From (Inverse) Optimal Control to Joint (Cooperative) Sequential Manipulation and Motion Planning
MetadataShow full item record
With this talk, my aim is to give a more integrated view on my research. I will start with the planning-as-inference framework and its implications in the context of stochastic optimal control. While this, in principle, allows us to exploit any probabilistic inference method, in concrete robotics applications I typically use the Laplace approximation, which leads back to a path optimization problem. I will discuss efficient, 2nd-order constraint path optimization methods that exploit the problem structure and our work on inverse optimal control based on this framework. Switching to higher-level reasoning, I will then briefly discuss our work on learning and planning in relational Markov decision processes—the core open problem for applying this in real-world robotics clearly is the symbol grounding or acquisition problem. Our recent work on logic-geometric programming, which proposes a joint-path optimization and task planning formulation, combines a first-order symbolic description of the kinematic structure of sequential manipulation into the mathematical program over a path. This joint formulation of symbolic decision and path optimization problems also guides our recent work on human-robot cooperative manipulation. Throughout the talk I will mention more applications we work on, including exploring the environment to find manipulable degrees of freedom, manipulation skill learning, manipulation POMDPs, inverse optimal control, and cooperative assembly of IKEA items.
- IRIM Seminar Series