Dr. Lerrel Pinto, New York University

Robot Learning in the Wild
September 17, 2021


While robotics has made tremendous progress over the last few decades, most success stories are still limited to carefully engineered and precisely modeled environments. Interestingly, one of the most significant successes in the last decade of AI has been the use of Machine Learning (ML) to generalize and robustly handle diverse situations. So why don't we just apply current learning algorithms to robots? The biggest reason is the complicated relationship between data and robotics. In other fields of AI such as computer vision, we were able to collect diverse real-world, large-scale data with lots of supervision. These three key ingredients that fueled the success of deep learning in other fields are the key bottlenecks in robotics. We do not have millions of training examples in robots; it is unclear how to supervise robots and most importantly, simulation/lab data is not real-world and diverse. My research has focused on rethinking the relationship between data and robotics to fuel the success of robot learning. Specifically, in this talk, I will discuss three aspects of data that will bring us closer to generalizable robotics: (a) the size of data we can collect, (b) the amount of supervisory signal we can extract, and (c) diversity of data we can get from robots.


Dr. Lerrel Pinto

Assistant Professor of Computer Science at NYU. His research interests focus on machine learning and computer vision for robots. He received a PhD degree from CMU in 2019; prior to that he received an MS degree from CMU in 2016, and a B.Tech in Mechanical Engineering from IIT-Guwahati. His work on large-scale learning received the Best Student Paper award at ICRA 2016 and a Best Paper finalist award at IROS 2019. He is a recipient of research grants and awards from ONR, Amazon Science, and Honda Research. Several of his works have been featured in popular media like TechCrunch, MIT Tech Review, Wired, and BuzzFeed among others. Information about his research can be found on his website: https://www.lerrelpinto.com/.