| Dynamical Systems for Motion Planning Relates to SENSOPAC Work Package 2: 2A.1, 2A.3, and Work Package 4: 4A people involved: Sebastian Bitzer, Sethu Vijayakumar Role within SENSOPAC
Work on natural (human) movement
statistics from WP1 will act as a seed for defining basic exploratory
movements. We will find dynamical systems representations for
these human motion patterns for
use in the simulation and the arm-hand robot system.
The dynamical systems formulation has the advantage that individual aspects of the motion plans such as speed and amplitude can be altered while the main characteristics of the movements stay in tact. This allows for easy adaptation of motion plans to changing context. DescriptionMovements are inherently dynamical. It therefore seems natural to represent them with dynamical systems. Ijspeert et al. (2003) describe an approach in which a trajectory is represented as an attractor of a nonlinear dynamical system. Such a representation of a motion plan has several advantages. For example, the plan becomes robust against small perturbations and it can be easily modulated in time and space while the main characteristics of the plan stay in tact. Exploratory movements have to be adapted to changes in situation/context in order to use sensors optimally, i.e. in their optimal working range. For example, if the weight of a held object needs to be precisely determined with force sensitive sensors, then too slow movements might not elicit a sufficient sensor response while too fast movements would saturate the sensors. With dynamical systems representations of prototypical exploratory movements we can adapt these movements to changing contexts by, for example, changing their speed or amplitude without loosing the characteristic shape of the movements that was originally observed in humans. Additionally we are working together with subproject Dimensionality Reduction to support learning of the dynamical systems by providing a compact representation of motion data prior to learning. For that purpose we are especially investigating compatibility between dimensionality reduction and dynamical systems representation, i.e. we are investigating the question whether modulation of a motion plan in the lower dimensional space leads to a corresponding modulation in the original space. ReferencesIjspeert, A. J., Nakanishi, J. and Schaal, S., Learning Attractor Landscapes for Learning Motor Primitives, Advances in Neural Information Processing Systems, NIPS, 15, MIT Press, pp. 1523-1530, 2003 |
Subproject overview |