Mobile Manipulation Research Platform
Final Report Abstract
We compiled a list of our publications and projects that demonstrate our use of the PR2, including embedded videos, here: https://ipvs.informatik.uni-stuttgart.de/mlr/pr2-report/ The most important research results and published work based on the PR2 is as follows: Manipulation Learning from Demonstration: Peter Englert developed a series of very succesfuls methods for manipulation learning from robot demonstrations. The typical setup was that a human uses kinesthetic teaching to demonstrate a manipulation task with the PR2, e.g. opening a door, or pushing a box. From this demonstration we developed methods to extract an underlying cost function that explains the demonstration (Inverse KKT), use Bayesian Optimization to improve on the demonstration, and combine both to have a system that learns manipulation skills from very few demonstrations. Throughout this work Peter used the PR2 as the experimental platform. PR2 as platform for Human-Robot interaction and symbol learning research: We more recently started work on human-robot interaction. In a first study, lead by Andrea Baisero, we investigated learning to relate words to geometric features and object identities. More recently, Ruth Schulz lead research where users interact with the PR2 for a joint task, e.g. build a bridge of blocks together. The PR2 with a high-level software interface to script interactive manipulation behaviors became our standard research platform for this. Planning Contact Trajectories: Vien Ngo lead some research on applying rigorous POMDP formulations (his field of expertise) for designing robot motions that are information gaining. E.g., the robot motions touch the environment and slide along tables to reduce uncertainty over table or object coordinates. Compliant Contact Control: The PR2 also helped us enormously to learn about robust real-world combined position, impedance and force control given an incorrect dynamical model (as every analytical model of the PR2 is incorrect). We coded a new 1000Hz onboard real-time controller. Active Learning & Dynamics Learning: In two further publications, both at ICRA‘15, we used the PR2 to collect data of its dynamics, and to test active learning methods.
Publications
-
Reactive phase and task space adaptation for robust motion execution. In Proc. of the int. conf. on intelligent robots and systems (iros 2014), 2014
P. Englert and M. Toussaint
-
Active exploration of joint dependency structures. In Proc. of the ieee int. conf. on robotics and automation (icra 2015), 2015
J. Kulick, S. Otte, and M. Toussaint
-
Inverse kkt — learning cost functions of manipulation tasks from demonstrations. In Proceedings of the international symposium of robotics research (isrr 2015), 2015
P. Englert and M. Toussaint
-
Pomdp manipulation via trajectory optimization. In Proc. of the int. conf. on intelligent robots and systems (iros 2015), 2015
N. A. Vien and M. Toussaint
-
Sparse gaussian process regression for compliant, real-time robot control. In Proc. of the ieee int. conf. on robotics and automation (icra 2015), 2015
J. Schreiter, P. Englert, D. Nguyen-Tuong, and M. Toussaint
-
Touch based pomdp manipulation via sequential submodular optimization. In Proc. of the ieee-ras int. conf. on humanoid robots (humanoids 2015), 2015
N. A. Vien and M. Toussaint
-
Combined optimization and reinforcement learning for manipulations skills. In Proc. of robotics: science and systems (r:ss 2016), 2016
P. Englert and M. Toussaint
-
Policy search in reproducing kernel hilbert space. In Proc. of the int. joint conf. on artificial intelligence (ijcai 2016), 2016
N. A. Vien, P. Englert, and M. Toussaint
-
Building a bridge with a robot: a system for collaborative on-table task execution. In Proc. of the int. conf. on human agent interaction (hai 2017), 2017
R. Schulz, P. Kratzer, and M. Toussaint
-
Constrained bayesian optimization of combined interaction force/task space controllers for manipulations. In Proc. of the ieee int. conf. on robotics and automation (icra 2017), 2017
D. Driess, P. Englert, and M. Toussaint
