Robots use cameras near their grippers to guide delicate actions, but fingers can block the view of an object. Images may also fail to show whether contact is being maintained.
Researchers added finger-mounted touch sensors, which measure contact directly, to robot control software that turns observations into actions. They compared it with camera-based control using the same training examples, without wrist-camera images for the touch-based version.
On three tasks involving hard-to-see contact, touch improved success by 31.7 to 60.0 percentage points, meaning the success rate rose by that many points. It exceeded vision with wrist cameras in observed success on all three tasks, suggesting touch can provide useful feedback when cameras cannot see the interaction.