Abstract
Robotic automation for pick and place task has vast applications. Deep Reinforcement Learning (DRL) is one of the leading robotic automation technique that has been able to achieve dexterous manipulation and locomotion robotics skills. However, a major drawback of using DRL is the Data hungry training regime of DRL that requires millions of trial and error attempts, impractical in real robotic hardware. We propose a multi-subtask reinforcement learning method where complex tasks can be decomposed into low-level subtasks. These subtasks can be parametrised as expert networks and learnt via existing DRL methods. The trained subtasks can be choreographed by a high-level synthesizer. As a test bed, we use a pick and place robotic simulator, and transfer the learnt behaviour in a real robotic system. We show that our method outperforms imitation learning based method and reaches high success rate compared to an end-to-end learning approach. Furthermore, we investigate the trained subtasks to demonstrate a adaptive behaviour by fine-tuning a subset of subtasks on a different task. Our approach deviates from the end-to-end learning strategy and provide an initial direction towards learning modular task representations that can generate robust behaviours.
Abstract (translated)
URL
https://arxiv.org/abs/2102.04022
