Every autonomous vehicle crash holds lessons that could make all such vehicles safer, yet the data rarely leaves the company involved. Cornell researchers set out to understand why. In a study published in the Proceedings of the ACM on Human-Computer Interaction and presented at the CSCW conference, a team including Cornell Tech doctoral candidate Hauke Sandhaus, Qian Yang of the Cornell Ann S. Bowers College of Computing and Information Science, and Wendy Ju of Cornell Tech interviewed twelve employees working on safety at autonomous vehicle companies.
The interviews revealed an industry of silos: firms rely on widely varying in-house datasets with little shared knowledge between them. Two barriers dominate. Crash data exposes details of a company's machine learning models and infrastructure, making release politically fraught even internally, and safety knowledge itself is treated as private competitive advantage rather than public good. Regulations in the United States and Europe require only basics, such as the crash month, the manufacturer and whether anyone was injured, missing contributing factors like pedestrian movements, weather or road debris.
The authors propose separating safety lessons from proprietary detail: sharing accident information without raw video, creating standardised test scenarios every vehicle must pass before using public roads, routeing collaboration through universities and independent institutions, and having a federal agency build a virtual testing city to which companies contribute their hardest scenarios. The research was funded by the National Science Foundation and Schmidt Sciences.