SNN-Based Learning of Internal Dynamic Models for Reaching Behaviour in Robot Arms
No Thumbnail Available
When performing reaching tasks, the brain relies on noisy and delayed sensory feed- back, but is still able to perform precise on-line movement corrections. Research in motor neuroscience suggests that this ability is due to internal forward models that make predictions about future states of the body, based on the current state and the outgoing motor signal. These predictions are compared to the delayed sensory feed- back such that a prediction error can be established, and the forward model can be kept unbiased. The forward predictions are used as a substitute for sensory feedback, allowing for fast and precise motor corrections. However, to our knowledge, no bio- logically plausible models of reaching behaviour based on forward modelling exist yet. In this thesis, we present a spiking neural network architecture for learning reaching behaviour in a robot arm, inspired by the forward modelling theory. The learning process does not require a predefined training set and only relies on an implicit motor babbling phase. We demonstrate that the inclusion of a forward model significantly improves the performance and learning capabilities of a robot arm.
Faculteit der Sociale Wetenschappen