---

---

---

---

Science Current Events and Science News RSS Feeds

---

---

Email a Friend Printer Friendly

Robots can learn from their mistakes

September 10, 2004

In the near future, intelligent, automated robots will help us in a number of different fields, such as transports, cleaning, mining, and agriculture. These robots will steer themselves without human involvement and will plan their own tasks. This makes great demands on safety, so that people, animals, or other things in the environment are not injured or damaged.

These robots will also have to be able to discover if their work is not proceeding according to plan. In other words, the robots will need to have the ability to notice that something is wrong and even to find out what went wrong.

The most common way of finding mistakes in robots today is to use models that predict the robot's position after a certain action. If the robot winds up in the predicted position, it is acting correctly and if it finds itself in another position, then something is wrong. But with fully automated robots, it is often difficult, not to say impossible, to set up precise models.

In his doctoral dissertation, Ola Pettersson at oerebro University has studied how mistakes can be discovered without models. Instead Ola has investigated the behavioral patterns of robots to see if they have acted incorrectly.

"Studying behavioral patterns in order to detect mistakes in mobile robots is a new method," says Ola Pettersson. "Research is therefore needed to find out whether the method really works. In our research we suggest that the robot should learn what is right or wrong behavior."

The method the Oerebro scientist is using will make it easier for humans to use robots. "We will simply be able to tell the robot that it's doing something right or wrong, instead of describing it in mathematical models," says Ola Pettersson.

The dissertation shows that it is possible to detect a mistake and even to determine what type of mistake it was, by studying the robot's patterns of behavior. Several different techniques have been tested, and their usefulness has been shown in a large number of experiments, using both simulated and real robots. Statistical methods have been used to compare the various results.

Ola Pettersson will defend his thesis in computer science on Friday, October 1, 2004, at 1:00 p.m. in Auditorium T, Oerebro University. The dissertation is titled Model-Free Execution Monitoring in Behavior-Based Mobile Robotics.



Vetenskapsrådet (The Swedish Research Council)