A Neural Network Based Strategy for Robot Navigation in Dynamic Environments

Meng, Hua (1993). A Neural Network Based Strategy for Robot Navigation in Dynamic Environments. PhD thesis The Open University.

DOI: https://doi.org/10.21954/ou.ro.0000ff15


This thesis studies the problem of robot navigation in the presence of unexpected environmental changes, which include unknown static obstacles and moving objects with unknown trajectories. Throughout this work, neural networks, as a new technique, are used to develop the functional components, which constitute the proposed navigation strategy. The neural network based navigation strategy we propose follows a two-level hierarchy and operates by integrating three network components (planner, navigator and predictor). At the higher level, the planner generates a nominal path from the initial position to the goal among the fixed known obstacles. At the lower level, the navigator incorporates the predictor to refine the coarse path by taking into account unexpected environment changes to achieve on line real-time guidance.

During this research, three neural network components were developed. The path planner was developed first by using a three-layer feedforward network to optimize the cost (collision penalty) function of a path. The first version of the navigator - Navigator-1 - was then implemented using a multilayer feedforward network in which steering commands for static obstacle avoidance were generated by directly converting sensor reading through the network. To enable the navigator to handle moving objects with unknown trajectories, on-line motion prediction was introduced. The predictor was developed using an Elman recurrent net. Following that, an enhanced version of the Navigator-1 - Navigator-2 - was developed using a structured network in which three sub-nets were used - two of the sub-nets were used to realise dynamic obstacle avoidance and static obstacle avoidance respectively, and the third sub-net was used to make final steering decision by reconciling the results from those two sub-nets. Finally, the overall navigation strategy was implemented in a simulation system. Simulations showed encouraging results. It demonstrates that the neural network based strategy is capable of achieving adaptive navigation in the presence of unexpected environmental changes.

Viewing alternatives

Download history


Public Attention

Altmetrics from Altmetric

Number of Citations

Citations from Dimensions

Item Actions