research, development, and real-world testing. The company says Pi is designed as a versatile platform for embodied intelligence, motion control, AI-powered interaction, and human-robot collaboration. What makes Pi [music] interesting is the way Jaka is trying to combine the brain side and the body side of robotics. The robot has 27 degrees of freedom, and it uses newly developed integrated joint modules that are 15 to 27% smaller than Jaka's previous generations. That is a big deal for a compact humanoid [music] because smaller joints help the whole body become lighter, cleaner, and easier to move. The knee joints can deliver up to 88.5 lb ft of torque >> [music] >> or 120 Nm, giving the robot enough strength for stable walking and controlled locomotion. Each arm can carry up to 6.5 lb or 3 kg, so Pi can handle objects, test manipulation tasks, and work through research scenarios where the robot needs to interact with the physical world instead of just walking around as a demo. Then there is the fusion brain architecture. Jaka built the core system on Intel's heterogeneous computing platform, and the idea is to separate high-level intelligence from low-level motion control. The company describes the higher-level side as the cerebrum. That part handles AI reasoning, vision perception, large language models, application logic, and the general understanding of what the robot is supposed to do. The lower-level side works more like a cerebellum. That part controls real-time movement through an EtherCAT-based [music] control network with millisecond-level latency. That separation matters because humanoids need two very different types of intelligence [music] at the same time. The robot has to understand spoken instructions, recognize the environment, decide on an action plan, and then move its body with stable timing and coordination. [music] Jaka is trying to give Pi the ability to hear a request, understand what is happening around it, plan a physical response, and execute that response through deterministic control systems that stay reliable at the movement level. This also shows how Jaka is expanding beyond its original identity. The company was founded in Shanghai in 2015 and became known for collaborative robots, or cobots, used in industrial automation. Its Zoo series includes models like the Zoo 3, Zoo 5, Zoo 7, Zoo 12, Zoo 18, Zoo 20, and Zoo 30 for jobs such as assembly, machine tending, palletizing, [music] and packaging. Then it has the Pro series, including Pro 5, Pro 12, [music] and Pro 16, built for harsh industrial environments with IP68-rated protection against [music] dust, oil, and water. For jobs that need careful physical contact, Jaka offers the S series, including the S5 and S12 with high accuracy force sensing and advanced force control. The AL and A series combine robotic manipulation with machine vision, which makes deployment easier in dynamic production environments. The company also has Mini Cobot and Mini 2 for education, research, hospitality, and small-scale automation, plus Jaka Lens 2D, Jaka Lens VPS vision systems, six-axis force sensors, RoboHub control platforms, and low-code programming tools. So, Pi is part of a much bigger shift for the company. Jaka already introduced K series humanoid platforms, including K1, K1L, and K1W. And now Pi brings large language models, machine vision, force control, real-time planning, and compact motion hardware into one new embodied AI platform. China has already started putting humanoid robots into postal logistics, >> [music] >> including automated parcel sorting in Guangzhou, and robots like Pi show how quickly Chinese robotics companies are moving from industrial arms into humanoid systems that can work in more flexible environments. Now, the second big development comes from Vietnam, and this one is interesting because Vin Big Dynamics has now officially entered the humanoid robotics race with Dino, the company's first humanoid robot. Dino made its international debut at ICRA 2026 in Vienna and Computex Taipei 2026. Vin Big Dynamics describes it as an intelligent humanoid robot built to serve as a versatile assistant for modern living environments. >> [music] >> The robot combines advanced AI, environmental perception, autonomous navigation, situational awareness, human-robot interaction, and dexterous manipulation in a single platform. >> [music] >> The company has not released detailed specifications yet, so we do not have exact height, [music] weight, or torque numbers. Still, the target use cases are already clear. Dino is being developed [music] for security and surveillance in urban areas, campuses, commercial spaces, and [music] service complexes, while also being built toward household assistance. That means Vin Dynamics is aiming at two very different categories at the same time, public-facing security and domestic [music] support. That is a difficult combination because a security robot needs awareness, navigation, long operating stability, and reliable interaction with people in open environments. A household assistant needs safer movement, more precise object handling, and the ability to deal with messy, unpredictable everyday spaces. Vin Dynamics says Dino uses a flexible arm architecture and an advanced manipulation system so it can interact with objects in dynamic environments and perform tasks that require careful handling. One of the most useful demonstrations came from Vinpearl Safari Phu Quoc, where Dino was tested as an autonomous robotic guide. In that pilot deployment, the robot operated in outdoor [music] conditions, used multilingual speech, responded to visitors through natural language interaction, and relied on real-time environmental awareness to answer [music] questions and guide people. That is a strong test for a service robot because outdoor spaces are full of movement, noise, crowds, lighting changes, >> [music] >> and unexpected behavior. According to Vin Dynamics, the deployment showed that Dino could keep interacting with people while functioning in unpredictable service environments. That part is important because humanoid robots can look impressive in a clean lab, while real service environments expose weak points in navigation, [music] speech interaction, perception, and timing. Alongside Dino, Vin Dynamics also showed the building blocks behind its humanoid program. One of them is the VDM 80 actuator joint, a compact motor that works like the robot's muscle system. It weighs less than 2.2 lb or 1 kg, runs on a standard 48-V power supply, [music] and supports industrial communication standards like CAN FD, RS485, and EtherCAT. It can reach speeds of up to 235 revolutions per minute, and it is designed for long-term reliability with an expected operational life of more than 10,000 hours. The company also presented a human-like robotic hand with 11 moving joints and six actively controlled degrees of freedom. Integrated force sensors help the hand control grip more accurately, which is necessary for delicate tasks and consistent object handling. When this hand, the actuator platform, and Vin Dynamics AI training data set are combined, the company gets the foundation for future humanoids that can learn practical skills and operate safely in real-world settings. Then we get another Vietnamese humanoid from the same larger Vin Group ecosystem, and this one is called VRH3. It comes from Vin Robotics, a Vin Group subsidiary, and it was also unveiled at ICRA 2026 in Vienna and Computex Taipei 2026. VRH3 is the company's third-generation humanoid robot, and it is aimed more directly [music] at industrial automation, material handling, operational support, and work in complex environments. It has more than 31 actuators for coordinated whole-body movement, giving it the ability to perform tasks that require dexterity, balance, and precision. The robot runs on two onboard edge computers, [music] so it can process sensory and operational data locally instead of relying only on remote systems. That is important for low-latency decision-making and real-time autonomous control, especially in industrial settings where delays can make a robot unsafe or useless. Vin Robotics says VRH3 can perceive its environment, interact with people, transport objects, handle payloads of around 13 to 17 lb or 6 to 8 kg, and perform assembly-related operations. Its sensing and control systems are designed to help it move through dynamic environments while staying [music] stable during manipulation tasks. A major detail here is that Vin Robotics says all key technologies were developed in-house. That includes the robot's mechanical architecture, real-time computing and communication infrastructure, electrical and electronic architecture, power distribution platform, battery management system, and full-body AI control framework. This kind of vertical integration gives the company more control over how the hardware and software work together, which can make a big difference when trying to improve balance, manipulation, energy use, and reliability over multiple generations. At ICRA 2026, Vin Robotics demonstrated VRH3's teleoperation system using motion capture technology built directly into a virtual reality headset. The operator could control the humanoid's movements in real time without needing extra external tracking equipment. That kind of setup could be useful for remote industrial operations, hazardous environments, maintenance work, and any scenario where a human needs to guide a robot from a safer distance while still using natural body movement. So, from Vietnam alone, we now have Dino from Vin Dynamics targeting security, service, guidance, and household roles, and VRH3 from Vin Robotics targeting industrial tasks, payload handling, assembly support, and teleoperated work. For a country building up its global humanoid presence, that is a surprisingly broad entry. And then there is Nvidia, which is approaching humanoid robotics from a completely different angle. Instead of launching one robot as a finished product, Nvidia introduced what it calls the first open humanoid robot reference design built on the Isaac GR00T development platform. The goal is to give researchers a full hardware and software foundation for building humanoid robots instead of forcing every lab to stitch together hardware, simulation tools, AI models, sensors, hands, middleware, and deployment systems from different vendors. This was announced at Nvidia GTC Taipei, and the reference design is built around a Unitree H2 humanoid robot. The Unitree H2 is a nearly 6-ft tall human-scale robot weighing around 150 lb. [music] The body has 31° of freedom, which gives it enough movement range for human-like motion and real-world testing. Nvidia paired it with dual Sharp Wave tactile five-finger hands, adding another 44° of freedom. All together, the platform reaches 75° of freedom. That is a huge number because the hands alone add a major layer of complexity. Walking is already hard. Full-body control is already hard. Add tactile five-finger manipulation, wrist cameras, [music] and hand-level control, and suddenly the robot can be used for far more advanced research than basic locomotion. The sensing setup includes a head-mounted stereo camera, wrist-mounted cameras for manipulation, and an inertial measurement unit for motion tracking. The robot can also support demanding physical [music] work with arm torque of up to 88.5 lb ft or 120 N m, leg torque of up to 265.5 lb ft or 360 N m, and payload support of up to 33 lb or 15 kg. The onboard AI brain is Nvidia's Jetson AGX Thor T5000 computing platform. This system uses a Blackwell-based GPU capable of delivering 2070 FP4 teraflops of AI performance. It also includes a 14-core ARM CPU [music] and 128 GB of unified memory, giving the robot enough onboard compute for real-time sensor processing, control, [music] and AI workloads. The software side is where NVIDIA's strategy becomes clearer. The platform is tightly connected to Isaac [music] GR00T, which covers the whole robotics workflow from data collection to simulation, [music] training, evaluation, and deployment. Researchers can use Isaac Teleyop to capture demonstration data, Isaac Sim and Isaac Lab for virtual training and testing, and Isaac ROS middleware to deploy trained policies onto physical robots. NVIDIA is also including open foundation models for humanoid reasoning and learning, so researchers can build robots that perform multitask behaviors [music] and gradually develop more advanced capabilities. Labs can adopt the full stack or take individual components and plug them into existing workflows. NVIDIA also says the same development platform will support the Unitree G1 humanoid, which makes it more accessible to institutions already using that popular research platform. Several major robotics institutions have already committed to using the reference design, including >> [music] >> AI2, ETH Zurich, the Stanford Robotics Center, and the Advanced Robotics and Controls Laboratory at the University of California, San Diego. Steve Cousins, the executive director of the Stanford Robotics Center, [music] said robotics moves fastest when researchers can build on open platforms, share code, and test ideas on real machines. The pace is getting serious now, >> [music] >> and the next wave of humanoid robots will probably come from the companies that can connect AI reasoning, sensors, hands, balance, and real-world training into one stable system. Also, if [music] you want more content around science, space, and advanced tech, we've launched a separate channel for that. Link's in the description. Go check it out. That's it for this one. Let me know what you think. Thanks for watching and I'll catch you in the next one.