The New Body Problem

by JS

David Vernon’s opening keynote at this year’s ICDL included a lot of information about the iCub, a serious effort to create a standard platform for developmental robotics. The talk was in part a review of a road map for developmental robotics that iCub researchers have been working on. They have a compelling story, one that I find very interesting.

After the talk there was a brief Q&A session, and one of the audience members asked an interesting question. If developmental robotics is so great, how do we demonstrate it for the vast majority of still skeptical roboticists? I didn’t like the answer Vernon gave, so I thought I’d provide my own. First, I don’t believe that in terms of what roboticists are trying to do at the present moment, developmental approaches are the way to go. To put this another way, if you want to build an autonomous tank, you are better off applying careful engineering than trying to structure some kind of developmental process that starts with a “baby” tank that grows into an “adult” tank. Second, as more and more problems get moved into the “solved” category through some combination of careful engineering and learning, the remaining questions will only get harder, and eventually the developmental approach will surpass careful engineering.

One possible future includes robots with innumerable differences in configurations (for example, consider the individual variations in humans). This differs from the view that robots will be constrained to certain models like automobiles. In one case, engineering each model would be feasible, in another, the development would rapidly eclipse the benefits of careful engineering.

These possible futures point to a problem we can consider in the present to demonstrate the necessity of the developmental approach, much in the same way the ‘kidnapped robot problem‘ crystallized research into autonomous mapping and navigation. For lack of a better name, I call this the “new body problem.” The basic problem statement is, starting with an uninterpreted sensory and motor interface, learn to complete simple tasks as instructed by a human.

To be concrete, consider that an uninterpreted sensor interface is just a time sequence of vectors (s_0, s_1, \ldots, s_n)_t where each s_i \in \mathbb{R}. A motor interface is a similarly general sequence of vectors (m_0, m_1, \ldots, m_k)_t where each motor variable can be set by the agent. Solving this problem as stated would require development, but we can certainly imagine modifications to the problem that would allow the agent to wake up with slightly more semantic sensor and motor knowledge that would still require substantial development before being able to take instructions and perform simple tasks. The point is that by framing the problem this way we are explicitly identifying dual goals, to increase the competencies of the learned robot at the end of development, and to decrease the necessary prior knowledge needed at the beginning of development. Developmental robotics considers both goals.