The Missing Data Stack for Physical AI

Niko West, the CEO of Rerun, introduces the concept of physical AI, emphasizing its current relevance and potential impact on various industries. He argues that while robotics is often tied to AI, many robotics applications are still lacking in true artificial intelligence capabilities. The conversation progresses to illustrate the role of physical AI in making technology more interactive and effective in real-world scenarios.

The discussion shifts to the Internet of Things (IoT), where Niko critiques the hype surrounding it over the last two decades, pointing out that it has not significantly transformed daily life. He contrasts this with physical AI, which he believes is on the verge of substantial economic impact due to its ability to address real-world complexities.

Niko describes the essential components for training physical AI systems, outlining the differences between online systems operating in real-time on robots and offline systems used for data analysis and training. He stresses the need for flexibility in AI applications to tackle the inherent messiness of real-world environments.

The complexities of data management in robotics are discussed, highlighting the challenges of integrating multi-modal data from various sensors and ensuring that this data is accessible for analysis and training. Niko emphasizes the importance of clear visualization tools to monitor data flows and identify issues in training pipelines.

Niko shares the origins of Rerun and the motivation behind open-sourcing their visualization and logging tools. He explains how the project aims to unify online and offline data systems for physical AI, allowing for better collaboration and debugging in robotic applications.

The podcast concludes with Niko outlining the future goals for Rerun and how the company aims to enhance the integration of its solutions into the robotics ecosystem. He mentions the importance of building a robust, flexible data stack that can handle the complexities of real-world data in robotics.