[2G1-ES-4] Robots and real worlds: Applied machine learning

[2G1-ES-4-02] Bootstrapping Baysian Inverse Reinforcement Learning in Robotics through VR Demonstration

〇Reed Sogabe^1,3, Tomoaki Kimura¹, Dinesh Malla^2,1, Masaru Sogabe², Katsuyoshi Sakamoto¹, Tomah Sogabe^1,2 (1. UEC, 2. Grid Inc., 3. K. International School Tokyo)

Keywords:Bayesian inverse reinforcement learning, robot arm, HTC-Vive demonstrations

Sparse rewards have been a persistent problem in reinforcement learning (RL). In many cases, one has to manually specify or shape the reward function, which greatly limit the application of RL to real-world tasks which are usually possessing long task horizon and high action dimensionality which makes manual setting of reward function extremely difficulty. In this work, we propose to overcome the sparse reward problem by using Bayesian inverse reinforcement learning which simulate and infer the reward from the suboptimal demonstration. We use deep deterministic policy gradients and hindsight experience replay algorithm along with HTV-Vive interface technique at the same frequency as displayed in the ROS environment to adaptive control 7-DOFCrane-X7 robot arm. We show the proposed method is able to solve various fetch tasks and learnt superior policy to the demonstrator policy.