These rules can be used to filter out worthless observations

What is interesting here is not the specific fut 15 coins feature language and rules. They are different for different robots, environments, and cost functions. What is important is that we can try to learn predictive models as to whether or not an observation can be expected to improve the state estimate. These rules can be used to filter out worthless observations.

In our view, research in state estimation is focusing too much on algorithm design and analysis and too little on system design. As we will apply state estimation in very com- plex autonomous robot applications, such as household robotics, where the estimated states are extremely high dimensional we won’t be successful unless we know how to parameterize the systems and how to provide them with the necessary probabilistic models. Therefore, our most important conclusion is an obvious one: the development of complex high-performance state estimation systems is a complex design problem with many design options. The choice of estimation algorithms is only one of these options but many other design dimensions have an equally large impact on system per- formance.


The design has to be tailored to the particular application at hand and the different design option interact with each other in obscure and opaque ways. We pro- pose empirical investigations and learning based on ground truth data as necessary tools for the development of successful state estimation systems. We have illustrated these is- sues using a probabilistic game state estimation in autonomous robot soccer. Our results are preliminary and a lot more has to be done to understand the design of complex state estimators properly.