Motional’s Blueprint for High-Performance ML Systems in Autonomous Driving | Ray Summit 2025 скачать в хорошем качестве

Motional’s Blueprint for High-Performance ML Systems in Autonomous Driving | Ray Summit 2025

At Ray Summit 2025, Dhananjai Sharma and Muhammad Taufik Tirtosudiro from Motional share how the company transformed its ML development lifecycle by rebuilding its data processing system to handle terabyte-scale autonomous vehicle data with speed, reliability, and efficiency. They describe the challenges posed by Motional’s legacy pipelines—manual processes, rigid architectures based on static data replication, and expensive general-purpose distributed frameworks that slowed feature engineering and model iteration by weeks. To overcome these limitations, Motional engineered a unified, horizontally scalable ML system built on Ray, enabling ML engineers to process terabytes of data in hours instead of weeks. In this talk, they present Motional’s blueprint for designing reliable and high-performance data processing pipelines for large-scale feature generation across Perception, Prediction, and Planning. They detail the company’s migration away from a brittle staging-based implementation toward a fully autoscaled Ray architecture that drastically reduced operational friction and costs while empowering ML engineers to own their workflows end-to-end. A highlight of the session is Motional’s novel “1-actor-per-node” Ray Actor pattern, designed to bring compute to the data and eliminate costly network communication. This approach not only boosts performance but also serves as a compelling candidate for inclusion in the official Ray Patterns documentation. Attendees will leave with practical, proven strategies for building scalable, performant, and resilient ML systems capable of tackling massive data challenges across industries. Liked this video? Check out other Ray Summit breakout session recordings at Subscribe to our YouTube channel to stay up-to-date on the future of AI!    / anyscale