Drop to Transfer: Learning Transferable Features for Robot Tactile Material Recognition in Open Scene

Tactile feedback is an important way for robots to perceive the environment. In the open scenes, many uncertain factors lead to the change in the external environment, such as the reinstallation of the robot end-effector and the change of exploratory motions. This situation unavoidably raises the domain shift problem, which makes it difficult for the robot to fit the new open environment if it makes decisions only based on the knowledge extracted from the closed environment. Thus, we propose a new deep unsupervised domain adaptive material recognition model, called drop to transfer (DTT), which adaptively learns tactile transferable features to achieve accurate material recognition in open scenes. DTT performs the clustering hypothesis on the target domain samples using the adversarial dropout (AdD). And in each training iteration, high-gradient features are progressively removed to enhance the expression of other transferable features. Meanwhile, the maximum mean discrepancy (MMD) metric is added in the optimization process to achieve feature alignment between the source and target domain distributions. In addition, we design an adaptive weighted optimization strategy to optimize each loss in DTT effectively. Experimental results show that the proposed method can effectively learn the transferable features and reduce the domain shift problem for the robot tactile perception task in the open scenes.