This half-day workshop will take place on Friday, May 20 (08:00-12:30 in Room 27) at ICRA 2016 – the IEEE International Conference on Robotics and Automation – in Stockholm, Sweden.
The nature versus nurture debate concerns how much an individual is defined by innate properties versus how much it is shaped by experience. Concerning humans, there is a consensus in psychology that we are the product of the interaction of both, nature and nurture. We believe that the same must be true for intelligent robots, where nature corresponds to engineered solutions and nurture corresponds to machine learning. Therefore, also in our field, it is essential to study the interaction of nature and nurture—or rather engineering and learning: How can we decide which parts of an intelligent robot should be engineered by humans and which should be left unspecified to be learned by the robot?
We think that this is an essential question that needs to be discussed to identify promising research directions towards creating intelligent robots. This matter is especially pressing right now, as machine learning approaches become more common in robotics and it is unclear how to best combine them with engineering approaches. With the recent advances in deep learning, other fields like computer vision seem to switch from an engineering paradigm to a learning-only paradigm. It stands to debate if this approach is promising.
This workshop consists of invited talks, poster presentations, and a panel discussion on the following questions:
1) Does robotics require only nature (i.e. engineering), only nurture (i.e. machine learning), or both?
Is it possible to create an intelligent robot whose response to the environment is completely predefined by a designer (pure nature view)? Is it possible to create an intelligent robot whose response to the environment is entirely learned by the robot without being provided any information about the environment by the designer (pure nurture view)? If neither is true, what is the right balance between these two extremes? In particular, which parts of the robot should be predefined and which should be learned from experience? What is the role of engineered solutions? What is the role of solutions obtained by the robot through machine learning? What is the best way to combine them? How does this depend on embodiment?
2) How does this balance depend on the amount of data and computation?
If we had infinite computing resources, would we still rely on engineered solutions as parts of the robot behavior? What if we also had an infinite amount of data? How much computation and data do we expect an intelligent robot to have access to? Given this restriction, what is the best balance between engineering and machine learning? How should we split up intelligent robot behavior into engineered and learned parts? Where does deep learning fit into this picture?
3) How does this balance depend on the problems which we are trying to solve?
What are the problems we are trying to solve in robotics? What is the right order to approach them? How can we exploit problem structure? What characteristics do robotics problems have? How do these characteristics relate to engineering and machine learning approaches? How does this balance depend on the complexity of the problem space? Which role does uncertainty play? Would we need machine learning if we had perfect understanding of the problems?
We were very surprised by the many enthusiastic responses of the speakers we had invited. We think that the quality of our confirmed invited speakers shows the demand to discuss “nature versus nurture in robotics”.
See you in Stockholm!
Rico Jonschkowski & Oliver Brock