Shenzhen-based DexForce is using massive virtual simulations to train humanoid robots. This approach aims to bypass the high costs and logistical hurdles of collecting real-world data.
The 10 trillion sample barrier to physical AGI
The robotics industry has long been caught between two worlds: the ability to perform impressive physical feats and the ability to handle mundane, practical tasks like picking up a screw or serving coffee. While many companies focus on the former, the latter is what drives market viability. However, teaching a robot to master these nuances requires an immense amount of data.
Jia Kui, a machine-learning professor at the Chinese University of Hong Kong, Shenzhen and CEO of DexForce, notes that achieving "physical artificial general intelligence" requires roughly ten trillion real-world samples. As Reuters reported,collecting this much data by hand is a logistical impossibility. To solve this, DexForce uses digital universes to simulate thousands of scenarios—altering lighting, angles, and physics—in a fraction of the time it would take a human worker to perform repetitive tasks.
A 1 billion yuan infusion from Lenovo Capital
Since its inception in 2021, DexForce has rapidly scaled its operations and its capital reserves. The 500-person company recently secured a Series B funding round totaling 1 billion yuan, which is approximately 147 million US dollars. This capital has significantly boosted the firm's valuation to over 10 billion yuan.
The financial backing for DexForce comes from heavy hitters, including Lenovo Capital and Lens Technology, alongside various state-owned funds. This investment allows the company to move beyond mere research and into widespread industrial application, providing simulation software to a growing list of global manufacturers.
From Samsung to Foxconn: The industrial adoption of DexForce
The utility of DexForce's synthetic data approach is already being tested by some of the world's largest manufacturing and consumer electronics giants. According to the report, companies including Toyota, Samsung, Panasonic, Foxconn, Haier, and Hisense are utilizing the DexForce platform to train their robotic systems.
Beyond the factory floor, the company's hardware is making a public appearance. DexForce's customer-service and coffee-making robots are currently operating in more than 15 cities across mainland China, serving customers in malls and tourist locations.. This deployment demonstrates a shift from controlled laboratory environments to the unpredictable reality of public spaces.
Navigating the 'sim-to-real gap' and public safety
Despite the efficiency of virtual training, significant technical and social hurdles remain. A primary concern for engineers is the "sim-to-real gap"—the difficulty of ensuring that knowledge learned in a perfect digital environment translates accurately to the messy, physical world. If the calibration is off, a robot trained in a simulation may fail when faced with real-world friction or gravity.
Furthermore, the deployment of humanoid robots in public areas raises unanswered questions regarding safety and regulation. While Jia Kui remains optimistic about an iterative approach to closing these gaps, the industry has yet to see a standardized regulatory framework for robots roaming freely in shopping centers or streets. It remains to be seen how DexForce will address these safety concerns as they scale their presence in China's urban centers .
Comments 0