Welcome to my home-page
Alfred Schurmann

How to learn the motivated autonomous robot?

          For practical reasons, the motivated autonomous robot (a household as well as an industry robot) must be able to learn and to adapt to changed surrounding. I assume that the robot has some known Programs for building visual and acoustic (may be also odor) patterns.

          The robot must have additionally the following adaptation abilities in order to handle intelligently also in a changing surrounding:
i.    recognize new features of an object,
ii.   build the object model for a new object and insert it into its knowledge base,
iii.  build situation models for new important situations,
iv.  adapt activity schemas for operating a new machine or device, may be in new rooms,
v.  build new activity schemas.

1. Learning new features of a known object

          The robot can recognize autonomously visual and acoustic (and odor) changes of an object, for which it has an object model, and build new visual and acoustic (odor, respectively) patterns using said programs for building patterns, and insert these new patterns in the model of this object (in the tree like knowledge base, see "1. Abstract object and situation models ...."in "Intelligence of the motivated...."). The robot can also recognize some habits or states of an object and insert them into the object model; examples of such features: "person x get up between 6:30 and 6:45 a.m.", person x has breakfast at 7:15 o` clock", "device g1 does not work", "no water is coming in the wash-machine when it is switched on",  person x is not in the flat from 7:50 a.m. to 5:00 p.m. from Monday to Friday".
         However, many new features of an object must be put into the robot by an authorized person (the robot inserts then these features into the object model); examples: "person y is ill", "child k1 is 11 years old", "children younger than 11 years must not take matches and lighter into hands".

2. Recognizing/learning an unknown object

          Object models in the knowledge base are stored in a tree like graph as its nodes; near the top of the tree are very abstract object models (e. g. for a machine, a container, a plant); near the bottom are concrete or less abstract object models (s. "1. Abstract object and situation models...." in "Intelligence of the motivated...."). An object model has all features of the predecessor model. When the robot notices an object, On, (visual or by means of other sensors), it searches top down in the tree for the smallest class, M(Ok), of objects to which object On belongs (i.e. finds the object model M(Ok) for which On is an instance, and no successor of M(Ok) is a model for On). In the case, if there is no concrete object model for On (i.e. M(Ok) has successors) then the robot considers On as a new (unknown) object and it builds
a) an object model, Ms(On), for the object On without help, or
b) object model, M(On), by means of a dialog with an authorized person.
         In the case (a), the robot builds the model Ms(On) on the basis of the model M(Ok); it enters in Ms(On) only these features of the object On which it has recognized. The model Ms(On) is inserted in the tree as a successor of the model/node M(Ok). The model Ms(On) has all features of the object models along the path from M(Ok) to the top; so the robot learns additional features of the object On.
        In the case (b), the robot builds the model MOn) for the object On in a dialog with an authorized person, Pz. First, Pz must accept that M(On) is an instance of M(Ok) (the robot could also recognize that object On is an instance of an other model, M(Or), too) and should be inserted as a successor of M(Ok). If Pz does not accept it, the robot propose an other object model, M(Od), for which On may be an instance. The person Pz may build also a new abstract model, M(Oa), as a successor of M(Ok) (M(Od), respectively) in a dialog with the robot, and then insert the model M(On) as a successor of M(Oa). On the basis of the model M(Ok) (M(Oa), respectively), the robot asks about features of the object On, and the person Pz replies. In this way is build the object model M(On).

3. Learning new situations and activities

             When the surrounding of the robot changes (e.g. the robot moves flat or factory building, it should operate a new machine), new situations appear to which the robot must react (i.e. perform activities). Thus, it must adapt some situation models and activity schemas to the new situations and activities. This adaptation (the learning of a new activity) is carried out as follows.
       Assume, the robot should operate a new machine, Mn. It searches for and finds a machine, Ma, similar to the machine Mn and which it can operate (i.e. has appropriate activity schemas); the object model for the machine Mn has been built. The robot adapts an activity schema, AMa,  for operating Ma to activity schema for operating the machine Mn by means of a dialog with an authorized person; it asks and does the following:
i)  with regard to a situation model in the activity AMa,
-  which object models in a situation model (in the activity schema AMa) must be replaced by new ones?
-  at which distances are the objects in the new situation model (location of objects)?
-  inserts the recognized moving characteristics (position, moving line, direction and speed) of objects in the new situation
ii) with regard to a sub-activity (of the activity schema AMa) connected with an directed arc (SM1,SM2) (a sub-activity
    contains a sequence of elementary activities and operations),
-  whether the sub-activity should be accepted, removed or modified?
-  in the last case, the robot asks, for each elementary activity/operation, whether it should be accepted, removed, replaced
    or modified,
-  in the last two cases, the person replies giving the elementary activity which should be done at this place.
       After the new activity schema has been adapted to the new Machine Mn, it must be applied experimentally to the machine Mn by the person and possibly carry on the adaptation.
    In this way the robot can also build/learn new activities with the help of an authorized person.

4. Should other learning methods be applied?

        Some readers may think, the robot learns not enough automatically - the new knowledge and activities are put in the robot mainly by a human. I think the outlined method for learning the robot are right for a robot which does not work on the Moon or 1000 m under water, because even intelligent pupils and students learn in a similar way - the knowledge and new abilities are put into the brains of pupils/students by specialists (directly as teachers or indirectly by means of books) with the active participation of these pupils/students; only the recording of knowledge in the considered (in (1), (2) and (3)) robot goes faster than into a human.

Are "robot schools" a solution?
                Also the developmental robots of J. Weng and Y. Zhang (in USA) and E. Prassler (in Germany) learn only to some degree autonomously, because this learning is partially supervised by a person; - also children do not learn automatically but with the help of adults. The results of learning robots of Weng and Zhang, and the expected learning results (in the project XPERO) of a robot of Prassler, are modest.
        No better results are possible, as long as the authors of such projects do not build in the foundation stones of intelligence (or intelligence algorithms) in a robot ( a dog can learn only what its primitive intelligence algorithms allow).Some foundation stones of intelligence are build in the considered (in (1), (2) and (3)) motivated autonomous robot (see Intelligence of the motivated....). By means of these intelligence properties, the robot can (as outlined above) not only classify/identify unknown objects and build its models, but also build new complex activities with the help of a person.
         From theoretical point of view, the developmental robot approach is interesting, because, may be very effective developmental algorithms could be created and implemented in robots, so that they could explore/recognize situations and do complex activities in its surrounding.
         From practical point of view, it is not controllable what an developmental robot learns, what it does not learn and to what activities it is able; such robots are not reliable, they even could behave aggressively towards humans; of course, not in the next 5 years. However, when in the next 40 years, as I said, very effective developmental algorithms (comprising all important  foundation stones of intelligence)  would be implemented in robots, then they could be dangerous for humans (but not so dangerous as Kalashnikovs).
          The supporters of reliable and human friendly robots ought to be on the alert. Copying developmental biology and developmental psychology theories is not the way to develop reliable robots.
           We humans need reliable, human friendly robots-servants - and the motivated autonomous robot, which learns as said above, is such a one.

Dr. A.Schurmann                                                                                                                     Modified
: November 29, 2006

Goto startsite    

Email: alfschurman@yahoo.de