A Software Suite for Sample-Efficient Robotic Reinforcement Learning

View the PDF file from the paper entitled Serl: a Software Suite to learn effective automatic enhancement of the sample, written by Jianlan Luo and 9 other authors

PDF HTML (experimental) view

a summary:In recent years, significant progress has been made in the field of automatic reinforcement learning (RL), enabling ways to deal with complex image notes, training in the real world, and combining auxiliary data, such as demonstrations and previous experience. However, despite these developments, Robotic RL is still difficult. It is recognized among practitioners that the special implementation details of these algorithms are often the same importance (if not more than that) for performance such as choosing algorithm. We assume a major challenge to ROBOTION RL, as well as further development of RL methods, is the lack of comparative access to these methods. To face this challenge, we have developed a carefully implemented library containing an active RL method outside politics, as well as ways to have bonuses and environmental reset, a high -quality control unit for a widely approved robot, and a number of difficult examples. We provide this library as a community supplier, half of its design options, and provide experimental results. Perhaps we find that our implementation can achieve a very effective education, gain policies to assemble a multi -chlorine vinyl panel, direct cables, transport objects between 25 to 50 minutes of training per average policy, and improve the latest results reported for similar tasks in the literature. These policies achieve perfect or almost perfect success rates, extreme durability even in light of the disturbances, and emerging recovery and correction behaviors show. We hope to provide these promising results and our high -quality open -quality implementation tool for a robot community to facilitate more developments in Robotic RL. Our code, documents and videos can be found in this URL https