The Physical Exploration Challenge: Robots Learning to Discover, Actuate, and Explore Degrees of Freedom in the World
Final Report Abstract
Actively seeking for information, exploring the environment and thereby acquiring a model of the environment is a crucial aspect of intelligent behavior. Such behavior is also described in terms of curiosity, or the intrinsic motivation to learn about the environment. The goal of this project was to develop methods to realize such behavior concretely in the context of a physical world, where a robot needs to physically explore and interact with its environment so as to uncover its physical and kinematic structure. In the course of this project we developed new approaches to represent the belief state of the agent, which represents the current state of knowledge and uncertainty about the structure and the state of the world. Based on this we developed efficient strategies to explore the environment to increase and refine this belief state. A focus of our work was on how and what to represent probabilistically to enable information gain metrics that drive exploratory robot behavior to discover DoFs. Further, to support the exploration and acquisition of information of the DoFs in the environment, we developed novel perceptual skills for the robot. In particular, we extended methods for online perception of articulated objects from interactions to segment and reconstruct the shape of the links of the articulated objects. We developed methods to plan robust actions that generate rich interactive information to reduce the uncertainty about the structure of the environment exploiting the probabilistic representation over kinematic structures. Finally, we contributed novel methods to acquire manipulation skills to explore and operate DoFs. Our research focused on sample-efficient Reinforcement Learning approach that require a low amount of human supervision and real-robot interaction time but generalize to a large range of different environments.
Publications
-
Entropy based strategies for physical exploration of the environment’s degrees of freedom. In Proc. of the Int. Conf. on Intelligent Robots and Systems (IROS 2014), September 2014
Stefan Otte, Johannes Kulick, Marc Toussaint, and Oliver Brock
-
Active exploration of joint dependency structures. In Proc. of the IEEE Int. Conf. on Robotics and Automation (ICRA 2015), 2015
Johannes Kulick, Stefan Otte, and Marc Toussaint
-
An integrated approach to visual perception of articulated objects. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 5091 – 5097, 05 2016
Roberto Martín-Martín, Sebastian Höfer, and Oliver Brock
-
Achieving robustness by optimizing failure behavior. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), pages 5806–5811, 2017
Manuel Baum and Oliver Brock
-
Building kinematic and dynamic models of articulated objects with multi-modal interactive perception. In AAAI Symposium on Interactive Multi-Sensory Object Perception for Embodied Agents, 2017
Roberto Martín-Martín and Oliver Brock
-
Inverse KKT: Learning cost functions of manipulation tasks from demonstrations. The International Journal of Robotics Research, 36(13-14):1474–1488, 2017
Peter Englert, Ngo Anh Vien, and Marc Toussaint
-
Opening a lockbox through physical exploration. In Proceedings of the IEEE International Conference on Humanoid Robots (Humanoids), 2017
Manuel Baum, Matthew Bernstein, Roberto Martín-Martín, Sebastian Höfer, Johannes Kulick, Marc Toussaint, Alex Kacelnik, and Oliver Brock
-
Learning manipulation skills from a single demonstration. The International Journal of Robotics Research, 37(1):137–154, 2018
Peter Englert and Marc Toussaint
-
Physics-based selection of informative actions for interactive perception. In IEEE International Conference on Robotics and Automation (ICRA), pages 7427–7432, 2018
Clemens Eppner, Roberto Roberto Martín-Martín, and Oliver Brock
