Bill Freeman at CVPR: frontier robotics is moving from bounding boxes to end-to-end policies

yunta_tsai · x · 2026-08-27

MIT's Bill Freeman opened his CVPR talk with a list of "bitter lessons", stating upfront that training end-to-end beats everything else.

Commentators note traditional vision ML is effectively blind outside annotated boxes — throwing away 90% of information, an ImageNet-era approach that cannot teach nuances. As a result, most frontier robotics work is moving from boxes to end-to-end policies: teaching a model how to act matters far more than labeling boxes.

Related event: Autonomous Driving Shifts to End-to-End Vision as Bounding Boxes Fall Short(2 posts)→

Original post →

More from Embodied

Embodied channel →