Method and system for robot manipulation planning

1. A computer-implemented method for planning a manipulation task of a robot, comprising the following steps:

learning a number of manipulation skills and generating a symbolic abstraction of each of the manipulation skills, by which a respective statistic model of a transition to a respective final state from a respective current state is determined for each of the learned manipulation skills;

determining a concatenated sequence of a subset of the manipulation skills selected from the number of learned manipulation skills based on their symbolic abstractions so that a given goal specification indicating a given complex manipulation task is satisfied, wherein:

the determining of the concatenated sequence includes generating a complete model corresponding to a combination, in the sequence, of all of the subset of the manipulation skills; and

the generating of the complete model is performed by computing, for each of a plurality of pairwise combinations of individual manipulation skills of the subset, a respective probability of a transition from a first one of the individual manipulation skills of the respective pairwise combination to a second one of the individual manipulation skills of the respective pairwise combination according to a divergence of emission probabilities between an end state of the first one of the individual manipulation skills of the respective pairwise combination and an initial state of the second one of the individual manipulation skills of the respective pairwise combination; and

executing the sequence of manipulation skills.

2. The method according to claim 1 , wherein:

the statistical model is a Task Parameterized Hidden Semi-Markov model (TP-HSMM); and

the learning of the number of manipulation skills includes, for each of the number of manipulation skills:

demonstrating and recording a plurality of manipulation trajectories;

determining a respective instance of the TP-HSMM depending on the plurality of manipulation trajectories of the respective manipulation skill; and

determining a respective one of the symbolic abstractions.

3. The method according to claim 2 , wherein the generating of the symbolic abstractions of the manipulation skills includes constructing a Planning Domain Definition Language (PDDL) model in which objects, the initial states, and a goal specification define problem instances, and in which predicates and actions define a domain of a given manipulation, wherein the symbolic abstractions of the manipulation skills use a classical PDDL planning language.

4. The method according to claim 2 , where the determining of the concatenated sequence of manipulation skills is performed such that a probability of achieving the given goal specification is maximized, wherein a PDDL planning step is used to find a sequence of actions to fulfill the given goal specification, starting from a given one of the initial states.

5. The method according to claim 2 , where the transition probabilities are determined using Expectation-Maximization.

6. The method according to claim 2 , wherein the symbolic abstractions are determined by mapping low-variance geometric relations of segments of the manipulation trajectories into a set of predicates.

7. The method according to claim 2 , wherein the determining of the concatenated sequence of manipulation skills includes optimizing a goal of minimization of a total length of a combined trajectory of the concatenated sequence.

8. The method according to claim 7 , wherein the determining of the concatenated sequence of manipulation skills includes selectively reproducing one or more of the manipulation skills of the sequence of manipulation skills so as to maximize a probability of satisfying the given goal specification.

9. The method according to claim 7 , wherein the determining of the concatenated sequence of manipulation skills further includes searching a most-likely complete state sequence between initial and goal states of the complex manipulation task as a whole using a Viterbi algorithm.

10. The method according to claim 7 , wherein the determining of the concatenated sequence of manipulation skills further includes searching a most-likely complete state sequence between initial and goal states of the manipulation task as a whole using a Viterbi algorithm that includes missing observations and duration probabilities.

11. The method according to claim 1 , wherein an individual one of the manipulation skills of the subset is included multiple times within the concatenated sequence.

12. A device for planning a manipulation task of a robot, the device comprising a processor, the processor configured to:

learn a number of manipulation skills, thereby generating a symbolic abstraction of each of the manipulation skills and by which a respective statistic model of a transition to a respective final state from a respective current state is determined for each of the learned manipulation skills;

determine a concatenated sequence of a subset of the manipulation skills selected from the number of learned manipulation skills based on their symbolic abstractions so that a given goal specification indicating a complex manipulation task is satisfied; and

instruct execution of the sequence of manipulation skills;

wherein:

the determination of the concatenated sequence includes generating a complete model corresponding to a combination, in the sequence, of all of the subset of the manipulation skills; and

the generating of the complete model is performed by computing, for each of a plurality of pairwise combinations of individual manipulation skills of the subset, a respective probability of a transition from a first one of the individual manipulation skills of the respective pairwise combination to a second one of the individual manipulation skills of the respective pairwise combination according to a divergence of emission probabilities between an end state of the first one of the individual manipulation skills of the respective pairwise combination and an initial state of the second one of the individual manipulation skills of the respective pairwise combination.

13. A non-transitory machine-readable storage medium on which is stored a computer program that is executable by a computer and that, when executed by the computer, causes the computer to perform a method for planning a manipulation task of a robot, the method comprising the following steps:

learning a number of manipulation skills and generating a symbolic abstraction of each of the manipulation skills, by which a respective statistic model of a transition to a respective final state from a respective current state is determined for each of the learned manipulation skills;

determining a concatenated sequence of a subset of the manipulation skills selected from the number of learned manipulation skills based on their symbolic abstractions so that a given goal specification indicating a given complex manipulation task is satisfied, wherein:

the determining of the concatenated sequence includes generating a complete model corresponding to a combination, in the sequence, of all of the subset of the manipulation skills; and

the generating of the complete model is performed by computing, for each of a plurality of pairwise combinations of individual manipulation skills of the subset, a respective probability of a transition from a first one of the individual manipulation skills of the respective pairwise combination to a second one of the individual manipulation skills of the respective pairwise combination according to a divergence of emission probabilities between an end state of the first one of the individual manipulation skills of the respective pairwise combination and an initial state of the second one of the individual manipulation skills of the respective pairwise combination; and

executing the sequence of manipulation skills.