Jump label

Service navigation

Main navigation

You are here:

Main content

Behavior

Robotic manipulators are normally assigned to tasks that require execution of recurring motions in quasi static workspace. Thus the motions are mostly calculated in advance. Those preplanned Motions are replayed during task execution.

Whenever a robot is acting or moving within a dynamic environment such static motion planning becomes insufficient. To act according to changes the robot need sensors to observe its environment. The decision which action will be chosen according to the current sensor input is called behavior.

With rising potential of modern autonomous robots they become able to solve more and more complex tasks. Improved hardware design and better mechanic of aktuators enables modern robots to choose between more and more different actions to react to different situations resulting in a more complex and difficult task to decide which action is the most rewarding one. Thus behavior research becomes more and more important in modern robotics.

In the past behavior problems were mostly solved by application of classical approaches of Artificial Intelligence. The application of logical rules to a world model becomes more and more unsuitable with higher numbers of logical possibilities and  rising complexity of the world model.

This the research focus of our institute is the design of new methods to define behavior decisions, aiming to intuitively design even complex tasks. The next step would be to allow a robot to design its own behavior allowing learning and selfadaption with the help of Computational Intelligence techniques.

 

XABSL

Extensible Agent Behavior Specification Language (XABSL) represents a programming language especially designed to specify robotic behavior. A set of state machines, the options, represents possible situations. An decision tree, containing all option, helps to choose the right action according to the current sensor input. These options can be developed independently and are easy to specify with the help of the XABSL syntax. Since the same input in the same state results in the same action the behavior cannot adapt or learn according to the situations.

Since the developed behavior is represented by a tree structure its easy to visualize. The possibility to visualize decisions with the help of transitions between tree nodes aids the programmer in debugging the behavior and detect faulty design of decisions.

The Robotics Research Institute has successfully utilized XABSL based behavior specifications for many years in different leagues of the RoboCup.

 

Adaptive Behavior

Many approaches to behavior design are approaches static in nature. A once specified behavior is not changed during execution, resulting in the same decisions every time. Even if a decision results in a faulty action in a specific situations the same action would be chosen if the same situation occures again leaving the robot without the possibility to correct it's error

Thus the  current research of our institute focuses on approaches to behavior design, which are more flexible allowing a self-adaptation and learning on the part of the robot. Thereby we have to distinguish between approaches that learn offline, before behavior execution, and online, during behavior execution.

Different methods of the research field of Computational Intelligence can be applied to design such behaviors. A very promising research project is the application of behavior-networks. Those make decisions about which action is to choose purposeful according to the current situation with the help of activation-funktions. Therefore a network containing aim- and competence-nodes is generated. The aim-nodes describe worldstates, the roboter wants to achieve. Competence-nodes describe the robots abilities under specific conditions. Each competence also contains expected effects which will most likely occur if the competence is activated. If an effect is expected to make aim true or false these nodes are connected. With the help of these edges an activation potential is distributed along the network which decides which action will be chosen.

This approach is a promising solution to integrate learning by demontartion  the design of the network is very flexible in itself and allows an easy adaptation by altering the activation weights.