A 4-ft Chinese robot with a brain split into two separate computing layers. A Vietnamese humanoid already tested as a live tour guide at a wildlife safari park. And NVIDIA handing researchers a 6-ft platform with 2,070 FP4 teraflops of onboard AI compute. All three came out in the same week. That is not a coincidence - it is a signal about where the humanoid robotics race is right now.
Jaka Pi: The Compact Chinese Humanoid With a Split AI Brain
Jaka Robotics, founded in Shanghai in 2015 and originally known for industrial collaborative robots, unveiled a humanoid called Pi at a compact 4 ft / 1.22 m tall and 92 lb / 42 kg. Its full footprint is 1,220 mm by 420 mm by 220 mm. This is not a human-scale machine. It is built for research, lab testing, and real-world development scenarios where a smaller, lighter body is actually more useful than a full-sized one.
Pi has 27 degrees of freedom. The newly developed integrated joint modules are 15 to 27% smaller than Jaka's previous generation, which matters because compact joints reduce total weight and make the whole body easier to control. The knee joints deliver up to 120 Nm of torque, enough for stable locomotion. Each arm handles up to 3 kg, so Pi can pick up and manipulate objects in research scenarios rather than just walking around as a demonstration unit.
The most technically interesting part is what Jaka calls the fusion brain architecture, built on Intel's heterogeneous computing platform. The system separates into two distinct layers. The cerebrum handles AI reasoning, vision perception, large language model processing, application logic, and high-level decision-making. The cerebellum controls real-time movement through an EtherCAT-based network with millisecond-level latency.
This separation exists because humanoids need two fundamentally different types of computation running simultaneously. Understanding a spoken instruction is a different problem from moving a leg joint with deterministic timing. Combining them in a single stack creates conflicts. Separating them means the robot can reason about what to do without that process interfering with how reliably its body moves.
| Spec | Jaka Pi |
|---|---|
| Height | 4 ft / 1.22 m |
| Weight | 92 lb / 42 kg |
| Degrees of Freedom | 27 |
| Knee Torque | 120 Nm |
| Arm Payload | 3 kg per arm |
| Computing Platform | Intel heterogeneous (cerebrum + cerebellum split) |
| Motion Control | EtherCAT, millisecond-level latency |
Jaka's existing product lineup - Zoo cobots, Pro series for harsh industrial environments, S series for force-sensitive tasks, and its K1 humanoid line - shows this is not a company making its first move into robotics. Pi is the next step in a longer arc from industrial arms toward embodied AI systems that work alongside people.
Vietnam Enters the Race: Dino and VRH3 From the VinGroup Ecosystem
Vietnam's VinGroup ecosystem produced two separate humanoid robots this week, from two different subsidiaries, targeting two very different markets. They both debuted at ICRA 2026 in Vienna and Computex Taipei 2026.
VinBig Dynamics - Dino is the company's first humanoid, described as a versatile assistant for modern living environments. VinBig has not released height, weight, or torque specifications yet. What the company has shared is the target scope: security and surveillance in urban areas, campuses, and commercial spaces, alongside household assistance. That is an unusually wide brief for a first-generation robot, and worth watching skeptically as specs emerge.
Dino's most substantive test so far came from a deployment at Vinpearl Safari Phu Quoc, where the robot operated as an autonomous multilingual guide. It handled outdoor conditions, responded to visitors through natural language, and used real-time environmental awareness to navigate crowds. Outdoor service environments are actually hard for robots - unpredictable lighting, noise, movement, and people who do not behave predictably. Getting through that test without obvious failures is a more meaningful signal than a clean lab demo.
VinBig also showed two hardware components that underpin its humanoid program. The VDM 80 actuator weighs under 1 kg, runs on a 48V supply, supports CAN FD, RS485, and EtherCAT communication standards, reaches up to 235 RPM, and is rated for more than 10,000 hours of operational life. The robotic hand has 11 moving joints, 6 actively controlled degrees of freedom, and integrated force sensors for grip accuracy. These are the building blocks - Dino's future capability will depend on how well these scale.
Vin Robotics - VRH3 is a different company under the same VinGroup umbrella, and a different machine with a different purpose. This is the third-generation humanoid from Vin Robotics, aimed at industrial automation, material handling, and operational support in complex environments.
VRH3 has more than 31 actuators for whole-body coordination, runs on two onboard edge computers for local processing without relying on remote systems, and handles payloads of 6 to 8 kg. Low-latency onboard compute is the right call for industrial settings - remote processing introduces delays that make robots unsafe in environments where timing matters.
The claim that stands out most: Vin Robotics says all key technologies were developed in-house, including the mechanical architecture, real-time computing infrastructure, electrical and electronic architecture, power distribution, battery management, and full-body AI control framework. Vertical integration at this level is unusual. It gives the company direct control over hardware-software interaction, which tends to produce better results on balance, manipulation, and energy efficiency across generations - but it also means every failure is entirely theirs to solve.
At ICRA 2026, VRH3 was demonstrated with a teleoperation system that uses motion capture built directly into a VR headset. No external tracking equipment needed. An operator can guide the robot's movements in real time using natural body motion - useful for hazardous environments, remote maintenance, and any situation where keeping a human at a safe distance while still using human judgment makes sense.
NVIDIA's Open Humanoid Robot Research Platform: 2,070 Teraflops and a Full Software Stack
NVIDIA's approach is structurally different from the other two. Rather than launching a finished robot, NVIDIA announced at GTC Taipei what it calls the first open humanoid robot reference design, built on the Isaac GR00T development platform. The goal is to give research institutions a complete, pre-integrated foundation - hardware, simulation, AI models, middleware, and deployment tools - rather than forcing every lab to assemble those components separately from different vendors.
The reference design is built around a Unitree H2 robot body - approximately 6 ft tall, around 150 lb, with 31 degrees of freedom. NVIDIA paired it with dual Sharp Wave tactile five-finger hands, each with significant articulation, bringing the total to 75 degrees of freedom across the platform. That number matters because hands are where humanoid robots currently fall apart in real tasks. Adding tactile five-finger manipulation to a full walking platform opens up a much wider range of research scenarios.
| Spec | NVIDIA / Unitree H2 Reference Platform |
|---|---|
| Height | ~6 ft / ~150 lb |
| Total DOF | 75 (31 body + 44 hands) |
| Onboard Compute | Jetson AGX Thor T5000 (Blackwell GPU) |
| AI Performance | 2,070 FP4 teraflops |
| CPU | 14-core ARM |
| Memory | 128 GB unified |
| Arm Torque | 120 Nm |
| Leg Torque | 360 Nm |
| Payload | 15 kg |
| Sensing | Head stereo camera, wrist cameras, IMU |
The software stack covers the full research workflow. Isaac Teleop handles demonstration data capture. Isaac Sim and Isaac Lab handle virtual training and evaluation. Isaac ROS deploys trained policies onto physical hardware. NVIDIA is also including open foundation models for humanoid reasoning, so labs can build on top rather than training from scratch. Institutions can take the full stack or adopt individual components into existing workflows.
NVIDIA says the platform will also support the Unitree G1, which extends its reach to institutions already using that platform. Four major research institutions have already committed: AI2, ETH Zurich, the Stanford Robotics Center, and UCSD's Advanced Robotics and Controls Laboratory. Steve Cousins from the Stanford Robotics Center noted that robotics accelerates when researchers can build on open platforms, share code, and test on real machines.
Why All Three Are Solving the Same Problem Differently
Strip away the marketing from each announcement and all three are wrestling with the same unsolved problem: how do you get a robot to understand what it is supposed to do and then reliably execute it with a physical body?
Jaka's answer is architectural separation. Keep high-level reasoning in one layer, real-time motion control in another, and let them run on different hardware with different latency requirements. This is a sensible engineering choice - the two types of computation have actually different constraints, and forcing them to share resources tends to make both worse.
Vin Robotics' answer is vertical integration. Control every layer of the stack in-house, from battery management up to the AI control framework, and iterate tightly across all of them. The risk is that in-house development on every component is slower and more expensive. The benefit is that when something needs to change - and it always does - you can change it without depending on external vendors.
NVIDIA's answer is to remove the integration problem entirely for researchers. Provide a pre-assembled, open platform with enough compute headroom that labs can focus on the AI and policy training, not on making sensors talk to actuators. 2,070 FP4 teraflops onboard is a significant number - it means real AI workloads can run on the robot itself rather than offloading to a remote server, which matters for latency and for testing in environments without reliable connectivity.
The geography is also worth noting. China, Vietnam, and the US are all moving in the same direction simultaneously. The Vietnamese entries are the most surprising - two humanoids from a single national corporate group, one targeting service environments, one targeting industry, both with specific hardware components already beyond prototype stage. That is a faster trajectory than most observers expected from that region.
My Take
The dual-brain architecture in Jaka Pi is the most underreported detail in this entire story. Every humanoid robotics team is dealing with this exact tension - AI reasoning and real-time motion control do not belong on the same compute stack, and the companies that figure out clean separation between them earliest will have a structural advantage. Jaka naming it explicitly and building around it is worth more attention than it is getting.
NVIDIA's play is obvious and probably correct. Fragmented tooling is actually the bottleneck for academic robotics right now. If Isaac GR00T becomes the default research stack the way ROS became the default middleware, NVIDIA ends up embedded in every humanoid lab in the world - and that is before the robots ever ship commercially.
Vietnam is the one nobody saw coming. Two separate humanoids, real hardware specs on VRH3, an actual field deployment for Dino, and a full in-house component stack. The VinGroup ecosystem is moving at a pace that does not match its current global profile in robotics. That gap probably closes faster than most people expect.
- Jaka Pi uses a cerebrum-cerebellum split - AI reasoning and real-time motion control on separate hardware layers, connected via EtherCAT at millisecond latency.
- VinBig Dino has already been tested as a live multilingual guide at Vinpearl Safari Phu Quoc - outdoor, crowd conditions, real deployment.
- Vin Robotics VRH3 claims full in-house development across all hardware and software layers - unusual vertical integration for a third-generation robot.
- NVIDIA's reference platform packs 2,070 FP4 teraflops into a 6-ft robot body - enough for onboard AI inference without remote server dependency.
- AI2, ETH Zurich, Stanford Robotics Center, and UCSD have already committed to the NVIDIA Isaac GR00T platform.
- None of these platforms have confirmed commercial availability dates or pricing.
FAQ
What is the difference between Jaka Pi's cerebrum and cerebellum systems?
The cerebrum handles high-level AI tasks - language model processing, vision perception, environment understanding, and decision-making about what the robot should do. The cerebellum handles real-time motion control through an EtherCAT network at millisecond-level latency, managing how the robot's joints and body actually move. The two systems run separately on Intel's heterogeneous computing platform so neither interferes with the other's timing requirements.
What institutions are using NVIDIA's Isaac GR00T humanoid platform?
AI2, ETH Zurich, the Stanford Robotics Center, and the Advanced Robotics and Controls Laboratory at the University of California San Diego have committed to using the reference design as of the GTC Taipei announcement.
How much compute does the NVIDIA humanoid robot research platform have onboard?
The platform uses a Jetson AGX Thor T5000 with a Blackwell-based GPU delivering 2,070 FP4 teraflops of AI performance. It also has a 14-core ARM CPU and 128 GB of unified memory. This is enough to run real AI inference locally on the robot without offloading to a remote server.
What is VRH3 and how is it different from VinBig Dino?
VRH3 comes from Vin Robotics, a VinGroup subsidiary, and is aimed at industrial automation, payload handling up to 8 kg, and assembly support. Dino comes from VinBig Dynamics, a separate VinGroup company, and targets service, security, guidance, and household roles. They are different machines from different teams within the same corporate group, unveiled at the same events.
Is the NVIDIA humanoid robot research platform available to buy?
No pricing or availability date was announced. The platform was introduced as a research foundation for institutions, not a commercial product. Interested labs should monitor NVIDIA's Isaac GR00T documentation for updates.
Conclusion
Three platforms, three countries, three different answers to the same engineering problem - and none of them with confirmed ship dates or pricing. The announcements are technically substantive, the underlying hardware specs are real, and the deployment at Vinpearl Safari is the kind of grounded test that separates lab robots from anything approaching production readiness.
The question worth sitting with: which architecture actually holds up when the robots move out of controlled settings? Jaka's split-brain approach, NVIDIA's raw compute headroom, or Vin Robotics' vertical integration? The answer probably looks different depending on the environment. And we will not know for certain until these platforms accumulate real operating hours outside the demo stage.
0 Comments