Discussions
Discussion Topics
[Learning Frameworks] The three major frameworks for LfD have been identified as 1) learning a state-action policy, 2) learning a cost/reward function and 3) learning a plan. How should we choose between these methods given different tasks/problems? How do we best combine different frameworks? Drawing inspiration from cognitive science and psychology, are there any other different frameworks we should consider?
[Human Models] Do we need more sophisticated models of humans than assuming they are near optimal? Is human behavior too complex/random to model otherwise? How does LfD benefit from theory-of-mind?
[Assumptions of LfD] Is LfD inherently limited by the performance of demonstrators? Can we extrapolate beyond the demonstration provider? How to account for the sub-optimality in human demonstrations? How to learn from multiple demonstrators? Do we need stronger inductive priors? If so, which is the right one? Physics? Causality?
[Additional Sensory Modalities] Are state-action pairs sufficient for imitation learning or are there fundamental ambiguity limits we will hit unless we consider additional information? Do we need to explicitly model data from additional sensory modalities in order to efficiently leverage the information? How do we collect such data without putting burdensome sensors on the demonstrator?
[Scaling-up LfD] Are there any fundamental limits to learning individual tasks from imitation? What are the major challenges in meta learning, transfer learning and continual learning from demonstrations for robotics? How do we leverage demonstrations provided in simulated/virtual environment?
Share your thoughts and questions with us! We will use these to moderate workshop discussion sessions!