Overcoming the Sim-to-Real Gap in Autonomous Robots

摘要

When designing and planning an autonomous, adaptive system utilizing the benefits of Artificial Intelligence (AI), both the specific field of AI and the desired use case must be addressed to achieve a reliable system. In the field of autonomous robotics, the authors focus on the aspect of training a robot not in reality, where training is reliant on real-time and the amount of hardware systems performing the training, but in a virtual environment, which can be parallelized and customized individually. In this paper, a functional simulation of a real robot was developed and trained using reinforcement learning in a virtual environment to successfully complete a specified task. The trained model was then transferred to the physical robot, which itself was able to complete the specified task in the real world, without the need of previous real-world training. Due to extensive observation and analysis of the physical system’s characteristics, the presented direct transfer approach does not rely on domain randomization, which is usually applied in this field.