How we Build and Evaluate Systems for Autonomous Earthwork

Excavation is a craft rather than a subject you can learn from a textbook. Skilled operators spend thousands of hours reading soil, weather, and the machine itself, optimizing dozens of variables at once and adjusting on the fly. Translating that judgment into something a team can specify, measure, and test against is one of the harder problems in autonomy. In this conversation, Meghan McDowell (Head of Program Management, Bedrock Robotics) talks with Dave Garber (Product Manager) and Vincent Gonguet (Head of Product) about how a task gets chosen, built, and evaluated. They cover why Bedrock started with mass excavation, the three things a partner needs from a machine on a live site (trustworthiness, dependability, and productivity), how simulation shortened the feedback loop from weeks in the field to hours, and why "done" may be the wrong frame for a system that keeps learning from every deployment.