Oct 8, 2021 - Karthik Narasimhan, Princeton (in-person)

Language-guided policy learning for better generalization and safety

Abstract: Recent years have seen exciting developments in autonomous agents that can understand natural language in interactive settings. As we gear up to transfer some of these advances into real-world systems (e.g physical robots, autonomous cars or virtual assistants), we encounter unique challenges that stem from these agents operating in an ever-changing, chaotic world. In this talk, I will focus on our recent efforts at addressing two of these challenges through a combination of NLP and reinforcement learning — 1) grounding novel concepts to their linguistic symbols through interaction, and 2) specification of safety constraints during policy learning. First, I will demonstrate a new benchmark of tasks we designed specifically to measure an agent's ability to ground new concepts for generalization, along with a new model for grounding entities and dynamics without any prior mapping provided. Next, I will show how we can train control policies with safety constraints specified in natural language. This will encourage more widespread use of methods for safety-aware policy learning, which otherwise require domain expertise to specify constraints. Scaling up these techniques can help bring us closer to deploying learning systems that can interact seamlessly and responsibly with humans in everyday life.

Bio:  Karthik Narasimhan is an assistant professor in the Computer Science department at Princeton University. His research spans the areas of natural language processing and reinforcement learning, with a view towards building intelligent agents that learn to operate in the world through both their own experience and leveraging existing human knowledge. Karthik received his PhD from MIT in 2017, and spent a year as a visiting research scientist at OpenAI prior to joining Princeton in 2018. His work has received a best paper award at EMNLP 2016 and an honorable mention for best paper at EMNLP 2015.

----------------------------

Nov 5, 2021 - Chuchu Fan, MIT

Nov 19, 2021 - Aaron Ames, Caltech

Dec 3, 2021 - Naomi Leonard, Princeton