The Genesis: Project BR1609 and Drone Vision
In 2016, within the innovation labs of DJI, a project codenamed BR1609 began. Few at the time could have predicted this internal research endeavor would evolve into Zhuoyu, one of China's most distinct autonomous driving companies. The project's roots are deeply intertwined with DJI's pioneering work in drone-based machine vision. At the forefront of this was the Phantom 4, which had already dominated the consumer drone market with its advanced machine vision capabilities, alongside the portable Mavic series that redefined the category. Shen Shaojie (沈劭劼), then an assistant professor at HKUST and a key DJI engineer focused on drone localization and planning, led the team that first integrated stereo cameras onto drones. They successfully adapted their HKUST research, a visual-inertial navigation system built on the VINS framework, from the laboratory into production drones.
This work on drones provided a critical foundation. The team's experience in developing robust visual-inertial navigation systems for drones, which require precise localization and path planning in complex, dynamic environments, proved directly transferable to the challenges of autonomous driving. The ability to process visual data in real-time, fuse it with inertial measurements, and make navigation decisions was a core competency honed on aerial platforms. This cross-pollination of expertise from drone technology to automotive applications is a testament to the underlying principles of perception and navigation systems.
Building the Autonomous Driving Stack
Zhuoyu's journey into autonomous driving was not an overnight transition. It began as a pre-research project within DJI, leveraging the company's strengths in perception hardware and software. The team's intimate knowledge of camera systems, sensor fusion, and real-time processing, developed for drone navigation, formed the initial building blocks for their autonomous driving efforts. This included the development of sophisticated algorithms for object detection, tracking, and scene understanding, all crucial for a vehicle to navigate safely and efficiently.
The early stages involved adapting and extending these drone-centric technologies to the automotive domain. This meant dealing with a different set of environmental conditions, sensor requirements, and operational complexities. While drones operate in three-dimensional space with relatively predictable flight paths, autonomous vehicles operate on two-dimensional roads, interacting with a much wider array of unpredictable agents like pedestrians, cyclists, and other vehicles. The transition required significant R&D to address these new challenges, including the development of advanced prediction models and decision-making logic tailored for road environments.
The Bold Decision: Deleting the Codebase
Perhaps the most striking aspect of Zhuoyu's evolution is its decision to completely delete its existing codebase. This was not a trivial undertaking; it represented a fundamental re-evaluation of their technological direction and a commitment to a new paradigm. This move signals a departure from incremental improvements and an embrace of radical rethinking. Companies in the competitive autonomous driving space often build upon existing frameworks, iterating and optimizing. Zhuoyu's choice to wipe the slate clean suggests a belief that a fundamentally different approach was necessary to achieve their goals.
The rationale behind such a drastic measure is likely rooted in the desire to escape technical debt and legacy constraints that often accumulate in long-term projects. Developing complex software systems, especially in rapidly evolving fields like AI and autonomous driving, can lead to architectural compromises and inefficiencies over time. By starting anew, Zhuoyu aimed to build a more elegant, efficient, and scalable system from the ground up, free from the limitations of its past development. This is akin to a chef discarding a slightly burnt dish to start over with fresh ingredients, ensuring a superior final product.
Implications for the Autonomous Driving Landscape
Zhuoyu's unique trajectory offers a compelling case study in innovation within the autonomous driving sector. The company's origin within DJI, a leader in visual sensing technology, provided a strong technological foundation. However, it was their willingness to make bold, unconventional decisions, such as deleting their entire codebase, that truly sets them apart. This suggests a strategic imperative to achieve a technological leap rather than merely an incremental advance.
This approach also raises questions about the long-term viability and scalability of such radical development cycles. While starting fresh can yield significant benefits, it also carries risks. The time and resources required to rebuild a complex autonomous driving system from scratch are substantial. Furthermore, the competitive pressure in the autonomous driving market means that companies must deliver results quickly. Zhuoyu's success will hinge on whether this bold gamble pays off, enabling them to develop a superior system that can outcompete rivals who have opted for more conventional development paths.
The company's story highlights a broader trend: the increasing importance of foundational technology and the willingness of innovative firms to challenge established development methodologies. As autonomous driving technology matures, we can expect to see more companies exploring diverse approaches to software architecture and development, driven by the pursuit of safety, efficiency, and performance. Zhuoyu's evolution from a drone vision project to a code-deleting autonomous driving startup is a potent example of this dynamic landscape.