Adaptive Absolute-Relative Rating for Noise Rejection in Behavioral Cloning based on Tsallis Statistics

In robot control from demonstrations, a sufficient dataset cannot be collected for many of the tasks that require experts with qualifications to special skills. Unfortunately, insufficient expert dataset would manifest various types of noise hidden in it. Since adding data is difficult as well, offline imitation learning needs to be robust to such a noise. In the conventional work, a behavioral cloning method based on Tsallis statistics has been developed. However, it weights each data with absolute rating with a fixed threshold, which would fail to imitate coarse/diverse motions. Therefore, this paper improves the conventional method by adding the function of relative rating for each data, which should enable robots to imitate non-noisy data even from coarse/diverse motions. This function can be obtained from a different derivation way of the optimization problem with Tsallis statistics. By integrating it with the conventional derivation way, the proposed method can adjust between the absolute and relative ratings. Finally, for more convenience, we design optimization tricks for the hyperparameters to maximize the variance of weights with avoiding extremely large weights. In numerical simulations and real-robot experiments, we demonstrate the robustness of the proposed method.