Tactile Learning

We introduce velcro peeling as a representative application for robotic manipulating of non-rigid objects in complex environments.

We design strategies for peeling velcro strips placed on various surfaces such as planes and cylinders. Specifically, we formulate the problem as a Partially Observable Markov Decision Process and solve it using a multi-step deep recurrent network for the cases of tactile-only and visual feedback. We present a custom simulation setup and a real experiment setup, which is used to validate the strategies and compare them against benchmark strategies, including the ideal, fully-observable case. Our results show that tactile input can be used to overcome geometric uncertainties present in the environment when designing robust robotic manipulation controllers.