While the robotics industry obsesses over VLA models and algorithmic “brains,” Benmo Technology is proving that without a highly capable, direct-drive physical “body,” true embodied intelligence can’t even reach the job site.
Who says all robots are starting to look the same?
Recently, a viral video left me utterly stunned. Two wheeled-bipedal robots aligned, docked, and then… they merged.
Moments later, the newly combined machine took a turn and stood up. And if you thought that was the end of it, theoretically, this robot can merge infinitely. If you wanted to, it could keep evolving, transforming into a silicon centipede. It was hard to believe my eyes—this was raw hardware in action, with absolutely zero AI-generated trickery.
But don’t laugh. While humanoid robots are still busy dancing and performing martial arts on gala stages, these unassuming little machines are already hauling water jugs and delivering documents. (P.S. When it comes to actual physical labor, wheeled-legged robots are the real MVPs.)
These robots, quietly doing the heavy lifting out of the spotlight, come from Benmo Technology—a company that has focused exclusively on the robotic “body” (the physical chassis and actuators) since day one.
Admittedly, when I first saw these videos, I was skeptical. In the current era of AI, focusing on the physical body—especially a non-humanoid one—doesn’t exactly sound like the fast track to Physical AGI. But after binge-watching their entire portfolio of field tests, my perspective shifted. While the “brain” is the hottest topic in embodied AI right now, the critical role of the “body” deserves far more attention.
Embodied AI is More Than Just a “Brain”
The “Brain” is undeniably the sexiest track in embodied intelligence today. Data shows that in the first half of 2026, domestic funding in the embodied AI sector reached approximately 44 billion RMB, with over half flowing into the “Brain” camp. The industry is in a frenzy over Vision-Language-Action (VLA) models, world models, and the race to train on millions of hours of data, making even traditional LLMs look like legacy tech.
Behind this boom is a welcome trend: the showboating is ending, and robotics companies are pivoting back to actual utility.
Suddenly, it’s a fierce competition of demos. Today a robot is unpacking boxes, tomorrow it’s folding laundry, and the next day it’s opening soda cans. But in front of the exhibition booths, rarely does anyone ask the crucial question: Can these demos actually be deployed in the real world?
Setting aside the issue of generalization, a massive number of these prototypes can only operate in highly optimized, standardized environments like controlled factory floors. Send one to a standard office park, and it gets stuck on a doorstep; send it down a sloped ramp, and it grinds to a halt. If a robot struggles just to reach its destination safely, how can it possibly do any real work?
There is a widely accepted dual-layer framework in the industry, which guides much of today’s popular research: System 1 and System 2, where the “brain” handles cognition and the “cerebellum” handles motor control. But most people forget the crucial third layer: The “Body,” which is responsible for “arriving.”
This is the fundamental difference between embodied AI and software LLMs. Stairs, gravel roads, and 80-kilogram payloads are hard physical constraints that cannot be solved simply by training a better model. No matter how brilliant the brain or how agile the cerebellum, if the body cannot access the environment, the result is zero.
At past trade shows, a common refrain was: “The physical body isn’t that important right now. We can max out the leverage on the AI brain; a 60-out-of-100 body is good enough.”
But the reality is that a “60-point” body doesn’t even qualify to enter the testing arena. If we constantly force the environment to adapt to the robot, how can we ever talk about true generalization?
This is exactly the bottleneck Benmo Technology is trying to break. Their mission is to push the “Body + Cerebellum” to the absolute limit, paving a smooth, frictionless road for the AI brain to finally land in the physical world.
Turning the “Body” into a Business
Benmo Technology was founded by Zhang Di, who completed his undergraduate studies in Mechanical Engineering at the Beijing Institute of Technology before earning his master’s in Robotic Systems and Control at the Hong Kong University of Science and Technology, studying under Li Zexiang, widely known as the “Godfather of DJI.”
Even as a student, Zhang was a hardware tinkerer. He once built a robot that could stand and balance on its head using JavaScript controls, which he promptly sold. This early student hustle perhaps foreshadowed the company’s enduring technical philosophy: Pick what actually works in the real world.
Founded in 2020, Benmo Technology entered the robotics joint market with a highly contrarian approach: integrated direct-drive technology, completely eliminating the need for gearboxes (reducers).
At the time, this went against industry consensus. Electric motors naturally produce high speeds and low torque. To reduce speed and amplify torque, a gearbox was considered as mandatory as a transmission in a gas-powered car. But Benmo saw an opportunity in this outdated paradigm.
Adding intermediate transmission gears increases leverage, but it also introduces energy loss, mechanical noise, and a shortened lifespan. In an industrial setting, this might be tolerable. But in a consumer robot—like a vacuum cleaner that rattles loudly and breaks down after a few years—those flaws are dealbreakers.
By removing the gearbox, Benmo solved these problems at the root. Their joints are more compact, operate in near silence, and boast a lifespan of 50,000 to 100,000 hours. This is the core reason Benmo bet everything on integrated direct-drive joints.
However, simply manufacturing hardware is just the entry ticket; getting people to use it is the real hurdle. To avoid the red ocean of being a mere component supplier, Benmo launched a second business line: complete robotic systems.
1. Xingtian: The world’s first commercialized direct-drive wheeled-bipedal robot.
A true heavyweight, Xingtian boasts a crawling payload and towing capacity of over 80 kg. It can carry adult humans, jump, climb slopes, and hold its position on inclines. It is essentially a robotic strongman, purpose-built for heavy labor in warehouses and industrial parks, such as material transport and logistics.
2. TITA:
Designed for vastly improved agility and robustness, TITA features 8 Degrees of Freedom (DOF) and a 10 kg dynamic payload. It can operate continuously in extreme temperatures ranging from -10°C to 45°C. Equipped with optical sensors, speakers, and cameras, TITA handles inspections, logistics, data collection, and mapping across various terrains. With an added navigation module, it performs point-to-point document delivery in libraries and office buildings.
3. D1: The modular, shape-shifting robot from the viral video.
Modularity is its defining feature. As the world’s first fully modular embodied intelligent robot, D1 can be split apart or merged back together in just 3 seconds. It can freely switch between wheeled-bipedal, four-wheeled, and flat-footed configurations depending on the task at hand.
While it looks like a sci-fi dream come true, the “merging” feature is actually highly practical for scenario adaptability. For lightweight tasks like area mapping, D1 can split into two independent units to run parallel operations. When it encounters stairs or obstacles, the units merge to guarantee mobility.
Furthermore, D1 is a game-changer for developer training. Because the bipedal and quadrupedal forms share the same underlying algorithmic framework, developers only need to toggle an SDK to switch locomotion requirements, drastically cutting down debugging time. During field testing, R&D teams no longer need to buy one robot for flat ground and another for stairs. One machine does both, effectively cutting hardware testing costs in half—freeing up budget to train the “Brain.”
Through these products, Benmo’s two business lines intersect perfectly: the direct-drive power modules provide the hardware, algorithms, and manufacturing foundation for the complete robots; meanwhile, the complete robots are thrown into real-world scenarios to collect the rarest, most valuable operational data, which in turn feeds back into the iteration of the joints and modules.
The market has already cast its vote:
- Direct-drive modules: In 2025, Benmo shipped approximately 8.5 million units, capturing a massive 61.1% market share in China’s consumer robotics direct-drive module sector.
- Wheeled-legged robots: Exported to over 50 countries and regions, making Benmo the highest-volume global shipper of wheeled-legged robots.
Embodied AI is a War of Attrition
Let’s return to the original question: Why, in the era of the “Brain,” do advancements in the “Body” still matter so much?
Many embodied AI companies are aggressively harvesting robotics data, but as of now, the industry is only scraping the surface with a few million hours of footage. Moreover, in crucial physical dimensions like tactile feedback, existing datasets are virtually blank.
How much high-quality data does Physical AGI actually require? The industry consensus is that achieving general autonomous capability requires a baseline of tens of millions of hours of data. This means the complete maturation and deployment of large-model “brains” will take years, if not a decade, of waiting.
But robots cannot afford to sit idle, and the industry cannot afford to pause.
Some companies specialize in decision-making, some in execution, and some in physical access. Only by advancing the Brain, the Cerebellum, and the Body simultaneously on three parallel fronts can we truly accelerate the countdown to full real-world deployment.
Against this backdrop, Benmo Technology’s laser focus on the “Body” provides a clear answer for the future of the hardware industry: No matter whether VLA, World Action Models (WAM), or native embodied models ultimately win the AI race, the final destination for those models must be deployed onto a reliable, highly capable physical body.
And someone has to build that body today.
Xingtian hauling 80 kilograms of supplies through a warehouse; TITA carrying documents through a corporate lobby; D1 splitting into two units to map a facility in parallel… These are not staged exhibition demos; these are daily operations already happening in industrial parks worldwide.
The endgame of embodied AI will undoubtedly rely on the Brain. But the ticket to enter the future? That’s the Body.