Real-to-Sim-to-Real: A Scalable Data Diet for Robot Learning

Real-to-Sim-to-Real: A Scalable Data Diet for Robot Learning

Abstract: Robotic automation, powered by machine learning driven methods, has the potential to build systems that change the future of work, daily life and society at large by acting intelligently in human centric environments. As with most modern machine learning methods, a key component in building such a robotic system the availability of data, abundant, diverse and high quality. In domains of nature language or computer vision, data of this form has scaled passively with internet scale, since people naturally interact through the medium of language or images. In contrast, robots are hardly deployed in human-centric settings and certainly are not collecting internet scale data passively. The key question I will ask is - how can we develop a data diet for robotic learning that scales passively? In particular, I will discuss how simulation, despite being fundamentally inaccurate, can provide a scalable source of data for robotic learning. We will discuss a class of real-to-sim-to-real methods that are able to construct simulation content on the fly from cheap real-world data, enabling scalable robust robot training. In doing so, I hope to shed some light on the unique challenge that data acquisition plays in robot learning and discuss how developing truly open-world robotic learning systems will necessitate a careful consideration of data quality and quantity.

Bio: Abhishek Gupta is an assistant professor in computer science and engineering at the Paul G. Allen School at the University of Washington since 2022. He lead the Washington Embodied Intelligence and Robotics Development lab focusing on robot learning and reinforcement learning. Previously, he was a postdoctoral scholar at MIT, collaborating with Russ Tedrake and Pulkit Agarwal. Prior to that he received his Ph.D. and B.S degrees from UC Berkeley, working with Sergey Levine and Pieter Abbeel. Abhishek is the recipient of Toyota Research Institute Young Investigator award and an Amazon Science Hub award, along with award nominations at several top conferences and workshops. His research interests lie in scalable reinforcement learning methods for robot learning, in particular methods for continual adaptation in the real world.