Workshop at R:SS, Thursday July 16, 2015

How do we build autonomous robots that are able to operate intelligently in a complex world?

Robots operating autonomously in unstructured environments need to acquire an understanding of how their world works and how to act within it. In a complex world that offers rich sensorimotor interaction, learning and interaction must take place on the basis of suitable abstractions. For example, knowing how to operate a specific handle for a specific purpose is of limited general use to the robot; it would be much better off having a generic concept of a handle and being able to recognize and operate various handles of a given type, including previously-unseen instances. Thus, perceptual and motor generalization are key features of reusable concepts.

Reusability is paramount if robots are expected to learn nontrivial concepts. Learning is only feasible to the extent that the learning problems are sufficiently pre-structured and constrained. For example, learning a concept such as setting a table offers too many degrees of freedom to be feasible. However, it can plausibly be learned on the basis of useful abstractions such as stackable, container, or to open, each of which can be learned in simpler ways.

This workshop discusses methods for the design and acquisition of reusable concepts such as generic sensorimotor skills that can be applied in diverse circumstances, as well as skills or abstract concepts in terms of which more elaborate skills or concepts can be expressed. Of particular interest are methods that allow robots to increase their knowledge and skills by learning new concepts with the aid of already-learned concepts. Thus, concepts are reusable in the sense that they can facilitate the learning of more advanced concepts, e.g. by constituting building blocks or guiding the learning process.

Issues of interest include but are not limited to:

Learning complex concepts in terms of simpler concepts already learned
Learning structure to guide future learning
Autonomous structuring of learning problems into subproblems
Representations and algorithms for composable concepts
Generalization based on functional parts, material properties, etc.
Skill generalization and transfer

Format

This workshop will feature a program composed of invited speakers. We have an exciting lineup of speakers, including senior authorities and rising stars, all with proven expertise in this area, and with strong and thought-provoking messages for us!

Confirmed Keynote Speakers

Frank Guerin, University of Aberdeen
George Konidaris, Duke University, Durham, NC
Ben Kuipers, University of Michigan, Ann Arbor, MI
Marc Toussaint, University of Stuttgart
Emre Ugur, University of Innsbruck
Paul Verschure, Universitat Pompeu Fabra, Barcelona
Gregor Schöner, Ruhr-Universität Bochum

Discussions

A key goal of our workshop is to ignite discussions among the participants (including the speakers), allowing the combined expertise and diversity of viewpoints present in the room to unfold, as opposed to conventional, one-way Q&A where the audience asks questions to be answered by the speaker.

Each keynote speaker will raise, during her or his talk, at least one major open research question, which will be brought up for discussion after the talk.
Instead of concluding their talks in the conventional way, speakers will lead directly into the discussion by asking provocative questions for the audience and/or by summarizing key points in controversial statements.
The workshop will conclude with a panel discussion with strong audience participation, exploring the core question of how to build autonomous robots that are able to operate intelligently in a complex world in the light of learning reusable concepts.

Posters

As an additional basis for discussion and to foster interaction on the initiative of participants, we will accept a limited number of contributed posters on a first-come/first-serve basis, subject to informal screening. (Submission details will appear here in due time.)

Organizers