Stardust OS Hackathon Beijing: 15 Teams Build 10+ Robot Apps in 36 Hours
Stardust Intelligence's Stardust OS Hackathon concluded its second stop in Beijing, billed as the industry's first hackathon centered on generating robot apps from one-sentence prompts. Fifteen teams produced more than 10 runnable robot applications in 36 hours.
Participants included many cross-disciplinary creators with no robotics R&D experience, as well as embodied AI researchers who had little prior experience with real humanoid hardware. The event emphasized low barriers, allowing teams to build without writing code from scratch.
The Beijing event was defined as one of the industry's first low-barrier development competitions focused on opening robot operating system capabilities. Participants did not need to write motion control, communication, or backend code from scratch; instead, they used Stardust AOS Skill packages, Meta packages, and natural language generation to turn one-sentence requests into schedulable, verifiable robot applications.
Traditional embodied AI development usually relies on a full engineering team and covers body drivers, camera calibration, SLAM, locomotion control, backend, and interaction systems, with cycles measured in weeks or months. Data from the Beijing stop showed that several teams had never operated real humanoid robots before, yet completed scenario definition, simulation validation, and parts of the real-machine pipeline within two days.
The CareLoop team, made up of two researchers, ran a full care scenario chain for the first time: delivering medicine, delivering water, collecting garbage bags, and assisting caregivers in turning bedridden elderly patients. The team said AOS skill packages and Agent assembly mechanisms shifted the focus from writing systems to defining scenarios and closing loops.
Three automation undergraduates from Beijing Institute of Technology built a mahjong robot named momo that can recognize tiles, draw tiles, and discard tiles, with emotional interactions such as celebrating victory and crying on a loss. They originally planned to use YOLO plus custom recognition logic, but ultimately called existing AOS SDK and motion control interfaces, significantly reducing development time.
The DJ Master team, composed of members with physics backgrounds and embodied project experience, realized an entertainment robot pipeline covering song upload, beat analysis, DJ actions, and rhythm-linked movement.
Unlike most solutions that layer large models onto systems as chat plug-ins, Stardust AOS places AI understanding, capability scheduling, body control, and application distribution within the same operating system. Developers face not a code repository but a set of composable capabilities: dexterous manipulation, mobile navigation, multimodal interaction, safety protection, and more.
The 15 applications included practical scenarios such as caregiving, restaurant AI management, chemical material sorting, home storage, and IoT integration, as well as entertainment scenarios like Pictionary, pitch-pot, and toast emoticon packs. A common feature was that no one wrote low-level systems from scratch.
Stardust Intelligence explicitly proposed a robot APP Store direction at the event: developers produce not one-off demos but packaged, distributable, reusable skill packages. In the future, robots may acquire new capabilities by downloading and licensing an app rather than commissioning a custom engineering team.
The finals are scheduled for October 24-25, 2026, in Shenzhen. Judges will focus on scenario value, loop completeness, real-machine feasibility, and app distributability, rather than purely algorithmic difficulty.