ArrayBot: Reinforcement Learning for Generalizable Distributed Manipulation through Touch

The notion of robotic manipulation [1, 2] easily invokes the image of a biomimetic robot arm or hand trying to grasp tabletop objects and then rearrange them into desired configurations inferred by exteroceptive sensors such as RGBD cameras. To facilitate this manipulation pipeline, the robot learning community has made tremendous efforts in either how to determine steadier grasping poses in demanding scenarios [3, 4, 5, 6, 7] or how to understand the exteroceptive inputs in a more robust and generalizable way [8, 9, 10, 11, 12, 13]. Acknowledging these progresses, this paper attempts to bypass the challenges in the prevailing pipeline by advocating ArrayBot, a reinforcement learning driven system for distributed manipulation [14], where the objects are manipulated through a great number of actuators with only proprioceptive tactile sensing [15, 16, 17, 18]. Conceptually, the hardware of ArrayBot is a 16 16 array of vertically sliding pillars, each of which can be independently actuated, leading to a 16 16 action space. Functionally, the actuators beneath a tabletop object can support its weight and at the same time cooperate to lift, tilt, and even translate it through proper motion policies.