Nvidia’s Robotic Strengths Spell Blissful Uncertainty for Chinese Firms

If the year 2023 is remembered as the year of generative AI, it’s no exaggeration to say 2024 will go down in history as the year for humanoid robotics to finally scale up.

We are witnessing a boom of general-purpose robotics, with humanoid robot becoming the face of this fast-paced industry. Since last year, heavyweights like Tesla or startups such as Figure AI have celebrated jaw-dropping milestones in the iteration of humanoid robot models, closing in on the mighty industry leader Boston Dynamics.

A handful of Chinese startups are also rising fast on the robotic scene, mounting a challenge to Western, primarily US, competitors by introducing their own dexterous, versatile humanoid models.

In the coming two to three years, as more of these robot developers come of age technically and commercially, we are poised to see a dramatic uptick in the mass deployment and adoption of humanoid robots across an array of scenarios.

With a global race around humanoid robotics heating up, Chinese companies are hankering for ways to emerge on top. A widely recognized benefit at their disposal is the nation’s vast supply network. It feeds into their production, keeping major parts like electric motor and roller screw linear actuator relatively affordable.

In the meantime, many Chinese firms find themselves at the deep end of a talent pool, with a steady supply of capable engineers to count on.

Strengths on the hardware side do not necessarily translate into advantages on the software front, though. Chinese roboticists have made enormous progress in areas such as teleoperation and complex motion control for humanoid robots. Videos abound online of models “learning” to fold clothes like humans or replicate an Atlas-style somersault.

These advances can be partly attributed to breakthroughs in AI algorithms. But this is also an area where Chinese robotic companies face what I call “blissful uncertainty” — blissful because a giant conveniently supplies them with the underlying technologies, uncertain because the giant could cut them off when they least expect it.

The name of the giant is Nvidia, whose innovations such as parallel-computing GPU (graphic processing unit), Isaac simulation platforms and other tools have largely underpinned the development work at many Chinese robotic startups.

Several domestic startups I spoke to told me that they utilize Nvidia’s iconic Isaac Gym or Isaac Sim platforms to conduct reinforcement learning (RL) training for humanoid or quadrupedal models.

With the Isaac system, the time required for validating RL algorithms for a particular robot prototype has been significantly reduced, shortening the transfer from a simulated environment to the real world. This allows for a substantial boost to its motion capabilities within a short period of time.

Nvidia’s Isaac system has even inspired the birth of domestic variations. In March, Robot Era, a Beijing-based startup, open-sourced what it calls Humanoid-Gym, a RL framework for training a massive number of robotic avatars in simulated environments.

According to Robot Era, this application, based on the Isaac software, is designed to streamline development processes, making it easier for humanoid robots to transition from a simulation avatar to a working prototype in the physical world.

The open-sourcing was intended to lower the threshold to developing RL algorithms for humanoid roboticists worldwide, providing them with a practical platform to build on so as to avoid “reinventing the wheel,” Robot Era says in a press release.

As a behind-the-scenes enabler, Nvidia has stepped up its investment over the years in humanoid robotics, pouring huge sums into innovations that have now become indispensable in RL and embodied AI.

At the GTC 2024 on March 18, Nvidia launched, among others, a new foundation model called GR00T for humanoid robotics. Simply put, this large multimodal model can parse human motions from dancing videos and drumming demonstrations, enabling a physical agent like a robot to understand natural language and emulate complex movements.

While marveling at Nvidia’s sustained ability to roll out powerful technologies, one cannot but express a reasonable dose of concern about over-dependence on this powerhouse. Could sudden curbs in access to Nvidia technologies cripple China’s nascent humanoid robotic space?

Such risks have been evident for a while now. The Verge, an online tech media outlet, reported in December 2023 that Nvidia released RTX 4090D in China, a slower variant of its RTX 4090 graphics card, to comply with US trade restrictions.

Considering that RTX 4090 is widely adopted in China by roboticists, this inevitably raises the question of what their responses would be if stricter sanctions were to come. Prior to the fresh US chip ban slapped on China in November last year, Nvidia commanded a 90% share of China’s AI chip market, according to Reuters. Further sanctions are set to dent but unlikely to shake up its domination.

To sum up, behind the facade of a burgeoning robotic sector lies ample reasons for circumspection. An increasingly powerful Nvidia is as much as an agent for change to Chinese robotic firms as a source of uncertainty.

At a time when full technical autonomy is still a long way off, continued demand for Nvidia products only underscores the lack of comparable alternatives from homegrown vendors. This doesn’t bode well for China’s budding robotic sector.

Ni Tao is the editor-in-chief of cnrobopedia.com, a website that focuses on Chinese robotics and automation.

SEE ALSO: NVIDIA’s Jensen Huang Responds to AI Chip Pricing Issue