Schedule

Abstract:

In this talk, I will discuss the pros and cons of using different temporal logics as formal task specifications for robotics. Specifically, I will describe what these formalisms can capture (goals, safety, constraints), what we need to assume (abstraction, grounding), and how we can use them to create provably-correct robot behaviors.

Abstract:

While end-to-end, fully autonomous learning is interesting to explore, for real-world applications, including robotics, the paradigm of human-in-the-loop learning has emerged as a practical way of guiding and speeding up the learning process - and in effect refining the task specification on the fly.  This talk will introduce some recent human-in-the-loop learning algorithms that enable robust navigation in challenging settings, such as in densely cluttered environments and over varying terrains.  While most of these algorithms take explicit input from human trainers, the talk will close with a new paradigm for reinforcement learning from implicit human feedback, specifically observed facial expressions.

Abstract:

Scalar rewards are a standard mechanism for capturing the goals of an agent as articulated by the reward hypothesis, which states: “all of what we mean by goals and purposes can be well thought of as maximization of the expected value of the cumulative sum of a received scalar signal (reward).” In this talk, I highlight the implicit commitments we make when we write down a reward function to capture the goals of an agent, and examine their consequences for task specification more generally.

Abstract:

Task specifications are always ambiguous. No instruction is interpreted in a vacuum, whether due to physical or social context. While we lack models that have this level of awareness, we are beginning to see some components taking shape. In this talk, I'll outline this larger vision for robots with awareness, and discuss some concrete initial work on key dimensions. Specifically, I'll touch on several recent works looking at instructing robots with language, instructing via video demonstration, and building robots that understand their own soundscape. This last piece adds a new twist to understanding your own impact on the environment around you. The bulk of the work in this talk will be based on papers from my student Vidhi Jain: https://vidhijain.github.io/

Abstract:

In this talk I will argue that general-purpose robots should not only pursue our goals but should also learn to represent and pursue their own – they should be autotelic. This approach is motivated by both scientific interests in modeling human open-ended development and the practical need for robots to adapt to diverse, unpredictable environments. I will discuss recent advances in goal-conditioned learning and automatic curriculum generation, emphasizing how self-generated goals can drive open-ended skill learning. After exploring existing methods for evolving goal representations, I will discuss open challenges towards truly open-ended autotelic learning.