Meta-learning Haptic Exploration of Simple 3D Objects

The vision of this work is to build a generalized model of haptic exploratory procedures that can be employed by a wide range of robotic platforms in different environments. Main challenge of combining deep learning and robotics is currently the time-efficiency of sample acquisition. Having to learn to perform a sequence of interactions with the environment, in which next action may depend on the previously acquired data, adds an extra complexity to the task of quickly teaching a robot to perform a haptic task. A meta-model for exploratory procedures built with data acquired by multiple robot platforms in simulation and online, that can be quickly adapted to a new robot or task, may offer a solution to this challenge. Here, we describe our approach to meta-learning of a haptic exploration policy offline, and demonstrate preliminary results achieved after only two gradient steps performed online with data acquired by a robot. We exemplify our approach on the task of learning to classify four simple three-dimensional objects. The meta-learner’s learning rule enables us to learn a generalized policy representing experience from different types of data available offline, and quickly adapt it to a new task – online classification of objects based on a sequence of tactile sensor measurements acquired by a robot. Tactile measurements are acquired through performing of three haptic glances per object at poses generated by the trained closedloop control policy. According to the principles of interactive perception, we optimize the closed-loop location control policy, integration of the data from the past iterations, and a classifier that classifies the objects all in one go. We employ the model resulting from a meta-training to control haptic exploration on the real robot. The classification accuracy achieved after two gradient updates of the meta-model performed with the data acquired by the robot is 93.7 % evaluated on a batch of 16 samples.