Robotics Teams Are Rebuilding the Data Stack from Scratch
Summary
The article discusses the data layer tax in robot learning, arguing that robotics teams must build bespoke data tooling due to multi-rate, multimodal data from sensors, cameras, and actions. It outlines policy evaluation, model training, sample construction, video decoding, dataset curation, and data ingestion challenges, and positions a unified data layer as essential to accelerate end-to-end robotics research.