An example of adaptive behavior

The Idea

To enable JANUS to work well in an open world, the heuristic-based methods must be also have the ability to learn by themselves. This is enabled by collecting experience of good behavior while executing tasks. Such knowledge can be used later to produce good behavior in new but similar situations.

The Task

In this example, JANUS has to execute a touch-mission with its left hand.

The left picture shows the initial position and the right picture the desired position. Unfortunately a red column lies directly between start and goal. Therefore measures must be taken to nogotiate the obstacle.

Novice Phase

At the beginning of this phase, JANUS does not know how to get around the red column in an elegant way. Therefore it tries a few times to find simply a valid way around.

Collecting experience

During this phase JANUS learns a lot about movements that are impossible if an obstacle lies in a similar position to the red column. This knowledge is stored in form of imaginary obstacles which are inserted in the thinking process if JANUS discovers that the arm has some problems carrying out a specific plan. The following films give a rough idea of how this knowledge is built up and used. The blue cubes mark the virtual obstacles and the red line shows the desired path (the plan) to be taken by the TCP. The schedule corresponds to the movements shown in the previous film

Remembering "bad regions"

All in all it does not look too elegant, but JANUS can solve the task after a reasonable period of learning using imaginary obstacles.

Expert Phase

Now JANUS is an "expert" for obstacles that have similar properties as the red column and which are located in a similar position. For a new task that has a similar (in terms of blocking effect) set of obstacles, JANUS reuses parts of the above knowledge and immediately constructs the following plan.

The execution of this plan looks like :

Applying knowledge

allowing JANUS to perform a reasonable touch straight away.

Is this a general solution to obstacle avoidance problems?

No of course not. The mechanism shown above works for surprisingly many situations, but not for all possible ones. In the context of the JANUS experiment this is to be expected, because the overall mechanism behind the control of JANUS is achieved using a set of many simple "90 percent good heuristics". We achieve a reasonable overall behavior by using combinations of heuristics and learning at various levels of behavior.