In this presentation, Sourav Panda presents Fast and Data Efficient RL from Pixels Using Non-Parametric Value Approximation, by Long et al. In it, the authors introduce a way to learn action selection in RL domains directly from pixel-level feature data that is highly sample efficient, which is a large worry for RL techniques.
Presentation Link: https://psu.mediaspace.kaltura.com/media/Group+Meeting/1_5hhsoyry