Humans are always learning new tasks from demonstrations, instructions, and even trial and error. In contrast, current AI systems are bound to the specific tasks for which they were originally programmed. We have seen great advances in the ability of AI systems to learn to do individual tasks at human expert levels (e.g., Chess, Go, and video games), but they are hopeless when presented with a new task. As AI systems become more prevalent, we need to start asking how can we teach them to do what we want. Join us for an evening to discuss research on how we are building AI systems that learn new tasks through natural language instruction and demonstrations.

What is the state-of-the-art in teaching robots new tasks through natural interaction?
What are the key challenges?
What are the potential applications and who will benefit?