Inhalt des Dokuments
Heuristic vs. model-based approaches for ball catching
Motivation Do heuristics provide better solutions than complex models? There is an ongoing debate in the fields of artificial intelligence, robotics and cognitive sciences how to solve problems of high uncertainty and of high complexity. In this thesis we want to compare different approaches for the problem of catching a ball flying high in the air, and compare them quantitatively and qualitatively.
Description of Work
In this project we want to study the relationship between model-based and heuristic approaches for the well-known outfielder problem. The problem is easily described as follows: an agent has to catch a ball that is flying through the air on a long and possibly perturbed trajectory. Although the problem formulation refers to the domain of sports, the outfielder problem is prototypical for many visual servoing tasks in robotics. Hence, understanding which strategies perform best under which conditions is of great practical relevance for deepening the understanding of intelligent behavior.
All experiments are conducted in a realistic physics simulation environment. In the experiments the agent has to catch a ball which is thrown in the air. A catch is considered successful if the agent is close to the ball when the ball hits the ground. The agent can use different strategies to catch the ball. These strategies should be compared with respect to their performance.
The strategies that should be implemented are:
- Heuristic strategy based on the viewing angle 
- Trajectory prediction strategy (e.g. based on parabola fitting)
- Explicitly modeled strategy using stochastic optimal control
In addition to the strategies different perturbations and disturbances should be considered. The perturbations either affect the trajectory of the ball, e. g. wind or spin, or they affect the sensorimotor apparatus of the agent, e. g. we add measurement noise.
Depending on the outcome of the simulation experiments, the development of a robotic demonstrator could be considered.
 Fink P., Foo P., Warren W. (2009) Catching fly balls in virtual reality: a critical test of the outfielder problem
 Chapman (1968) Catching a baseball