Introducing robots into human environments requires them to handle settings designed specifically for human size and morphology, however, large, conventional humanoid robots with stiff, high powered joint actuators pose a significant danger to humans. By contrast, “anthropomimetic” robots mimic both human morphology and internal structure; skeleton, muscles, compliance and high redundancy. Although far safer, their resultant compliant structure presents a formidable challenge to conventional control. Here we review, and seek to address, characteristic control issues of this class of robot, whilst exploiting their biomimetic nature by drawing upon biological motor control research. We derive a novel learning controller for discovering effective reaching actions created through sustained activation of one or more muscle synergies, an approach which draws upon strong, recent evidence from animal and humans studies, but is almost unexplored to date in musculoskeletal robot literature. Since the best synergies for a given robot will be unknown, we derive a deliberately simple reinforcement learning approach intended to allow their emergence, in particular those patterns which aid linearization of control. We also draw upon optimal control theories to encourage the emergence of smoother movement by incorporating signal dependent noise and trial repetition. In addition, we argue the utility of developing a detailed dynamic model of a complete robot and present a stable, physics-based model, of the anthropomimetic ECCERobot, running in real time with 55 muscles and 88 degrees of freedom. Using the model, we find that effective reaching actions can be learned which employ only two sequential motor co-activation patterns, each controlled by just a single common driving signal. Factor analysis shows the emergent muscle co-activations can be reconstructed to significant accuracy using weighted combinations of only 13 common fragments, labelled “candidate synergies”. Using these synergies as drivable units the same controller learns the same task both faster and better, however, other reaching tasks perform less well, proportional to dissimilarity; we therefore propose that modifications enabling emergence of a more generic set of synergies are required. Finally, we propose a continuous controller for the robot, based on model predictive control, incorporating our model as a predictive component for state estimation, delay-compensation and planning, including merging of the robot and sensed environment into a single model. We test the delay compensation mechanism by controlling a second copy of the model acting as a proxy for the real robot, finding that performance is significantly improved if a precise degree of compensation is applied and show how rapidly an un-compensated controller fails as the model accuracy degrades.
| Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:589077 |
| Date | January 2013 |
| Creators | Diamond, Alan |
| Publisher | University of Sussex |
| Source Sets | Ethos UK |
| Detected Language | English |
| Type | Electronic Thesis or Dissertation |
| Source | http://sro.sussex.ac.uk/id/eprint/47191/ |
Page generated in 0.0017 seconds