Abstract:
Various solutions to the Simultaneous Localisation and Mapping (SLAM) problem have been proposed
over the last 20 years. In particular, extending the fundamental solution of the SLAM problem has
attracted a great deal of attention. Most extensions address shortcomings such as data association,
computational complexity and improving predictions of a vehicle’s state. However, nearly all SLAM
implementations still depend on analytical models to provide estimates for state transitions.
Learning data-derived non-analytical models for use during localisation and mapping provides an
alternative that could significantly improve estimates and increase the flexibility of models. A methodology
to learn motion models without knowledge of the higher-order dynamics is therefore proposed
using tapped delay-line neural networks (TDL-NN). Incorporating the learned Nth-order Markov
model into a recursive Bayesian estimator requires that the learned model be assumed independent of
the transitional model, forming a black box estimator. Both real-world and simulated training data
were evaluated, along with changes to the input data’s format, to determine the best vehicle motion
predictor. Furthermore, an evaluation methodology is defined to asses how well the models could learn each
motion type. A comprehensive analysis of the one-forward prediction using various statistical measures
was used to determine the most appropriate metric. The methodology evaluated the predictions at
different levels of depth, providing supplementary information on the type of motions that are learnable.
Outcomes of the experiments revealed that inherently learning a vehicle’s dynamics cannot be achieved
using TDL-NNs. Currently the best that such an approach can learn is the delta between the vehicle’s
states. Consequently, modifications are required to the learning algorithms as well as the input data’s
format that will force the strategies to learn the higher-order dynamics.
