Learning Repetitive Robot Programs from Demonstrations using Version Space Algebra

M. Pardowitz, B. Glaser, and R. Dillmann (Germany)


Programming by Demonstration (PbD), Task Learning, Version Space Algebra


Robots are expected to move from industrial environments to the household domain. Programming those robots can not be done in conventional ways since the wide variety of tasks the user needs to be accomplished can not be foreseen by the robot manufacturer. One solution to this problem is the Programming by Demonstrion (PbD) paradigm, where the user himself programs the robot by demonstrating the task to be performed. In this setting, the robot observes the user acting and infers the program that fits the users intention. One problem not adressed by current PbD systems is how to infer loop structures that are necessary to per form repetitive tasks from the demonstrations. This pa per proposes a theoretical framework for efficient multi hypotheses tracking for hierarchical robot programs cov ering generic loop statements called Version Space Alge bra (VSA). A version space design is implemented for and evaluated with tasks from the household domain, i.e. the tasks of unloading a dishwasher basket and laying the table for multiple persons.

Important Links:

Go Back