blog
Blog
Thoughts on robotics data infrastructure, compression, and the future of physical AI.
CompressionDataloadingGPU UtilizationRobotics Data InfrastructureTraining Pipeline
The Quiet Tension Between Compression and Dataloading That Decides Your Training Run
Compression saves storage but costs decode time. Dataloading needs speed but starves on bloated data. The interplay between these two forces shapes every robotics training pipeline.
March 22, 2026Read →
Robotics Data InfrastructureData PipelineMLOps for Robotics
Why Frontier Robotics Teams Keep Losing to Teams With Better Plumbing
The best robotics teams don't win because of better models. They win because of better data infrastructure. Here's why plumbing beats papers in the race to ship manipulation policies.
March 11, 2026Read →
Data QualityTeleoperationImitation LearningRobotics Data InfrastructureBehavior Cloning
Your Robot is Only as Good as Your Worst Demonstration
In robot learning, data quality is not a nice problem to have. It is the problem. Bad demonstrations do not dilute your dataset. They poison it.
March 4, 2026Read →