According to the abstract, FoveaVLA is a Vision-Language-Action (VLA) model that replaces the wrist camera with learned foveation actions. Wrist cameras are commonly used in robot manipulation for close-up detail, but they limit the field of view to the end-effector's immediate surroundings.
The abstract notes that external cameras are part of the setup, but the text cuts off before explaining how they are used. The key idea is that foveation actions—presumably actions that control where the camera looks—can provide the needed detail without a physical wrist camera.
The work directly addresses a known trade-off in robot vision: the tension between close-up detail and a broader view. The available excerpt does not include experimental results or further implementation details.