Chat@SSI: Foundation Models for Decision Making
Recent foundation models (e.g., DALL-E, CLIP, GPT-3, ChatGPT, ...) broke new ground and brought excitement in natural language understanding, computer vision, and various other domains. ChatGPT passed the Turing test ... almost, until we see its answer to a simple grade-school maths application question: perfect in the language, perfect nonsense in maths. Would foundation models enable intelligent decision making for robots operating in the physical world? How?
- They may act as advisors. How will the robot inform the advisor of the situation at hand and the task objective in order to solicit useful advice?
- They may supply common-sense knowledge. How will the robot incorporate this knowledge into its formal decision model and benefit from it?
- Or they may serve as a total replacement of the robot's decision model. How will the robot verify that these decisions are relevant, correct, or at least, free from harmful effects? Keep in mind GhatGPT's tendency to generate verbally-fluent, reasonable-sounding "garbage".
We are excited to announce this upcoming open discussion to explore the potential of foundation models for robot decision making in the physical world: both the opportunities and the perils. We will have two distinguished visitors joining us: Profs. Leslie Kaelbling and Tomas Lozano-Perez from MIT. Profs. Lee Wee Sun and David Hsu will participate in the discussion as well.
Everyone from the department is invited to join the talk and interact with us via Slido! The Slido session has already been activated. If you have any burning questions and/or interesting thoughts for discussion, feel free to post them on Slido starting today. We look forward to your active participation!