{"id":793331,"date":"2021-11-16T08:00:04","date_gmt":"2021-11-16T16:00:04","guid":{"rendered":"https:\/\/www.microsoft.com\/en-us\/research\/?post_type=msr-research-item&p=793331"},"modified":"2021-11-08T10:45:23","modified_gmt":"2021-11-08T18:45:23","slug":"keynote-learning-from-observation-small-data-approach-to-human-common-sense","status":"publish","type":"msr-video","link":"https:\/\/www.microsoft.com\/en-us\/research\/video\/keynote-learning-from-observation-small-data-approach-to-human-common-sense\/","title":{"rendered":"Keynote: Learning from observation: Small-data approach to human common sense"},"content":{"rendered":"
Speaker: Katsushi Ikeuchi, Sr. Principal Research Manager, Microsoft Research Redmond<\/p>\n
Learning-from-Observation (LfO), a robot-teaching paradigm, aims to build a robot system that understands what humans do through a small number of human observations and map them to action. Unlike the more popular Learning-from-Demonstration paradigm, in which the robot itself is directly manipulated and the demonstration is repeated many times, Learning-from-Observation uses accumulated robotics knowledge from a small number of demonstrations through explicit steps and formulates it into task models to achieve a goal. In this session, we\u2019ll briefly introduce the basic design concepts of specific task models, shared and formulated through earlier system designs. We\u2019ll then discuss the formulation of common sense, required to achieve household tasks such as wiping a tabletop or bringing a cup of tea without spilling, into explicit task models. We\u2019ll also discuss task recognition through these task models from visual and verbal cues and the implementation of robust execution modules pre-trained by reinforcement learning.<\/p>\n