Summary of GRASP Lab Days 2022 - Faculty Research Talks presents Dr. Dinesh Jayaraman

00:00:00 - 00:20:00

The video discusses research being conducted at the GRASP Lab into methods of improving the flow of information between the world and a robotic agent. The research covers areas such as active observation, representation of information, and decision-making. The projects highlighted in the video are focused on improving these three aspects of the flow of information.

• 00:00:00 Assistant Professor Dinesh Jayaraman discusses his research in computer vision and robotics. He discusses the importance of perception in general purpose robots, and how to scale visual perception with model-based controllers.
• 00:05:00 The video discusses the work of researchers at the GRASP Lab, who are developing algorithms that can abstract scenes and reproduce movements accurately. The researchers explain that key points are a valuable tool for representing scenes, as they can be used to identify important details in a scene without having to manually specify them.
• 00:10:00 The presenter discusses how to create a training objective that includes desiderata and recidrata, which is a key point pyramid structure that captures all the information about a body. The presenter also discusses how to enforce compactness and informativeness with graded course-finding.
• 00:15:00 Jason Ma and Jason Yan present research on task policies from data logs. The idea is to use an offline policy to minimize divergence between the set of states that is induced by the policy and the set of states produced by an expert. This is done by setting up a divergence minimization objective.
• 00:20:00 The video discusses research into machine learning techniques that can be used to improve the flow of information between the world and a robotic agent. It covers research into active observation and the representation of information, as well as the ability for agents to learn and make decisions. The three projects highlighted in the video are focused on improving these three aspects of the flow of information.