While the world chases robot 'brains,' a Shenzhen startup sold 8.5 million robot joints

In 2025, a small Shenzhen company called Direct Drive (本末科技) shipped roughly 8.5 million direct-drive power modules, the brushless motors that move a robot's hips, knees, and ankles. That is the count qbitai reports from the company itself, which now claims 61.1% of China's consumer-robot direct-drive power-module market and exports wheel-legged robots to more than 50 countries.

While the rest of embodied AI spent 2024–2026 chasing bigger vision-language-action (VLA) models, world models, and training data, Direct Drive invested in a piece of hardware that almost every robot maker had already standardized: the reducer, a precision gearbox that sits between an electric motor and a robot joint to multiply torque and smooth motion.

Direct Drive left it out.

"Direct drive" means the motor turns the joint directly, with no gearbox in between. That trades some torque and absolute precision for quieter operation, a smaller form factor, and a stated service life of 50,000 to 100,000 hours. For a humanoid or quadruped walking on gravel, those are the specs that matter: backlash-free joints respond faster to a foot landing wrong, and the absence of gears removes the lubrication and wear that fail first in a robot deployed outside a lab.

The bet looked out of step in 2020, when founder Zhang Di quit a research track at the Hong Kong University of Science and Technology under Zexiang. The lab orthodoxy at the time, and the orthodoxy that still dominates 2026, is that the gating layer in embodied AI is software: better policies, more demonstrations, larger models. In H1 2026, Chinese investors put roughly 44 billion yuan (around $6 billion at prevailing 2026 exchange rates) into embodied-intelligence startups, and more than half of that capital went to "brain camp" efforts: VLA models, world models, and the data pipelines that train them, according to qbitai's tally.

Direct Drive's argument is not that the brain does not matter. Zhang and his team treat a robot as a layered system (body, cerebellum, brain) and the company's pitch is that the body layer is the one currently failing field tests. A VLA model that cannot command a joint across a doorstep, lift a real load, or survive a 20% grade is not a working robot.

The flagship 刑天 / Diablo is a heavy-load wheel-legged platform rated for 80 kg-plus payloads. TITA, an 8-DOF mid-size platform, runs from -10°C to 45°C and carries 10 kg dynamically, with integrated optics, speakers, and cameras for inspection and logistics. D1, the modular unit that reconfigures in roughly three seconds on Direct Drive's Bilibili channel, morphs between dual-wheel, quad-wheel, electric-foot, and flat-foot forms, all driven by the same SDK. The hardware is the point: each variant uses the same direct-drive joints, so a customer writing software for one form factor can port the policy to the others.

The 61.1% market share, the 8.5 million units, and the "world's highest wheel-legged shipment volume" claim all originate with the company and pass through qbitai's profile piece; none have been independently audited by a body such as GG Robot, IDC, or the IFR. The "body-first" framing also has a counter-position the brain camp will press: as VLA and world models scale, the argument goes, software will eventually subsume the kind of joint-level engineering Direct Drive specializes in, the same way high-level compilers absorbed most hand-tuned assembly.

What the company can point to is five years of installed units, the currency that argument cannot yet match. A 70 kg robot walking a real staircase is harder to fake than a benchmark curve, and that gap is the open question for 2026: whether the next generation of embodied-AI capability compounds on top of working hardware, or whether the brain camp has to wait for hardware bets like this one to mature first.