Navigation methods using multi-sensor systems are considered for terrains whose geometric models are known only in certain parts. Exact navigation algorithms, as well as those that obtain approximate paths with lesser computation time while compromising the optimality are studied. Navigational methods for terrains, where polygonal approximations are inefficient, are explored based on fractal interpolation and approximate path planning. The navigation process involves the problem of extracting information about the geometry of the terrain by suitably combining the sensor outputs. Techniques for (a) extracting integrated information, and (b) fusing the new information into the existing terrain model, are investigated. Two new paradigms for sensing: (a) active sensing where sensor operations are repeated using sequential statistical methods, and (b) learning fusion rules from examples where the system is trained with examples using the methods of empirical and structural risk minimization, are proposed. These two methods, together with the existing Bayesian-based methods, provide the design paradigms for the multi-sensor system. For computational purposes, a general distributed sensor system specified by a set of first order propositional sentences is proposed. This system encompasses, as special cases, the distributed versions of Bayesian inference, and geometric reasoning with logic and algebra. In the case of sensors whose outputs are real-valued vectors, the entire process of sensor integration is shown to be implementable in hardware using combinational circuits and comparators. This implementation is particularly suitable for real-time applications. Also implementations on several other computational systems are studied.
