Granted Patent B2

US 12,576,519 · App. 18/365,228 · Granted Mar 17, 2026

Inventors: Daniel Tanneberg (Offenbach, DE); Michael Gienger (Offenbach, DE)

Assignee: Honda Motor Co., Ltd.

B25J9/163B25J9/1661B25J11/008G06F18/23

Abstract

Claims (112)

1 . A computer-implemented method for learning operators for planning a behavior of an autonomous device, the method comprising:

obtaining a set of training data including observations on a plurality of skills executable by the autonomous device in an interaction with at least one entity for addressing a first task;

obtaining a set of predetermined data defining a hierarchy of entities including the at least one entity;

learning a set of individual skills based on the obtained set of training data;

generating an enhanced set of skills by generalizing the individual skills of the learned set of individual skills based on the predetermined set of data that defines the hierarchy of entities and adding the generalized individual skills to the set of individual skills, and

repetitively performing the step of generating the enhanced set of skills by generalizing the individual skills of the learned set of individual skills based on the predetermined set of data until a convergence criterion is met;

determining that the convergence criterion is met when, in the step of generalizing the individual skills of the learned set of individual skills, no new generalized skill can be found; and

performing behavior planning of the autonomous device for addressing a second task different from the first task based on the enhanced set of skills.

2 . The method according to claim 1 , wherein

the step of learning a set of individual skills based on the obtained set of training data includes:

calculating, for each demonstration tuple of the training data, a grounded effect by calculating a difference between an initial state and a resulting state, wherein the each demonstration tuple includes an initial state before the action was executed, the action, and the resulting state after the action is executed;

clustering the demonstration tuples based on the skill and the calculated grounded effects into effect clusters; and

lifting the demonstration tuples included in each of the effect clusters by replacing the at least one entity with at least one variable that includes the at least one entity.

3 . The method according to claim 2 , wherein

the step of learning a set of individual skills based on the obtained set of training data further includes extracting preconditions for each of the determined effect clusters.

4 . The method according to claim 2 , wherein

the step of learning a set of individual skills based on the obtained set of training data further includes

calculating preconditions as the intersection of all initial states of the demonstration tuples in each of the determined effect clusters,

generating a learned individual skill including the skill, the calculated preconditions, the calculated effect, associated with the learned individual skill, and

adding the learned individual skill to the set of individual skills.

5 . The method according to claim 4 , wherein

repetitively performing the step of generating the enhanced set of skills by generalizing the individual skills of the learned set of individual skills based on the predetermined set of data includes

generating pairs of skills with similar effects from the set of individual skills to generate possible variable type generalizations,

calculating generalized parameters from the skills based on the predetermined set of data defining the hierarchy of entities for each of the generated pairs of skills,

generating the generalized skill s g by replacing skill parameters of the generated pairs of skills with the calculated generalized parameters,

updating the calculated preconditions based on an intersection of the preconditions of s i and s j , and based on all elements of a powerset (P(⋅)) of the difference of their preconditions to generate a set of candidate preconditions based on

{

(

s

i

pre

⋂

s

j

pre

)

⋃

pre

|

pre

∈

P

⁡

(

d

)

}

;

with

d

=

(

s

i

pre

⋃

s

j

pre

)

-

(

s

i

pre

⋂

j

pre

)

calculating a score v g as a recall based on demonstration clusters from s i and s j ,

calculating a score v i as a recall based on the demonstration clusters from s i ,

calculating a score v j as a recall based on the demonstration clusters from s j ,

determining whether replacing the individual skills s i and s j , by the generalized skill s g based on the calculated scores v g and v i and v j .

6 . The method according to claim 5 , wherein

determining to replace the individual skills s i and s j , by the generalized skill s g in case of determining that a performance of the generalized skill s g is equal to or exceeds an average performance of the individual skills s i and s j .

7 . The method according to claim 5 , wherein

determining to replace the individual skills s i and s j , by the generalized skill s g in case of determining that

v

ℊ

≥

v

i

+

v

j

2

.

8 . The method according to claim 1 , wherein

the step of performing behavior planning of the autonomous device for addressing the second task different from the first task based on the enhanced set of skills further includes

determining whether unreachable goal predicates exist by calculating predicates in an obtained planning task goal description that cannot be reached with the enhanced set of skills starting from an initial state.

9 . The method according to claim 8 , wherein

in case of determining that unreachable goal predicates exist that cannot be reached with the enhanced set of skills starting from an initial state,

the step of performing behavior planning of the autonomous device for addressing the second task further includes

calculating repeatedly for each unreachable goal predicate potentially suitable skills for generalization from the enhanced set of skills, wherein the potentially suitable skills have the unreachable goal predicate in their effect set,

generalizing each potentially suitable skill by determining a lowest common ancestor of variables in the unreachable goal predicate and a matching effect predicate of the potentially suitable skill,

generalizing all variables of the potentially suitable skill to a lowest common ancestor of itself and the available entities in the planning task to generate a parameter generalization,

generating a new skill with the parameter generalization, and add it to the enhanced set of skills,

determining whether predicates in preconditions of the generated new skill can be reached, and adding determined unreachable predicates to the set of unreachable goal predicates,

propagating the proposed parameter generalizations to all skills that have a same set of parameters to generate further generalized skills and adding the further generalized skills to further enhance the enhanced set of skills, and

resuming performing behavior planning of the autonomous device for addressing the second task based on the further enhanced set of skills.

10 . The method according to claim 1 , the method comprising generating a control signal for controlling at least one actuator of the autonomous device to perform the second task based on the determined behavior, and outputting the generated control signal to the at least one actuator.

11 . A non-transitory computer-readable storage medium storing a program comprising instructions executable by a computer or a digital signal processor, which, when the program is executed by the computer or the digital signal processor, cause the computer or the digital signal processor to carry out the method of claim 1 .

12 . A system for learning operators for planning a behavior of an autonomous device for solving a predefined task, the system comprising:

a first acquisition unit configured to obtain a set of training data including observations of a plurality of skills executable by the autonomous device in an interaction with at least one entity for addressing a first task;

a second acquisition unit configured to obtain a predetermined set of data defining a hierarchy of entities including the at least one entity;

a processor configured to;

learn a set of individual skills based on the obtained set of training data;

generate a set of generalized skills by generalizing the learned set of individual skills based on the predetermined set of data that defines the hierarchy of entities, and

repetitively perform generating the enhanced set of skills by generalizing the individual skills of the learned set of individual skills based on the predetermined set of data until a convergence criterion is met;

determine that the convergence criterion is met when, no new generalized skill can be found; and

perform behavior planning of the autonomous device for addressing a second task different from the first task based on the generalized set of skills.

13 . A robotic system comprising the system according to claim 12 and at least one autonomous device.

Assignments (2)

Priority Claims (2)

Continuity (1)

References Cited (48)