Interactive Touch for Manipulation

Abstract

Towards helping people in daily life, robots need to better interact with our physical world and inevitably make contact with various objects. Touch provides contact geometry and forces information during interactions, which can be challenging to observe from vision due to occlusions or inherent limitations.

This thesis will focus on how to let robots leverage touch for manipulation in interactive means. We demonstrate several hardware platforms equipped with tactile sensing and integrated perception and control frameworks to apply interactive touch to real-world manipulation tasks. (1) We use touch for manipulating deformable objects like cables, using real-time tactile feedback during sliding. The robot can slide and pull the cable into different directions based on the tactile feedback to prevent falling. (2) We perform tactile exploration for learning the physical features of unknown objects. The extracted physical features are further applied to predict the forward model and swing up the in-hand object to a target pose by dynamic motions. (3) We embed tactile sensing with active rollers and design a 6-DoF roller grasper for better in-hand tactile dexterity. We demonstrate that the tactile-enabled roller grasper can robustly perform manipulation tasks for various objects, such as planar object reorientation, rolling along cables with tension, picking and singulating thin objects, etc. We hope applying interactive touch for manipulation can lead us closer to intelligent robot automation and the transformation of our physical world.