The Stack Overflow Podcast Building brains for bulldozers
17 snips
Mar 6, 2026 Kevin Peterson, CTO of Bedrock Robotics and veteran robotics engineer from Waymo, talks about upgrading heavy construction gear with autonomy. He discusses model choices and training approaches, the role of simulation versus real data, and the engineering challenges of rugged on-machine compute and sensors. He also covers scaling safety and how automation can tackle construction labor shortages.
AI Snips
Chapters
Transcript
Episode notes
From Psychology Major To Self Driving Pioneer
- Kevin Peterson began in psychology at Carnegie Mellon and pivoted after seeing a poster for a dune buggy with lasers, which led to his first self-driving car project.
- He later worked on self-driving trucking, spun tech out to Caterpillar to automate a mine truck, and amassed millions of miles of productive robotic operation.
Train Robots Like Language Models Using Action Tokens
- Modern robotics models borrow LLM-style thinking: actions are the tokens and multimodal inputs matter as much as architecture (CNN vs transformer).
- Kevin emphasizes training on sequences of actions (e.g., how to move an arm or dig) rather than only static perception.
Excavators Are The Most Interesting Manipulators
- Bedrock targets construction because excavators are complex manipulators that change terrain and require diverse behaviors, not single-mode decisions like driving.
- The team blends imitation learning and reinforcement learning to scale diverse digging and manipulation skills.