Video Action Models: Leveraging Pre-Trained Video Models for Robotics

What if robot learning could pre-train on freeform human video instead of relying solely on expensive robot action data?

Elvis Nava, co-founder and CTO of mimic robotics, introduces Video Action Models, a model class that repurposes generative video model backbones for general-purpose dexterous robot control.

You'll leave with a view of a new frontier for robot learning.