CAREER: Concurrent Robot Learning from Simulation and Real for Closing the Sim-to-real Gap

像机器狗或人形机器人一样，有腿的机器人可以很好地完成室内和室外的任务，比如重新布置家里的家具或监控工厂。然而，控制这些机器人是困难的，并且伴随着机器人摔倒的风险。近年来，人工智能（AI）在通过计算机模拟教机器人如何行走来解决跌倒挑战方面显示出了希望。然而，由于仿真与真实物理机器人之间的差异，在仿真中学习到的机器人行为在现实世界中往往表现不佳。这个学院早期职业发展（Career）项目支持研究一种同时从模拟和物理实验中学习的新方法。在这个项目中，这种新颖的方法旨在利用模拟的优势，如可扩展性，同时使用来自真实机器人的数据。最终，该项目将最大限度地发挥人工智能和机器人学习算法的潜力，使机器人更有能力、更安全。此外，该项目还结合了教育活动，让学生参与到真实的机器人学习环境中，培养对机器人的更广泛理解。该项目将研究一种新型的学习算法，从多元宇宙中学习（LfM），它同时从大规模、廉价的基于物理的模拟和昂贵的现实世界中学习，以弥合众所周知的“模拟到真实”的差距。该项目的基本原理包括从模拟和真实经验中无缝和持续地学习，以及对实验数据进行结构化的数学推理。这种方法不同于大多数现有的学习算法，这些算法仅仅依赖于模拟或现实世界的经验。本研究涉及三个主要组成部分的开发：(i)现实世界中的自主和安全学习环境，（ii）同时利用模拟和现实的新型机器人学习算法，以及（iii）可解释的人工智能来理解硬件实验。该项目将促进对安全至关重要的有腿机器人（如带机械手的四足机器人和两足机器人）具有挑战性的运动技能的开发，同时确保机器人有效、高效和安全的学习。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Legged robots, like robot dogs or humanoid robots, offer the possibility of being able to move well in indoor and outdoor tasks, such as rearranging furniture at home or monitoring factories. However, controlling these robots is difficult and comes with the risk of robots falling. In recent years, artificial intelligence (AI) has shown promise in addressing the falling challenge by using computer simulations when teaching robots how to walk. However, the robot behaviors learned in simulation often underperform in the real world due to the difference between the simulation and the real physical robot. This Faculty Early Career Development (CAREER) project supports research that investigates a novel approach that simultaneously learns from both simulated and physical experiments. This novel approach in this project aims to leverage the advantage of simulations, such as scalability, while using the data from real robots. Ultimately, the project will maximize the potential of AI and robot learning algorithms to allow more capable and safer robots. Additionally, the project incorporates educational activities to engage students with real-world robot learning environments, fostering a broader understanding of robotics.This project will investigate a novel class of learning algorithms, Learning from Multiverse (LfM), which concurrently learns from both large-scale, inexpensive physics-based simulation and expensive, real world ground-truths to bridge the well-known “sim-to-real” gap in legged robots. The fundamental principle of the project involves seamless and continuous learning from both simulated and real experiences, as well as structured mathematical reasoning of experimental data. This approach differs from the majority of existing learning algorithms that rely solely on either simulation or real-world experience. This research involves the development of three main components: (i) autonomous and safe learning environments in the real world, (ii) novel robot learning algorithms that simultaneously leverage simulation and reality, and (iii) explainable artificial intelligence to understand hardware experiments. This project will facilitate the development of challenging motor skills for safety-critical legged robots, such as quadrupedal robots with manipulators and bipeds, while ensuring effective, efficient, and safe robot learning.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.