DeepRoute.ai Pivots to Physical AI Infrastructure

DeepRoute.ai has shifted its strategic focus from serving as an advanced driver-assistance systems supplier to building foundational artificial intelligence infrastructure for the physical world. The company unveiled this new direction at the Beijing Auto Show, where it showcased its unified foundation model and emphasis on large-scale real-world data.

Credit: DEEPROUTE.AI

Chief Executive Officer Maxwell Zhou used the event to argue that the future of autonomous driving lies in creating AI capable of operating across various embodied agents, rather than focusing solely on vehicle hardware. By not displaying any vehicles at its exhibition, the organisation highlighted its goal of becoming the brain builder for physical systems.

The company currently powers urban navigation systems in over 300,000 vehicles across China. These deployed units serve as a data flywheel, having generated more than 1.3 billion km of driving data and 44.8 million hours of active usage over the past year.

DeepRoute.ai aims to reach 1,000 km in miles per critical intervention by the end of the year. Zhou noted that achieving this level of reliability requires a foundation model approach, as smaller, specialized models face limitations that prevent exponential improvement.

A key development in this transition is the addition of former DeepSeek research head Ruan Chong as chief scientist. Ruan is overseeing a move away from cognitive fragmentation, which occurs when systems rely on multiple small, specialized models for tasks such as trajectory planning, traffic light detection, and pedestrian identification.

The new foundation model unifies three core capabilities. A driver model manages vehicle actions, an analyst model handles language and data annotation, and a critic model learns from negative data to avoid dangerous driving patterns. This architecture has reduced the research and development iteration cycle from approximately five days to 12 hours.

The firm anticipates that as the number of vehicles on the road increases, the marginal cost of acquiring additional data will decrease. The company plans to reach one million vehicles on the road to secure its position as a provider of physical AI infrastructure.

---

• The company has pivoted its strategy to build foundational artificial intelligence for the physical world instead of solely providing automotive software.

• Its foundation model integrates driver, analyst, and critic capabilities to shorten research iteration cycles to 12 hours.

• A fleet of 300,000 vehicles serves as a data flywheel, producing over 1.3 billion km of real-world driving data.

Source: PANDAILY