Thomas Bührmann - Research Student
|
Phone Location Webpage Supervisor |
++44 (1273) 87-2949 JMS 3D8 T.Buehrmann@sussex.ac.uk http://www.funkalism.com/ Ezequiel DiPaolo |
| Search the CCNR bibliography for papers by Thomas Buehrmann | ||
Research Interests
The fundamental question in my thesis concerns the organization of movement in animals and robots. I believe that this problem should be approached by realizing that movement is the outcome of coordination between a number of coupled dynamical systems (skeleton, muscles, reflex circuits, central commands) that changes depending on context. I see this as being in contrast with a more static approach inspired by engineering that treats the problem mainly as one of transformations from visual coordinates to joint coordinates.
In practical terms, the aim of my thesis is to build a robot control system based on system-level analysis of biological movement control. By identifying functional and network properties of the different levels in the hierarchical organization of movement generation I try to create a neural network model using techniques from evolutionary robotics that qualitatively captures observed biophysical phenomena and offers insight into concepts underlying coordination.
The main motivation for using a hierarchical approach is that each level can exploit the possibly complex and non-linear dynamics of lower levels (such as a muscle's force-length and force-velocity relationships). I will argue that many of the invariants and phenomenological properties of natural movements (e.g. bell-shaped velocity profiles or roughly straight lines in cartesian space) can possibly be understood from the interactions between a neurocontroller and a complex musculoskeletal system. Rather than being explicitly calculated features of a movement trajectory, they could be "emergent" properties of the coupled dynamical system.
Non-linear relations between a muscle's force, length, velocity and activation for example play crucial roles in allowing quick movements with precise breaking and implementation of so-called pre-flexes (zero lag resistance to perturbations). It has also been shown that these properties simplify the control signals necessary to move a limb quickly without overshooting. Similarly, spinal circuits implement low-level dynamics which could be exploited by higher level sytems. Being the interfaces for integration of afferent sensory feedback, reflexes, spinal pattern generators and control signals originating in higher levels, they provide a means e.g. to selectively activate and deactivate parallel control pathways, thus effectively reducing complexity (similar to Bernstein's idea of potentiation/depotentiation to account for the degrees-of-freedom problem).
In short, taking into account the dynamical properties of the musculoskeletal system and spinal circuitry, voluntary movement can then be understood as "the selection of subsets of spinal interneurons whose excitability is regulated in part by afferent feedback throughout the movement" (McCrea, Behavioral Brain Sci. 15). In my research I try to take these ideas seriously and use robotics as a testbed to test their explanatory power and real-world applicability. To this end I'm currently co-evolving muscle designs and neural reflex circuits that when coupled to a physically simulated robot arm create movement that exhibits certain desirable properties (e.g. independent control of position and stiffness). I'm also about to experiment with learning architectures (specifically liquid/echo state machines) replacing or working in parallel with evolutionary algorithms.