Large behavioral models in AI - Tom Burggraf

We dig into how Toyota Research Institute trained a new kind of robot model — not a language model, but a large behavioral model that learns physical tasks from demonstration. No scripting step-by-step movement. No hard-coded routines. Just “Set the breakfast table,” and the robot figures it out.

Tom Berggraf, Head of Product for AWS Batch, explains how TRI used batch + SageMaker integration to train these models efficiently at scale, and why this matters for real-world robotics: elder care, home tasks, industrial maintenance, and any future where robots work alongside people.

This is a clear shift: robots that learn by watching, adapt, and execute complex sequences smoothly. It’s the closest we’ve seen yet to “useful domestic robots” that anyone could command.

If you have ideas for technical topics you’d like to see us cover in a future show, let us know by emailing us at ask-hpc@amazon.com.