Using genetic programming to evolve vision-based controller for PPRK robot

My M.Sc. student, Pawel Gajda, developed under my supervision an environment for evolving vision-based controllers for simple PPRK robot.

System components

• The PPRK robot:
  
  See detailed description of PPRK robot.
• The CMUCAM2+ camera:
  
• Software that evolves the contollers encoded as Genetic Programming individuals.
  • See my paper describing the approach on image interpretation using genetic programs that process visual primitives.
  • The software is based on the Visfast library by Wojciech Jaskowski and Bartek Wieloch, and uses ECJ (Evolutionary Computation in Java).

Experiment results

Two frames from the training video sequence and their representation in terms of visual primitives:

A typical individual evolved using the approach:

See the movie showing the robot navigating towards the target (WMV format).

The complete text of Pawel Gajda's thesis describing the system (in Polish).