Yang Guangyao spent years in DJI's autonomous driving unit (since spun out as Zhuoyu) before founding Xiaoqiuqiu, an AI coach that watches video and tells amateur players how to fix their swings.
A stack of uncut match videos on a phone, no way to track which swing pattern fixed which error, and no hitting partner in sight. That is the sensory problem Yang Guangyao says pushed him out of autonomous-driving engineering into amateur racket-sport coaching, and the founding pitch for a year-old app called Xiaoqiuqiu.
He frames it as the same problem he spent years solving on roads: multimodal input, large-model architecture, and a data flywheel that compounds with each user. That is the bait. The bigger story is structural. Yang pegs the coach-supply gap at roughly one registered tennis coach per 1,754 amateur players in China — a figure derived from roughly 11,000 registered coaches and approximately 20 million participants — well above the global norm of about 584-to-1 cited in the ITF Global Tennis Report. By the General Administration of Sport's own April 2025 count, the ratio has improved to roughly 572-to-1 (43,981 coaches / 25.19 million participants), reflecting both a rising coach count and a larger participation base. The structural gap Yang cites — and the market opportunity it represents — predates the official update but is directionally consistent with it. Whether the autonomous-driving playbook survives contact with a tennis court is the open question.
The pedigree is the ByteDance-then-DJI-then-NIO arc familiar from Chinese tech, with a specific twist. Yang's last AV role was at DJI's autonomous-driving group, since spun out as Zhuoyu, where he led visual perception, system decision-making, and software-hardware co-design, per a Leiphone exclusive. By his own account there, the team's underlying thesis was straightforward: streets and tennis courts both generate multimodal streams a model has to turn into actionable feedback, and both reward whoever compounds the data loop fastest. The company frames the autonomous-driving pipeline, with its perception logs, simulation tools, and replay infrastructure, as the missing substrate for amateur sports training.
The product is built around that loop. The app, launched globally about a year ago, pairs a phone-side lightweight model with a cloud-side large model that runs 3D body reconstruction, temporal action modeling, and scene understanding, then auto-generates highlight clips plus action feedback and coaching advice, per the founder's interview. An AI tennis racket is in mass production. A smart racket, a "ball buddy" sensor, and a court camera sit at prototype stage. Yang's stated approach is software-first: solve how training data gets generated before trying to compete on imaging with an incumbent like DJI. That hierarchy explains why the app is treated as the core data entry point, with hardware feeding it rather than the reverse.
The structural wedge looks real. China's tennis population reached 25.19 million by August 2024, up 28% on the ITF 2021 baseline of 19.67 million participants, serviced by 43,981 registered coaches per the General Administration of Sport. Xinhua and the Global Times frame the surge as a domestic participation boom. Xiaoqiuqiu is betting the gap between participation and coaching supply widens faster than the country can train new coaches.
Founder-disclosed early traction is modest in volume. Cumulative users sit in the tens of thousands after about a year, with roughly a third of that growth arriving in the past month, per the founder's interview. The strongest app-store chart positions are in small or niche markets: top-50 in China's sports-paid charts, Macau at number 5, Tajikistan at number 1, plus Singapore around number 60 and Slovakia at number 40. Those numbers are startup-reported and not independently verified against third-party app-intelligence snapshots.
Two caveats are worth flagging. First, the founder's pitch that AI coaching can take share from racquet-sports equipment incumbents like Wilson, Head, and Yonex is a strategic aspiration rather than a measured outcome. Those companies sell gear rather than training software, and the source does not benchmark Xiaoqiuqiu against the existing field of racquet-sport video-analytics tools. Second, the prototype-stage hardware is forward-looking and may or may not arrive on the announced roadmap. A funding round Yang describes as deferred is also unresolved.
What to watch next: whether the multimodal-perception thesis produces measurable improvement in user testing, since the source carries no outcome data; how many AI rackets actually ship, given that mass production has been announced but volumes have not been disclosed; and the funding-round resolution. The substrate thesis is worth tracking, the structural gap is real, and whether the data flywheel compounds inside it is the open question.