2020/08/26 by Melisa Idil Sener, Yukie Nagai, Sener, Melisa +5
Computer Science · Engineering · #FOS: Computer and information sciences #Reinforcement Learning in Robotics #Robotic Path Planning Algorithms #Robotics (cs.RO) #Robotics and Automated Systems
paper · pdf · doi:10.48550/arxiv.2008.11503
openalex publication_date 2020/08/26 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
One effective approach for equipping artificial agents with sensorimotor\nskills is to use self-exploration. To do this efficiently is critical, as time\nand data collection are costly. In this study, we propose an exploration\nmechanism that blends action, object, and action outcome representations into a\nlatent space, where local regions are formed to host forward model learning.\nThe agent uses intrinsic motivation to select the forward model with the\nhighest learning progress to adopt at a given exploration step. This parallels\nhow infants learn, as high learning progress indicates that the learning\nproblem is neither too easy nor too difficult in the selected region. The\nproposed approach is validated with a simulated robot in a table-top\nenvironment. The simulation scene comprises a robot and various objects, where\nthe robot interacts with one of them each time using a set of parameterized\nactions and learns the outcomes of these interactions. With the proposed\napproach, the robot organizes its curriculum of learning as in existing\nintrinsic motivation approaches and outperforms them in learning speed.\nMoreover, the learning regime demonstrates features that partially match infant\ndevelopment; in particular, the proposed system learns to predict the outcomes\nof different skills in a staged manner.\n