ActiveScale: Scaling Active Perception for Robots
across Model, Data, and Hardware

ActiveScale presents a scalable way to study active perception for robot manipulation across model, data, and hardware.
It equips a vision-language-action model with temporal camera observations and camera-pose prediction,
and introduces the Active-perception Mobile-manipulation Platform (AMP) for coordinated viewpoint and manipulation control.

Checkpoints

Checkpoint	Description
midtraining	Mid-training checkpoint
bag	Bag task checkpoint
drawer	Drawer task checkpoint
pot	Pot task checkpoint
box	Box task checkpoint
under	Under-table task checkpoint

Code and usage: github.com/ShuaiZhou302/ActiveScale