- TBA, graduate students (University of Stuttgart)
- Optionally Vien Ngo (University of Stuttgart)
Marc Toussaint is professor for Machine Learning & Robotics at University Stuttgart. The “team” will be very small, perhaps only include another undergraduate student and optionally a post-doc of mine. This is intended, as this team will focus on testing and evaluating a very specific set of methods and existing code developed by Toussaint (see below). These methods focus on trajectory optimization and optimal control (including grasping from a control perspective) and are therefore ideally suited to address challenge 2.
Optionally, we will also develop/transfer new methods for belief planning (similar to the recent excellent work of Platt, Lozano-Perez, Kaelbling et al) to cope with that part of the challenge that involves perceptual uncertainty. In this case, a post-doc of mine, Vien Ngo (expert in Bayesian RL) will join us to work on this.
In our lab we have a PR2, which we also use as test platform of our control and motion optimization methods. We expect that the transfer to the infrastructure of challenge 2 requires modest effort.
We will focus on exploring and testing constrained optimization methods for optimal control and real-time motion optimization. Concerning trajectory optimization, we will specifically focus on applying Toussaint's k-order Markov Motion Optimization code (KOMO, the code of which is already available here: http://ipvs.informatik.unistuttgart. de/mlr/marc/source-code/14-KOMO.pdf), which formalizes motion optimization problems in a rather general form and implements state-of-the-art constrained optimization methods, esp. a novel Augmented Lagrangian method, to solve these problems efficiently..
Concerning control, we equally follow the path to formulate them as rigorous constrained optimization problems, which are quick so solve in a 1msec cycle. Both, the trajectory optimization and control methods are, as a side-story, already published in Toussaint et al.: Dual Execution of Optimized Contact Interaction Trajectories. (IROS 2014). But the code is much more general than what is described in this publication.
Is our resources permit (in terms of time of Vien Ngo), we will also transfer the above mentioned belief planning approaches to our KOMO code.