Why Physical AI Needed a Completely New Data Stack

What It Covers

Nico West from Rerun.ai discusses building logging infrastructure for robotics and embodied AI, covering data visualization challenges, robotics breakthrough progress, and designing systems for multimodal physical world data.

Key Questions Answered

• •Data Model Design: Physical AI requires custom data formats supporting multimodal, multirate, episodic data that traditional tabular databases cannot handle, necessitating complete infrastructure redesigns from scratch using Arrow-based systems.
• •Robotics Progress Indicators: Advanced manipulation tasks like laundry folding transformed from impossible to routine within one year through combining imitation learning with reinforcement learning and end-to-end neural approaches.
• •Open Source Strategy: Making visualization tools open source while monetizing cloud infrastructure creates adoption advantages, enabling integration into other projects and building trust without limiting core functionality access.
• •Production Deployment Reality: Successful robotics companies deploy tens to hundreds of robots in manufacturing for pick-and-place tasks, but focus on practical implementation over impressive demos to achieve working products.
• •Data Pipeline Bottlenecks: Robotics teams spend excessive time writing custom parallel jobs for basic queries that should be simple SQL operations, highlighting the need for specialized query engines for physical data.