This paper presents a method to improve imitation learning in robots by focusing on data with a relatively low correlation to the desired output. The proposed approach inputs such data into each layer of the neural network, thereby amplifying its influence. This allows for more effective incorporation of diverse data sources into the learning process. The method has been tested through a simple pick-and-place operation using raw images and joint information as input, and has shown significant improvements in success rates, even with data from short sampling periods.
Publication date: 19 Jan 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2401.09691
