Wed, Jun 17, 2020 @ 04:00 PM - 05:30 PM
Ph.D. Defense - Max Pflueger 6/17 4:00 pm "Learning from Planners to Enable New Robot Capabilities"
Ph.D. Candidate: Max Pflueger
Date: Wednesday, June 17, 2020
Time: 4:00 pm - 5:30 pm
Committee: Gaurav S. Sukhatme (chair), Joseph Lim, Sandeep Gupta, Ali Agha
Title: Learning from Planners to EnableNew Robot Capabilities
Solving complex robotic problems often involves reasoning about behavior multiple steps in the future, yet many robot learning algorithms do not incorporate planning structures into their architecture. In this dissertation we show how we can harness the capabilities of planning algorithms to learn from the structure of the robotic problems we wish to solve, thus expanding beyond what was available from baseline planners. We consider problems in multi-arm manipulation planning, path planning for planetary rovers, and reinforcement learning for torque controlled robots, and show how in each case it is possible to learn from the behavior of planning algorithms that are limited and unable to solve the full generalized problem. Despite not being full solutions these planners provide useful tools and insights that can be leveraged in larger solutions.
In multi-step planning for manipulation we develop a high level planner that can find solutions in difficult spaces by solving sub-problems in sub-spaces of the main planning space. For planetary rovers show how to use inverse reinforcement learning to learn a new planning algorithm that can function on different (and generally cheaper) input data. Reinforcement learning algorithms often suffer from unstable or unreliable training, we show how this can be mitigated by augmenting the robot state with a state embedding space learned from planner demonstrations.
Planning and control algorithms often rely on rigid and prescribed assumptions about the nature of robot problems, which may not be suitable for the generalized and versatile robot systems we wish to build. However, as this dissertation argues, those structures are still useful in informing the behavior of more flexible families of algorithms.
Meeting ID: 938 1684 0971
Google Meet (ONLY A BACKUP - IF WE EXPERIENCE PROBLEMS WITH ZOOM):
WebCast Link: https://usc.zoom.us/j/93816840971
Audiences: Everyone Is Invited
Contact: Lizsl De Leon