Collaborative Research: Differentiable and Expressive Simulators for Designing AI-enabled Robots

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

This project develops a novel principled and systematic approach to closing the so-called sim-to-real gap in robotics. The work delivers a breakthrough in future robot technologies, as well as Computer Aided Engineering, which anchors US industries such as automotive, aerospace, and space exploration. Therefore, results from this award bring benefits to the society and increase US competitiveness. This project produces knowledge and establishes algorithms that enable computer simulation to reduce design time and costs in engineering intelligent and safe robots. Artificial intelligence (AI) is poised to endow the next generation of robots with mobility and decision-making skills. However, experimentation and acquisition of large amounts of training data encapsulating physical interaction between the robot and its environment can be costly and sometimes risky to the robots and practitioner. Existing simulation techniques provide virtual testbeds in which the robot can learn efficiently and safely, but have one major drawback - robots designed via computer simulation operate differently, often less effectively, when deployed to the real world, which is called the sim-to-real gap. This award supports fundamental research to reduce, and whenever possible eliminate, the sim-to-real gap, unlocking the full potential of AI-enabled robots. The new modeling techniques and numerical algorithms developed are able to accurately capture the phenomena of the real world by combining well-known physics equations and empirical traits learned from data. The project also advances the state of the art in robotic technologies and broaden the participation in computing of high-school students from underrepresented groups. Presently, robotic simulators are hampered by two main problems. First, the robot models used are not accurate and expressive enough. Second, the methods that seek to improve these models are both laborious and robot specific. This research aims to create a new generation of expressive and differentiable simulators to address both problems. A simulator is expressive if it has the ability to capture the physics of interest. A “differentiable simulator” is one that in addition to predicting the time evolution of a robot, can also produce gradient information, thus allowing the simulator to be automatically adjusted to better match the real world. The research team focuses on establishing the mathematical foundation, prototyping robotics simulators with learnable components, investigating practical ways to leverage these simulators in robotic applications, and evaluating the impacts of expressive and differentiable simulators on closing the sim-to-real gap. This project is supported by the cross-directorate Foundational Research in Robotics program, jointly managed and funded by the Directorates for Engineering (ENG) and Computer and Information Science and Engineering (CISE). 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.