About This Talk
Training robotics foundation models on a million hours of multimodal data is not simply a matter of adding more GPUs. The bottleneck continually moves across the system as datasets grow, pipelines become more dynamic, and training spans larger clusters.
This session explores the infrastructure behind Dyna-2, from robotics-native episode storage and large-scale ingestion to dataset curation, cluster-local caching, distributed training, and job resilience. It examines how targeted redesigns increased ingestion throughput by 31x, reduced training startup from roughly 48 hours to under a minute, and sustained high GPU utilization across multi-node clusters.
The session also shares recent progress using Lance as a unified data substrate for search, iterative data enrichment, and training, reducing the need to move or rematerialize data across separate systems. It looks ahead at the next set of infrastructure challenges as robotics datasets push toward greater scale and diversity, including tighter feedback loops between data collection and training, faster iteration on continuously evolving datasets, and infrastructure that can support increasingly multimodal, long-horizon foundation models.












