Perceptual Information

VLA reasoning models and other state-of-the-art perception systems achieve high performance in the domains they were trained on. The progress is real, but they don't fully generalize out of distribution. Even small variations in conditions or tasks can still break them. This is the area where we work.