Anticipation and Attention in Robot Control

| Published | English
Johansson, Birger
Cognitive Science
Functional zoology
Lund University Cognitive Science (LUCS)-lup-obsolete
Ikaros: An infrastructure for system level modelling of the brain
Research Group:
Lund University Cognitive Science (LUCS)-lup-obsolete
Anticipatory and predictive models are becoming very important features of robot systems. This thesis investigates some aspects of predictive modeling. Is prediction always a good thing? How important is it to anticipate what will happen in the future? Is it better to anticipate far into the future or to focus on the next few seconds? What are the requirements for predictive models?

Predictive models are hard to analyze. In daily life there could be an infinitive number of future states and to be able to analyze the predictive models, both the task and the environment must be simplified. This thesis uses small robots in simple navigation tasks. A typical task could be to predict what another robots will do and act according to this prediction.

The first two papers describe a model for predictive eye movements in infants. Results presented in paper 3 indicate that a multi robot system will benefit from anticipation compared to a system without anticipation. The implementation of the system is described in papers 4-6. Experiments with a more advanced multi agent system are reported in paper 7. These experiments studied how far ahead agents must predict events in its environment to benefit from the prediction. The result indicate that the prediction time is heavily correlated with the task. An optimal prediction time will change depending on the task and current state. To not take this into account can reduce a system's performance even with longer prediction time.

The system developed for these experiments is called AARC (Anticipation and Attention in Robot Control) and is described in detail in paper 8. When running the whole system in full scale, 17 computers are communicating between each other and steering 6 robots. The system can run both in real robot mode or as a pure simulation. Two different systems have been developed in this thesis. The first used a more traditional AI planning strategy and the second used a mix of traditional AI and predictive models (Linear associaters and Kalman filtering). The thesis presents these systems, together with a paper describing the Ikaros framework used to build AARC and the color tracking used for robot tracking.

Further experiments are described in paper 9 where the AARC architecture is used to anticipate how a number of moving obstacles will behave to allow a simulated robot to select an appropriate strategy to negotiate the obstacles. Large scale simulations are used to test different strategies and parameters.
eye-movements ; visual field ; tracking ; occlusion ; cognitive modelling ; boebot ; e-puck ; color tracking ; multi-agent ; epistemic action ; architecture ; agent ; robot ; anticipation ; attention ; prediction
ed606683-08e4-46bb-9ea7-0f219fb79466 | Link: | Statistics

Cite this