Learning to Control a Biomechanical Model of a Human Arm

One way to understand how an animal learns to control its motor system is by hypothesizing a model and comparing its behavior to that of the animal. Such a model will consist of three components: a detailed model of the animal's musculoskeletal system, a dynamical simulator to model the physics of motion, and a control architecture. Recent work in biomechanics has provided a precise musculoskeletal model of the human upper limb (see Figure 1). It is composed of an integrated skeletal [1] and muscle [2] model derived from CT tomography data of a human male cadaver; the model consists of seven bones with 13 degrees of freedom and 42 muscle bundles representing 26 muscle groups. My work involves embedding this model in a dynamical simulator and designing control architectures to achieve complex behaviors such tossing a ball or throwing darts.

The complexity of the musculoskeletal model poses significant challenges to traditional learning architectures. In particular, the configuration space of the arm is a 13-dimensional continuous space, and the action space is a 42-dimensional continuous space. Moreover, biomechanical constraints significantly constrain the space of reasonable actions. For example, abduction of the upper arm is achieved (in humans) by four muscle groups whose effects are partially, but not completely, redundant; it makes little sense for a controller to use only one of these muscles to effect an abduction of the upper arm. Techniques that may prove useful in designing controllers include reinforcement learning, hierarchical models (in which complex controllers are built by composing simpler controllers), and continuous function approximators such as neural networks and support-vector machines.

Figure 1: A visualization of the Garner and Pandy (2000) upper limb model. (Image courtesy of the Visualization Lab, University of Texas at Austin.)

[1]

B. Garner and M. Pandy, "A Kinematic Model of the Upper Limb Based on the Visible Human Project (VHP) Image Dataset," Computer Models in Biomechanics and Biomedical Engineering, Vol. 2, 1999.

[2]

B. Garner and M. Pandy, "Musculoskeletal Model of the Upper Limb Based on the Visible Human Male Dataset," Computer Models in Biomechanics and Biomedical Engineering (to appear).