Technical reports

Symbol grounding:
Learning categorical and sensorimotor predictions for coordination in autonomous robots

Karl F. MacDorman

May 1997, 170 pages

This technical report is based on a dissertation submitted March 1997 by the author for the degree of Doctor of Philosophy to the University of Cambridge, Wolfson College.

Abstract

To act intelligently, agents must be able to adapt to changing behavioural possibilities. This dissertation proposes a model that enables them to do this. An agent learns sensorimotor predictions from spatiotemporal correlations in sensory projections, motor signals, and physiological variables. Currently elicited predictions constitute its model of the world.

Agents learn predictions for mapping between different sensory modalities. In one example a robot records sensory projections as points in a multidimensional space. It coordinates hand-eye movements by using closest-point approximations to map between vision and proprioception. Thus, one modality elicits predictions more closely identifiable with another. In a different example, an agent generalizes about a car’s sensorimotor relations by weighting sensorimotor variables according to their mutual influence: it learns to navigate without any a priori model of the car’s dynamics.

With feedback from miscategorization, an agent can develop links between categorical representations and the relevant objects they distinguish. Wavelet analysis provides a neurologically plausible means of accentuating invariance that can subserve categorization. In some experiments, categorical representations, derived from inter-category invariance after wavelet analysis, proved to be efficient and accurate at distinguishing different species of mushrooms.

In a simulation of fish chemoreception, agents learn sensorimotor predictions that uncover salient invariance in their environment. Predictions are formed by quantizing a sensory subspace after each dimension has been weighted according to its impact on physiological variables. As these predictions also map from motor signals to likely changes in sensory projections, the agent can chain backwards from desired outcomes to form plans for their attainment.

Full text

PDF (16.7 MB)

BibTeX record

@TechReport{UCAM-CL-TR-423,
  author =	 {MacDorman, Karl F.},
  title = 	 {{Symbol grounding: Learning categorical and sensorimotor
         	   predictions for coordination in autonomous robots}},
  year = 	 1997,
  month = 	 may,
  url = 	 {https://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-423.pdf},
  institution =  {University of Cambridge, Computer Laboratory},
  doi = 	 {10.48456/tr-423},
  number = 	 {UCAM-CL-TR-423}
}