RoboticsRobot learningStory 02
A robot learns where to zoom next
What changedFoveaVLA replaces the wrist camera with learned high-resolution crops from external cameras, predicted alongside each action chunk. The robot keeps a wide scene view while requesting fine detail only where manipulation needs it.
The useful part
Why it matters
The robot keeps a wide scene view while requesting fine detail only where manipulation needs it.
Keep in mind
Good to know
The results are limited to the tested cameras, datasets, and manipulation suites and have not been independently replicated.
Evidence