Multi-objective evolution of RF-signal homing behavior in simulated autonomous wheeled robots using differential evolution
Although there are more than 1 million robots occupying the world today in the automotive and materials handling industries, a large majority of these robots are fixed robots which are equipped with hand-engineered, pre-programmed routines to function within a static, predictable environment. Only a...
No Tags, Be the first to tag this record!
|Summary:||Although there are more than 1 million robots occupying the world today in the automotive and materials handling industries, a large majority of these robots are fixed robots which are equipped with hand-engineered, pre-programmed routines to function within a static, predictable environment. Only a small fraction (0.15%) of this total comprises of autonomous mobile robots that have artificial intelligence and which can be adaptive to changing, dynamic environments. This is mainly due to the difficult task of synthesizing effective yet robust controllers for autonomous
mobile robots. As such, evolutionary robotics (ER) has been introduced as a new methodology to overcome these limitations by applying artificial evolutionary optimization algorithms for the automatic generation of robotic controllers. Over the last decade, a number of successful studies have been reported in the application of ER. However until very recently, only single-objective evolutionary algorithms have been utilized in ER. In the few investigations that have utilized evolutionary multi-objective algorithms (EMO), the studies have only been conducted on highly abstract, legged robots. Hence, the motivation for this thesis is three-fold; firstly to
investigate whether EMO can be successfully applied to ER on simulated but actual, real-world physical wheeled robots, secondly whether EMO can be applied to ER for generating radio-frequency (RF) localization behaviors, and lastly whether EMO can be applied to ER for generating useful behaviors in multiple robots working as a collectively-intelligent group. The experiments are implemented to focus on five main research objectives: (1) to obtain a fitness function for generating the wheeled robot's RF-localization behavior in an inherently noisy environment; (2) to
evaluate the EMO's performance in evolving the required robot's controllers to solve the task environment; (3) to test the evolved controllers' robustness; (4) to verify
the EMO's ability to generate useful controllers in a collective task; and (5) to analyze the evolved controllers' internal processing structure in terms of Hinton graphs. The results showed that: (1) a fitness function was successfully generated for the wheeled robot's RF-localization behavior; (2) the EMO performed reliably in synthesizing the required controllers for solving the task environment; (3) the
evolved robot controllers were robust to the different, previously unseen testing environments that were different from the evolution environment; (4) the EMO was able to evolve controllers for solving a collective box-pushing task for multiple robots; and (5) based on the Hinton graph analysis, there were noticeably strong excitatory as well as inhibitory synapses present in the most optimal evolved
controllers that produced the desired robot behaviors. Therefore in conclusion, this thesis has shown that EMO is a useful and promising technique to employ in ER for automatically generating robust RF-localization behaviors in simulated autonomous wheeled robots as well as for collective behaviors in multiple robot environments.|