Robotics founders are pursuing two distinct paths in physical AI development, each with different commercial implications. The data-first camp, including Physical Intelligence and Waymo, follows the large language model playbook of scaling with massive datasets. These teams typically come from machine learning backgrounds and focus on controlled environments where skills can be measured and benchmarked.

The architecture-first approach takes a different tack, building systems that handle real-world complexity from day one. Field AI and its team of NASA and Google Deep Mind veterans use Bayesian methods to quantify uncertainty and adapt to unpredictable conditions. Rather than constrain robots to lab settings, this method drops them into unstructured environments where they must operate safely without prior infrastructure.

Commercial traction currently favors the architecture-first builders. Their systems generate operational data across diverse real-world conditions, creating a self-reinforcing flywheel that data-hungry models lack. Customers gain operational intelligence that integrates with existing workflows, not just impressive demos.

The evidence on the ground suggests that operational deployments, not laboratory demonstrations, will determine which approach wins in bringing AI to physical environments.