Fu Zhang, a HKU assistant professor and former eight year DJI senior consulting scientist, says the drone industry's 'second half' is onboard autonomy, meaning flying robots that perceive, decide and physically intervene in GPS denied environments.
A quadcopter can hover next to a damaged bridge and stream video of a loose bolt to an engineer miles away. It cannot tighten the bolt. That gap between aerial observation and aerial intervention is the bet behind SPARO, a six-month-old Chinese startup whose founder argues the drone industry's "second half" belongs to autonomy, not to better airframes.
The founder is Fu Zhang, an assistant professor at the University of Hong Kong and head of its Multi-Agent Robotics and Systems Lab. Zhang spent roughly eight years as a senior consulting scientist at DJI, contributing to flight control, multi-sensor fusion, and LiDAR work, and his open-source SLAM and perception code is already used by Unitree, Meituan, and "dozens" of other Chinese robotics firms, according to 36Kr. His HKU faculty profile lists a bachelor's from USTC in 2011, a PhD from UC Berkeley in 2015, and an HKUST postdoc before he joined HKU as an assistant professor in August 2018. On Google Scholar his publication record sits at roughly 10,000 citations, with papers such as FAST-LIVO2, the IEEE Transactions on Robotics Best Paper winner in 2025, and SUPER, published in Science Robotics the same year. Per ScholarGPS, he ranks fifth worldwide and first among Chinese scholars in robotics over the last five years, and falls within Clarivate's top 1 percent of highly cited researchers.
That pedigree is the basis for the company-side argument that aerial autonomy is a category of its own. SPARO calls the field "general aerial intelligence," a framing that puts perception, decision, and physical action inside a single software stack that runs on the drone rather than in the cloud. The company says the stack hits centimeter-level positioning without GPS, stable perception in low light, smoke, or other degraded scenes, roughly ten times the compute efficiency of traditional pipelines, and obstacle-avoidance latency under five milliseconds, versus the roughly 100 milliseconds a human pilot typically needs to react. None of those performance numbers have been independently benchmarked. They are founder-and-company claims, surfaced in a 36Kr exclusive.
The capital signal is unusual. SPARO closed four consecutive rounds inside its first six months, totaling "several hundred million RMB," a Chinese phrasing that, in founder interviews and trade press, covers tens of millions to the low hundreds of millions. The seed came from Yaotu Capital, with Jinqiu Fund, Alibaba, Hongyi Investment, GLP Yinshan Capital, and Yunshi Capital joining on follow-ons, according to 36Kr and Quantum Bit's parallel write-up. Alibaba's exact investing entity is not specified in the public reporting. Fast, repeat-tap rounds are common when late investors want to avoid being left out of a hot deal; they are also common when founders are intentionally staging capital for a longer runway. Both readings are plausible here.
Zhang's structural argument, attributed, is that the consumer and industrial drone industry spent a decade optimizing hardware specifications and unit price. The next decade, he argues, will be decided by which company owns the autonomy layer on top of those commodity airframes. DJI's continuing dominance in flight control and airframe manufacturing does not, in this framing, solve the harder problem of letting a small quadcopter perceive a partially occluded environment, decide what to do, and do it without a pilot in the loop. The bet is that onboard compute, perception, and decision can be packed into the same weight and power envelope that today's cameras and gimbals already occupy.
The honest caveats are real. SPARO has announced commercialization and is "scaling delivery" but has not disclosed independent customer counts, revenue, or field-deployment metrics, and the company is six months old. The DJI-versus-autonomy framing is Zhang's positioning, not industry consensus. The performance claims, which include centimeter-level GPS-denied positioning, sub-five-millisecond obstacle avoidance, and ten-times compute efficiency, are company-stated and self-reported. And the falsifier for the thesis is concrete: if onboard compute does not reach the efficiency needed to run full autonomy in a sub-kilogram payload budget, the intervention story stalls, and ground-based embodied-AI companies with cheaper compute move into aerial-adjacent tasks first.
What to watch next is whether SPARO's full-stack autonomy stack ships into a publicly named industrial customer with measurable reliability data, and whether the open-source perception and SLAM tools that Zhang's lab has already released become the substrate on which that stack runs.