Kasper Stoy on self-reconfigurable robots

Luke Muehlhauser: In Larsen et al. (2013), you and your co-authors write:

Using a bottom-up, model-free approach when building robots [is] often seen as a less scientific way, compared to a top-down model-based approach, because the results are not easily generalizable to other systems… In this paper we will show how the use of well-known experimental methods from bio-mechanics are used to measure and locate weaknesses in our bottom-up, model-free implementation of a quadruped walker and come up with a better solution.

From looking at the paper I could see how your experimental method allowed you to find a better solution for your walker robot, but I couldn’t understand how you addressed the challenge of generalizing the solution to other systems despite the bottom-up, model-free approach. Could you explain that part?

Kasper Stoy: The problem is that researchers who are doing cutting edge robotics want to explore how materials and their interaction can aid the movement of the robot. For instance, researchers have been working on passive-dynamic walkers for a long time now that exploit the mechanical system to achieve walking without using sensors or actuators. The energy comes from walking downhill and the control is open-looped – the mechanical system itself is self-stabilising. For these systems our current engineering approach is ok, but not great. We can just about model these kinds of walkers so we can get a good guess of the initial parameters to get the system walking, but there is still an extended phase of tinkering before the system actually walks. It is in this tinkering phase that our precise motion capture as used in biomechanics comes into play. Given measured paths of all parts of the robot we can analyse the data and come up with reasonable hypotheses about how to improve the robot. Hence, the tinkering becomes much more systematic even though the underlying physical processes are too complex to be modelled. It may not be apparent, but just modelling the impact of a foot with the ground which takes all types of frictions into account, the deformation of the foot, the spring effect, and so on is largely intractable. Hence, the underlying assumption of our work is that all models of locomotion are fundamentally wrong. They may give a high-level picture of what is going on, but fundamentally they cannot be used to predict what will happen just two steps later. However, if we turn this around and we have gotten our robot to walk and we record the data of how it walks. Although difficult, we can build a model that match this data which is based on the ground truth and where there is no modelling bias on the part of the researcher. We now have a better informed model that can be used to build the next generation of robots. Hence, the model is a generalisation of our specific implementation which you can copy, but you can also replace elements of which you have better implementations. In locomotion research like many other fields the models and physical systems are drifting apart because it is researchers with different skill and interests who work on them. In our work, we are clearly working on physical systems and just provide a hint to how our experimental results can be generalised in a way that is meaningful to modelling oriented researchers. We hope.

Luke: You co-authored a book on self-reconfigurable robots in 2010. What do you mean by that term, and what are some recent examples of practical self-reconfigurable robots? (Perhaps, more recent than the book?)

Kasper: Self-reconfigurable robots are an idea inspired by multi-cellular organisms. The idea is that instead of building a robot as a single expensive, monolithic, and fragile piece of hardware you would built a robot from many relative simple robotic cells. We refer to these robotic cells as modules and the research community as modular robotics. Self-reconfigurable robots then take this a step further in that not only is the robot multi-cellular, the modules can also rearrange the way they are connected automatically. Hence as the modules wander around each other the robot as a whole change shape. While this is rare in biology there is a small aquatic animal called hydra which has this ability. In fact, it has been shown that if you cut it in half it transform itself into two smaller replicas of the original hydra. The advantage of this type of robot is extreme robustness and versatility. Features that are particular important in for instance extra-terrestrial exploration. Regarding, application this was from the beginning curiosity driving research. Is it possible to build a robot that is multi-cellular and able to chance its own shape? The answer is yes. The concept has been implemented successfully in about five systems internationally. However, modules are relatively expensive, big (10cm diameter), complex, and quite heavy. Hence, direct applications are still lacking given that the cost-benefit of these robots is not good enough for a commercial market. However, the more modest modular robots have seen some recent success on KickStarter with the Modbots. A modular construction kit for building your own robots.

Luke: What are each of the “about five systems” you refer to?

Kasper: The most developed system is the M-TRAN III self-reconfigurable robot developed at AIST in Japan. The runner ups are ATRON which was developed mostly by my former colleagues at University of Southern Denmark. However, both of the above systems are not being developed further. Another runner up is the Roombots developed at EPFL in Switzerland. Recent exciting development are the M-Blocks coming of of MIT and SMORES coming ot of UPenn. There is a large number of robots which come very close, but not quite achieving convincing three dimensional self-reconfiguration.

Also, check out this Wikipedia article for a more comprehensive list.

Luke: What are some specific breakthroughs or kinds of progress that you expect to see in the next 10-15 years of research into self-reconfiguring modular robots?

Kasper: The field has very much been about whether it is possible to build robots that can change their own shape or not. Within the last decade, we have established that this is indeed the case. However, currently the robotic cells are too complex, heavy, and expensive for most applications. Hence, one of the goals is to reduce all of these obstacles, but given the current state of the art in mechatronics this is quite difficult. Hence, I see a mainstream area of research focussing on various simplifications of the mechatronics to make it attractive for specific applications in space, robotic prototyping tools and the like. However, it also clear that other implementations have to be considered. For this some have turned to DNA-based implementations or using micro fabrication technology. Another unexplored opportunity is in soft robotics that is gain traction in the robotic community as a whole. The question is then where this would lead in 10-15 years. I think all areas have potential so if for instance the soft modular robotics effort is successful this could lead to robotic multi-cellular systems much closer to biological cells in terms of material properties. Where would this find applications is anyones guess, but maybe in furniture, clothing or in healthcare in the form of adaptive casts.

Luke: Thanks, Kasper!