Yata Memorial Lecture : Boris Sofman : Journeys from Research to Commercialization

In this talk

1. 4:18

   CMU research and lasting foundations

   Sofman traces his robotics career to Carnegie Mellon, where he studied as an undergraduate and completed a PhD in 2010. His thesis explored ways for robots to improve perception and navigation using signals from their own operation. On the off-road Crusher project, he worked on extending perception from roughly 20 meters to 50–70 meters in difficult terrain. He says the problem-solving habits and relationships built at CMU lasted longer than any particular technology.

2. 9:06

   Anki’s consumer robots and business constraints

   Sofman co-founded Anki with CMU colleagues Hans Tappenier and Mark Palatucci to bring robotics into consumer products, using smartphone components to lower costs. Cozmo combined a compact robot with AI, games, and animation techniques developed with artists experienced in films such as WALL-E and Toy Story. The products found an audience, but holiday-driven sales, high inventory costs, and a lack of recurring revenue created severe cash-flow pressure. After financing challenges, Anki sold its assets, and Waymo hired much of its robotics and AI team.

3. 36:29

   Waymo’s path from research to driverless service

   At Waymo, Sofman led trucking and perception work, including technology used across cars and trucks. He describes how machine-learning systems trained on large amounts of driving data helped Waymo handle complex city conditions and expand from San Francisco to other cities and highways. The company also faced layoffs, paused trucking, and pressure to demonstrate commercial progress while spending heavily. Sofman says dependable service required extensive work beyond early demonstrations, especially in safety, operations, and evaluation.

4. 45:58

   Bedrock applies autonomy to heavy machinery

   Sofman co-founded Bedrock Robotics to automate heavy equipment, beginning with excavators used in construction. Bedrock retrofits existing machines, which can cost $500,000–$600,000, and aims to provide contractors with digital operators amid labor shortages. Its supervised systems collect data while performing earthmoving tasks, with a first driverless deployment planned for later in the year. Sofman sees potential to extend learned capabilities from excavation to tasks such as demolition and to other machine types.

5. 58:28

   Scaling through data, complexity, and trust

   In the Q&A, Sofman argues that robotics can be difficult to replicate because each application involves distinct machines, sensors, data, and safety requirements. He describes data reuse across Waymo cities and expects similar learning benefits as Bedrock expands to new tasks and equipment, while acknowledging that the amount of transfer remains uncertain. For construction sites, he also stresses making a machine’s awareness and intentions understandable to people nearby.