METHOD, COMPUTER PROGRAM AND SYSTEM FOR CONTROLLING A PLURALITY OF ROBOTS, AND COMPUTER-READABLE STORAGE MEDIUM

Information

  • Patent Application
  • 20190070725
  • Publication Number
    20190070725
  • Date Filed
    September 14, 2017
    7 years ago
  • Date Published
    March 07, 2019
    5 years ago
Abstract
A method for controlling a plurality of agents to complete a mission, including deriving a decomposition set of decomposition states in a set of possible states of an automaton, wherein the automaton characterizes the mission, deriving a sequence of actions to be carried out by the plurality of agents depending on the decomposition set, where each action is to be carried out by at most one of the plurality of agents.
Description
CROSS REFERENCE

The present application claims the benefit under 35 U.S.C. § 119 of German Patent Application No. DE 102017215311.3 filed on Sep. 1, 2017, which is expressly incorporated herein by reference in its entirety.


FIELD

The present invention relates to a method for controlling a plurality of robots, a computer program and a system configured to carry out the method and a computer-readable storage medium.


BACKGROUND INFORMATION

Linear Temporal Logic (LTL) is a mathematical specification logic which is able to capture temporal relationships. It originally results from the field of model checking and verification.


SUMMARY

An example method in accordance with the present invention may have the advantage that it automatically generates optimal action-level behavior for a team of robots (or agents). In accordance with the present invention, separable tasks are optimally allocated to the available agents or robots, while avoiding the need of computing a combinatorial number of possible assignment costs, where each computation would itself require solving a complex planning problem, thus improving computational efficiency, in particular for on-demand missions where task costs are unknown in advance.


Further advantageous aspects are described herein.


LTL can be applied to robotic behavior planning. It then provides a formalism to specify an expected behavior in an unambiguous way. As such, an LTL specification can be used to describe a result of the expected behavior, while the way to achieve this result can be automatically derived by the system.


An LTL formula ϕ can be defined over a set of atomic propositions Π. A single atomic proposition is notated π∈Π. Each atomic proposition can be either true custom-character or false custom-character. To express temporal relationships, the semantics of the formula ϕ can be defined over a sequence σ of propositions. Conveniently, the sequence σ is defined as a function of time index t, and σ(t)⊆Π for each t.


A proposition may be expressed in terms of concatenations of atomic proposition using the Boolean operators and (“Λ”) and/or or (“V”).


Boolean operators custom-character (“not”) and Λ (“and”) and temporal operators custom-character (“next), custom-character (“until”) and custom-character (“release”) can be used to recursively define a satisfaction relation custom-character as follows:

    • σ(t)custom-characterπiff π∈σ(t)
    • σ(t) custom-charactercustom-character991 iff custom-character(σ(t) custom-characterϕ1)
    • σ(t) custom-characterϕ2 Λϕ2 iff σ(t) custom-characterϕ1 Λσ(t) custom-characterϕ2
    • σ(t) custom-characterϕ1 iff σ(t+1) custom-characterϕ1
    • σ(t) custom-characterϕ1custom-characterϕ_2 iff ∃t2≥t such that σ(t2) custom-characterϕ2 and ∀ti ∈[t,t2) it holds that σ(t1) custom-characterϕ1
    • σ(t) custom-characterϕ1custom-characterϕ2 iff t1=∞ or ∃t1≥t sucht that σ(t1) custom-characterϕ1 and ∀t2 ∈[t, t2) it holds that σ(t2) custom-characterϕ2.


A Non-deterministic finite automaton custom-character is characterized by a tuple custom-character=(Q, Q0,α, δ,F) consisting of

    • a set of states Q,
    • a set of initial states Q0⊆Q,
    • a set of Boolean formulas a over the set of atomic propositions Π,
    • transition conditions δ:Q×Q→α, and
    • a set of accepting (final) states F⊆Q.


Note that the term nondeterministic finite automaton custom-character is used in the broad sense that also encompasses deterministic finite automata, i.e., every deterministic finite automaton is also a nondeterministic finite automaton in this sense.


For two states qi,qj∈Q, the absence of a transition between these two states is denoted by δ(qi,qj)=⊥. Accordingly, there exists a transitions between these two states if δ(qi,qi)≠⊥, and the Boolean formula δ(qi,qj) denotes the transition condition.


A sequence σ over propositions when applied to the nondeterministic finite automaton custom-character describes a sequence of states q∈Q, called a run ρ:custom-character∪{0}→Q. The run ρ is called feasible if it starts in an initial state ρ(0)=q0 with q0 ∈Q0 and if all transition conditions are satisfied along the run σ(t) custom-characterδ(ρ(t−1), ρ(t)) for all t. A run ρ is called accepting if it is feasible and ends in an accepting state qn ∈F. Sequence σ is called to violate the specification if it does not describe a feasible run.


If sequence σ describes a feasible but not an accepting run, it does not satisfy the specification. If sequence σ forms a prefix of an accepting run and can be extended to a sequence satisfying the specification, it is said that σ partially satisfies ϕ.


A given mission custom-character that is to be completed by a set of agents can be expressed in terms of an LTL formula ϕ or equivalently in terms of a nondeterministic finite automaton custom-character. It may be given as a set of tasks custom-character={custom-character, . . . , custom-character}. The tasks custom-character are independent parts of the mission that can be allocated to the agents. The above-mentioned set of tasks is called a decomposition of the mission custom-character. This implies two decomposition properties which are fulfilled by all tasks. The tasks have to be mutually independent, i.e. execution or non-execution of a first task custom-character must not violate a second task custom-character. Furthermore, completion of each of the tasks custom-character, . . . , custom-character implies completion of the mission custom-character.


This enables acting agents to act independently, without any coordination, and execution does not have to be synchronized between the agents.


A task custom-character may be specified by an LTL formula ϕ(i) or a nondeterministic finite automaton custom-character(i). The conditions of mutual independence and completeness can be expressed by saying that any strategy that satisfies each LTL formula ϕ(i) that specifies task custom-charactercustom-character for a strict subset of tasks custom-character⊂{custom-character, . . . , custom-character} partially satisfies the LTL formula ϕ that specifies the mission custom-character.


Consequently, completing the subset of tasks custom-character can be associated with reaching a certain state in the nondeterministic finite automaton custom-character that also specifies the mission custom-character. However, not every state implies completion of a set of tasks when requiring the above properties.


Therefore, a first aspect of the invention makes us of a decomposition set custom-character of the nondeterministic finite automaton custom-character that specifies the mission custom-character. The decomposition set custom-character contains all states q which can be associated with completing the subset of tasks custom-character which is a subset of the decomposition {custom-character, . . . , custom-character} of the mission custom-character.


Based on the decomposition set custom-character, a team model that can be augmented to contain all possible decomposition choices. This team model can then be used for efficiently planning an optimal decomposition and a corresponding allocation of tasks to agents. It can also be used for at the same time planning action sequences to execute the mission.


To make the relation between formula ϕ(i) that specifies task custom-character and the LTL formula ϕ that specifies the mission custom-character clear, we let {ϕ(i)} with i=1, . . . , n be a set of finite LTL specifications for the tasks and {custom-character} and {σi} denote sequences that satisfy the tasks {ϕ(i)}, i.e. σicustom-characterϕ(i)∀i∈{1, . . . , n}. The tasks {custom-character} are a decomposition of the mission custom-character if and only if σj1 . . . σji . . . σjncustom-characterϕ for all permutations of ji∈{1, . . . , n} and all respective sequences σi. If tasks {custom-character} are a decomposition of the mission custom-character, they fulfill the decomposition properties of independence and completeness regarding the mission custom-character.


The several aspects of the present invention avoid the need of computing a combinatorial number of possible assignment costs, where each computation would itself require solving a complex planning problem, thus improving computational efficiency, in particular for on-demand missions where task costs are unknown in advance.


Therefore, in the first aspect, the present invention includes a method for controlling a plurality of agents to complete the mission custom-character, comprising the steps of:

    • deriving the decomposition set custom-character of decomposition states in the set of possible states Q of the automaton custom-character, wherein the automaton custom-character characterizes the mission custom-character,
    • deriving a sequence βfin of actions (a1,a2, . . . , an) to be carried out by the plurality of agents depending on the decomposition set custom-character, where each action (a1,a2, . . . , an) is to be carried out by at most one of the plurality of agents.


Preferably, the method may further comprise the step of controlling the plurality of agents in accordance with the derived sequence βfin of actions (a1,a2, . . . , an).


In another aspect of the present invention, the method further comprises the step of generating the decomposition set custom-character by exploring an essential sequence σe of an accepting run ρi through one or more candidate decomposition states qi.


Preferably, this method further comprises the step of adding the one or more candidate decomposition state qi to the decomposition set custom-character depending on whether a complementary sequence {circumflex over (σ)}e to the explored essential sequence σe around the respective one or more candidate decomposition state qi is accepting.


Even more preferably, the decomposition set custom-character consists of all those states qi in the set of possible states Q of the automaton custom-character, for which the complementary sequence {circumflex over (σ)}e to the explored essential sequence σe around the respective state qi is accepting.


In another aspect of the present invention, the method further comprises the step of generating a team model custom-character based on the automaton custom-character that characterizes the mission custom-character and based on automata custom-character(r) that each characterize the capabilities of one of the plurality of agents.


Preferably, it may be envisaged that the team model custom-character comprises a set of actions custom-character that comprises switch transitions custom-character which change the acting agent from one of the plurality of agents to another one of the plurality of agents.


That is, individual agents are assumed to act independently and based on the decomposition set, special transitions (the switch transitions custom-character) indicate the options to split the mission at some state and allocate the rest to a different agent. In other words, the switch transitions custom-character are purely virtual transitions that by themselves do not lead to any actions of the agents.


More preferably, these the switch transitions custom-character are configured to each change the acting agent from one of the plurality of agents to a next one of the plurality of agents. This is particularly useful because it implies that, starting in a state associated with a first agent r, no state associated with an agent r′<r can be reached by any path in the team model.


As indicated above, preferably the switch transitions custom-character are configured such as to only act if the automaton custom-character is in a decomposition state.


In another aspect of the present invention, the method further comprises the step of deriving the sequence βfin of actions (a1, a2, . . . an) to be carried out by the plurality of agents by a label-setting algorithm in which each state s of a set of states custom-character of the team model custom-character is associated with labels l that are characterized by a sequence β of action leading to the respective state s. That is, the label-setting algorithm searches for a final label lfin. Finding the final label lfin is equivalent to finding the respective sequence βfin of actions that satisfies the mission.


Preferably, this method further comprises the step of constructing a reachable set of temporary labels Lt,s for each state s and a set of permanent labels Lp,s.


Even more preferably, this method further comprises the step of constructing, for each selected label l*, a set V of consecutive labels v by extending an action sequence β associated to the selected label l* by all available actions a and adding the resulting labels lv to the reachable set of temporary labels Lt,s.


Preferably, each label l comprises at least one component that characterizes a cost ĉβ under the corresponding sequence β of actions a.


Even more preferably, it may be envisaged that the derived sequence βfin of actions (a1,a2, . . . , an) to be carried out by the plurality of agents is the one out of all actions that satisfy a characterization ϕ of the mission custom-character that minimizes a team cost {circumflex over (κ)} which depends on the component that characterizes the cost ĉβ.


Preferably, only actions a resulting in Pareto-optimal labels lv at their target state v are added to the reachable set of temporary labels Lt,s. This is a very efficient implementation.


In another aspect of the present invention, the component that characterizes the cost ĉβ under the corresponding sequence β of actions a depends on costs ca,r associated with each of these actions a with one component each for each one of the agents.


Preferably, the component that characterizes the cost ĉβ under the corresponding sequence β of actions a is stored in memory by way of a data structure that comprises at least one component cβ,r that characterizes costs associated with a selected one of the agents 11, 12, 13 and at least one component ∥(cβ,1, . . . , cβ,r-1)T, ∥(cβ,1, . . . , cβ,r-1)T1 that characterizes the costs associated with a group of agents that precede the selected one of the agents.


This makes use of the surprising fact that, starting in a state associated with agent r, no state associated with a preceding agent r′<r can be reached by any path in the team model custom-character, i.e., no action associated with any r′ will occur in a continuation of the corresponding sequence γ.


In another aspect of the present invention, each label 1 comprises at least one component that characterizes a resource status γ at the respective state s under the corresponding sequence β of actions.


Preferably, the characterization ϕ of the mission custom-character comprises an inequality constraint that restricts the at least one component that characterizes a resource status γ to a predefined region.





BRIEF DESCRIPTION OF THE DRAWINGS

The present invention is explained in more detail below with reference to figures.



FIG. 1 shows a robot control system according to a first aspect of the present invention.



FIG. 2 shows a flow-chart diagram that illustrates a preferred method according to a further aspect of the present invention.



FIG. 3 shows a flow-chart diagram which relates to a preferred algorithm to determine the decomposition set custom-character.



FIG. 4 shows a flow-chart diagram which relates to a preferred method to construct the team model custom-character.



FIG. 5 shows a flow-chart diagram which relates to a preferred method to plan an optimal action sequence.



FIG. 6 illustrates an example of the structure of the team model.





DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS


FIG. 1 shows a robot control system 10 that is configured to plan and allocate tasks to a plurality of robots 11, 12, 13 such that the plurality of agents, preferably robots 11, 12, 13, by fulfillment of their respective tasks, jointly achieve a common goal, thus achieving a predefinable mission custom-character. Robot control system 10 is equipped with communication means (not shown), e.g., a wireless networking transceiver, to communicate with each of the robots 11, 12, 13 via a communication link 22, 23, 24. Similarly, each of the robots 11, 12, 13 is equipped with corresponding communication means.


In a preferred embodiment, the robot control system 10 comprises a computer 20 with memory 21, on which a computer program is stored, said computer program comprising instructions that are configured to carry out the method according to aspects of the present invention described below if the computer program is executed on the computer 20.


In further aspects of the preferred embodiment, the robots 11, 12, 13 comprise a computer 11b, 12b, 13b each, said computer being equipped with a computer memory each (not shown) on which a computer program is stored, said computer program comprising instructions that are configured to carry out some or all of the method according to further aspects of the invention described below if the computer program is executed on the computer 11b and/or 12b and/or 13b. Preferably, the robots 11, 12, 13 each comprise actuators 11a, 12a, 13a that enable each of the robots to physically interact with an environment in which the robots are placed.



FIG. 2 shows a flow-chart diagram that illustrates a preferred method according to a further aspect of the present invention. In a first step 1000 the method receives an agent model custom-character1,custom-character2,custom-character3, . . . each for each of the agents or robots 11, 12, 13. The agent models can for example be read from a dedicated location in computer memory 21.


These agent models custom-character1,custom-character2,custom-character3, . . . are preferably each given as an automaton custom-character=(custom-character,custom-character, custom-character,Π, λ) consisting of

    • a set of states custom-character that the corresponding agent or robot can be in
      • an initial state custom-charactercustom-character
      • a set of possible actions custom-charactercustom-character×custom-character that the corresponding agent or robot can carry out
      • a set of propositions Π
      • a labeling function λ:custom-character→2Π.


Modeling the agent models custom-character1,custom-character2,custom-character3, . . . as an automaton as described is convenient because it is intuitive to model the internal state and the actions of the agents as a state machine. Furthermore, it is convenient to model an abstraction of places in the environment as a topological map.


Independently of step 1000, the method receives a specification of the mission custom-character in step 1100. Preferably, this mission specification custom-character is an LTL specification, e.g. a set of tasks {custom-character, . . . , custom-character}. In a following step 1200, this mission specification custom-character is converted into a nondeterministic finite automaton custom-character. Note that steps 1100 and 1200 are optional. Alternatively, the method may directly receive the mission specification custom-character as the nondeterministic finite automaton custom-character. Then, in step 1300, the method determines the decomposition set custom-character depending on the automaton custom-character. A preferred embodiment of this determination procedure is explained in detail in FIG. 3.


Following steps 1000 and 1300, the method constructs a team model custom-character depending on the automaton custom-character, the decomposition set custom-character and the agent models custom-character1,custom-character2,custom-character3 . . . in step 1400. A preferred embodiment of this construction procedure is explained in detail in FIG. 4.


In the following step 2000, the method carries out a procedure of planning an optimal action sequence βfin based on the team model custom-character which is explained in detail in FIG. 5.


In step 3000, the optimal action sequence βfin is translated into executable commands for the agents or robots 11, 12, 13, for example by means of a lookup table that may be stored in computer memory 21. The executable commands are each associated with one of the agents or robots 11, 12, 13 and distributed to the respective agent or robot 11, 12, 13 via one of the communication links 22, 23, 24. The respective agent or robot 11, 12, 13 then executes this command and preferably upon completion of the command sends a confirmation message to the robot control system 10. In case that the execution of this command is not possible, the respective agent or robot 11, 12, 13 may send an error notification to the robot control system 10, which may react accordingly. In case it receives a confirmation message that a command has been executed, it may send a next command to a next respective agent or robot 11, 12, 13. In the case it receives an error notification, it may enter a dedicated mode, e.g., a shut-down mode of all agents or robots 11, 12, 13.



FIG. 3 shows a flow-chart diagram that depicts a further aspect of the present invention, which relates to a preferred algorithm to determine the decomposition set custom-character. This method starts in step 1310 with a step of reading the aforementioned automaton custom-character. An index i that will be used to label all the states {qi}=Q of the set of states Q that is associated with the automaton custom-character. This index is initialized to an initial value, e.g. i=1. The decomposition set custom-character is initialized as custom-character=Ø.


Then, in step 1320, the method constructs an accepting run ρi that passes through state qi corresponding to the present value of the index i. State qi is the candidate decomposition state. Such an accepting run ρi may for example be constructed by exploring the graph defined by the transition conditions δ associated with the automaton custom-character and constructing a first partial run ρf from state qi to initial state q0 associated with automaton custom-character while considering inverted transitions and a second partial run ρl from state qi to a final state f∈F associated with automaton custom-character. The accepting run ρi that passes through qi may then be constructed by concatenating the inverted first partial run ρf and the second partial run ρl.


In the following step 1330, the method generates an essential sequence σe associated with the accepting run ρi.


A sequence σ is called essential for nondeterministic finite automaton custom-character and associated with a run ρ if and only if it describes the run ρ in custom-character and σ(t)\{π}custom-characterδ(ρ(t−1), ρ(t)) for all t and propositions π∈σ(t), i.e., σ contains only the required propositions.


For example, the essential sequence σe may be generated from the accepting run ρi by converting all propositions of corresponding transition conditions δ(ρi(t+1),ρi(t)) of the accepting run ρi to their respective disjunctive normal form and successively adding all propositions of each one conjunctive clause to the essential sequence σe for all t.


In the following step 1340, the method generates a complementary sequences {circumflex over (σ)}e of the essential sequence σe. To this end partial sequences σ1 and σ2 are generated with σ1 being the part of essential sequence σe from its initial state to state qi and σ2 being the remaining part of essential sequence σe from state qi to its final state, i.e., essential sequence σe1σ2 is a concatenation of these two partial sequences σ1 and σ2. The complementary sequence {circumflex over (σ)}e is then generated by reversing the order of these two partial sequences σ1 and σ2, i.e., {circumflex over (σ)}e2σ1.


Next follows step 1350, in which it is checked whether or not the complementary sequence {circumflex over (σ)}e corresponds to an accepting run (which amounts to a simple iterative check whether the propositions of the complementary sequence {circumflex over (σ)}e satisfy the transition conditions δ). If it is determined that the complementary sequence {circumflex over (σ)}e is an accepting sequence, the method branches to step 1360 in which state qi is added to the decomposition set custom-character, after which the method continues with the execution of step 1370. If it is determined that the complementary sequence {circumflex over (σ)}e is not an accepting sequence, the method skips directly to step 1370.


In step 1370, it is checked whether index i has already iterated over all states qi of set Q, preferably by checking whether i=∥Q∥. If not, the method branches to step 1380 in which index i is incremented by an increment of 1 and the method continues with a next iteration in step 1320. If, however, it is determined that the index i has already iterated over all states qi of set Q, the method branches to step 1390, in which this part of the algorithm for determining the decomposition set custom-character ends.



FIG. 4 shows a flow-chart diagram that describes a preferred embodiment of a still further aspect of the present invention. This aspect relates a method for constructing the team model custom-character from the nondeterministic finite automaton custom-character, the decomposition set custom-character and the agent models custom-character1,custom-character2,custom-character3 . . . .


The method starts in step 1410, in which the nondeterministic finite automaton custom-character, the decomposition set custom-character and the agent models custom-character1,custom-character2,custom-character3, . . . , custom-charactern. are received. For ease of notation, in the context of the discussion of FIG. 4 the models will be labelled with a generic labelling superscript (r), i.e the agent models will be denoted custom-character(r).


Next, in step 1420, a corresponding product model P(r) will be created for every agent model custom-character(r). By combining the agent model automaton custom-character(r) with the nondeterministic finite automaton custom-character of the mission custom-character, the product model custom-character(r) can be constructed to capture both the agent capabilities encoded in the agent model automaton custom-character(r) and the specification of the mission custom-character encoded in the nondeterministic finite automaton custom-character. Dropping the superscript (r) for ease of notation, conveniently, the product model custom-character may be given by custom-character=custom-charactercustom-character=(custom-character,custom-character, custom-character) comprising

    • a set of states custom-character=Q×custom-character
    • a set of initial states custom-character=Q0×{custom-character}
    • a set of actions Ap={((qs,ss),(qt,st))∈custom-character×custom-character:(ss,st)∈custom-characterΛλ(ss)custom-characterδ(qs, qt))}.


For a plurality of agents, especially a plurality of robots, the respective agent models may differ from each other, each representing the capabilities of the respective agent, while the nondeterministic finite automaton custom-character Y is determined by a particular specification of the mission custom-character. As such, the product model custom-character may be constructed separately for each of the agents. It describes for each of the different agent how the mission custom-character can be executed by the agent to which the agent model custom-character(r) corresponds.


Therefore, in a preferred embodiment for each r∈{1, . . . , N} the corresponding product model custom-character(r) is constructed as custom-character(r)=custom-charactercustom-character(r) as defined above.


In order to combine a plurality of agents it is possible to construct a team model automaton custom-character from the individual product models custom-character(r). This is done in step 1430.


The team model automaton custom-character is conveniently constructed as a union of all the local product models custom-character(r) with r∈{1, . . . , N} as follows: The team model automaton custom-character is constructed as custom-character=(custom-character,custom-character,custom-character,custom-character), comprising

    • a set of states custom-character={(r,q,s):r∈{1, . . . , N}, (q,s)∈custom-character}
    • a set of initial states custom-character={(r,q,s)∈custom-character:r=1,(q,s)∈custom-character}
    • a set of final states custom-character={(r,q,s)∈Scustom-character:q∈F}
    • a set of actions custom-character=∪r custom-character.


In following step 1440 a set of switch transitions custom-charactercustom-character×custom-characteris determined. The set of switch transitions custom-character is defined as the set of all those transitions custom-character=((rs,qs,ss),(rt,qt,st)) between a starting state (rs,qs,ss)∈custom-character and a terminal state (rt, qt,st)∈custom-character which

    • connect different agents, i.e. rs≠rt,
    • preserve the progress in the nondeterministic finite automaton custom-character, i.e. Rs=qt,
    • point to the next agent, i.e. rt=rs+1
    • point to an initial agent state, i.e. st=custom-character, and
    • start in the decomposition set custom-character, i.e. qs custom-character.


Conveniently, the set of switch transitions custom-character may be constructed by traversing all states qs in the decomposition set custom-character and all starting agent indices rs={1, . . . , N−1}. For this choice of state qs and starting agent index rs, traversing all states ss for which (qs,ss)∈custom-character fixes rt,qt,st and thus yields the set of switch transitions custom-character.


An example of the structure of the team model custom-character is depicted in FIG. 6, which shows an example of a system comprising three agents. The team model custom-character has an initial state (bottom left corner) and three final states (right side). Between the agent automata, directed switch transitions custom-character to the next agent connect states of the decomposition set custom-character.


In step 1450 following step 1440, the set of switch transitions custom-character is added to the set of actions custom-character, i.e. custom-charactercustom-charactercustom-character. This concludes the algorithm shown in FIG. 4. FIG. 5 shows a flow-chart diagram that describes a preferred embodiment of an even further aspect of the invention. This even further aspect of the invention relates to derive an action sequence βfin which minimize a team cost κ for given agent models custom-character(r) of the team of agents 11, 12, 13, a cost function C, initial resources γ0≥0 and the specification ϕ of the mission custom-character such that the specification ϕ is satisfied. An action sequence β is called satisfying if the associated state sequence σ satisfies the specification ϕ.


Generally, an action sequence β is preferably defined as β=s0a1s1 . . . ansn which is a run in custom-character with sicustom-character and aicustom-character. In order to distribute β among the involved agents, custom-character for agent r is preferably obtained by projecting β onto custom-character(r).


Conveniently, the cost function C may be defined as follows. Each action of the team model custom-character is assigned a non-negative cost, i.e. C:custom-charactercustom-character≥0. For switch transitions custom-character, preferably the associated cost C(custom-character) is chosen as zero to reflect the fact that switch transitions custom-character are purely virtual and will not appear in the action sequence β(r) executed by the agents 11, 12, 13.


For modelling the multi-agent character of a cost, it is convenient to extend the cost C(a) associated with an action a∈custom-character to a vector of the same dimensionality N as the number of agents 11, 12, 13, i.e. C(a)∈custom-character≥0N where each agent r=1, . . . , N represents one dimension.


To reflect the fact that each action a with non-zero cost ca=C(a) is associated with a particular agent by the fact that custom-character\custom-character=∪rcustom-character, it is convenient to define







c

a
,
i


=

{





C


(
a
)


,


if





i

=
r








0
,
otherwise














and custom-character=0. Consequently, the costs cβ associated with an action sequence β can be computed as cβa∈βca.


Given a set of action sequences, a Pareto front of all cost vectors cβ for satisfying action sequences β then forms a set of potentially optimal solutions. In order to prioritize these solution, in a preferred embodiment one may compute an overall team cost κ as κ(cβ)=(1−ε)∥cβ+Π∥cβ1, where ε∈(0,11 may be chosen fixed but freely. This conventiently reflects an objective to minimize the maximal agent cost ∥cβ, e.g. minimizing a completion time of mission custom-character, and an objective to avoid unnecessary actions of the agents 11, 1213 via a regularization term ∥cβ1.


To save memory requirements for storing the cost vector cβ, preferably the cost vector cβ is stored as a compressed cost vector ĉβ which is three-dimensional, independent of the number of agets, by recursively choosing











c
^

β

=


(




||


(


c

β
,
1


,

,

c

β
,

r
-
1




)

T



||








||


(


c

β
,
1


,

,

c

β
,

r
-
1




)

T



||
1







c

β
,
r





)

.





(
1
)







This definition exploits the mathematical truth discovered as part of the work leading to the invention that given a fixed but arbitrary agent r, the team cost κ of the action sequence β can already be avaluated for all agents r′<r since no action associated with any of these agents r′ will occur in a continuation of β.


This makes it possible to simplify the computation of the team cost κ by instead computing a compressed team cost





{circumflex over (κ)}(ĉβ)(1−ε)∥(ĉβ,1β,3)T+ε∥(ĉβ,2β,3)T  (2)


with ĉβ,i denoting the i-th component of the compressed cost vector ĉβ. This representation not only removes a dependency of the team cost cβ on the team size N, it also a more efficient representation during planning. The reason for this efficiency gain is that additional cost vectors are Pareto-dominated as will be discussed below in the discussion of step 2100, and can thus be eliminated from the set of potential solutions much earlier in the planning process.


Furthermore, in addition to the specification ϕ which allows to model discrete constraints, in an optional further development is possible to consider constraints of the agents in continuous domains, like for example constraints on resources γ. A change of resources γ may be modeled by a resource function Γ:custom-charactercustom-characterM where M indicates the number of resource dimensions that models the change of resources γ under a given action a∈custom-character. Conveniently, the resource function can take both negative and positive values to reflect the fact that resources can be modified in both directions.


For the action sequence β, the resulting status of resources γβ is given by γβ0a∈βΓ(a) The set of satisfying action sequences is constrained to sequences β=s0a1s1 . . . ansn such that at any state sx∈β and a truncation β′ of sequence β until this state sx, i.e. β′=s0a1 . . . axsx it holds that γβ′,u>0 for each component i=1, . . . , M. In other words, the action sequences β are constrained such that the inequality constraint of the resources γβ holds at any time during the execution of the action sequence β.


Note that it is also possible to express constraints of the from γβ,i≥0 within this framework by choosing a fixed offset ξ smaller than the minimal change γΔ,i. of the resource component γβ,i under an exchange of any one action ajcustom-character for any other action akcustom-character, i.e. γΔ,i=min(aj,ak)|Γ(aj)i−Γ(ak)i|. The constraint γβ,i≥0 can then be modeled as an equivalent inequality constraint γβ,i+ξ>0.


While it would be possible to capture interval constraints of the form γβ,i ∈I=(Il,Iu) by a set of two inequality constraints, a more preferred solution that introduces a smaller number of Pareto optimal labels as explained below is to remodel the interval constraint as








γ

β
,
I


-



I
u

-

I
l


2


>
0




where







γ

β
,
I


=

||




I
u

-

I
l


2

+

I
l

-

γ

β
,
i



||





denotes a distance measure of γβ,i from the center of the interval I.


The actual algorithm for the planning problem discussed above is based on a label-setting approach which can be thought of as a multi-criteria generalization of the Dijkstra shortest path search. Instead of operating on states with associated costs, the label-setting algorithm constructs a set of labels for each state. For each state s∈custom-character, a label l will be given as l=(ĉββ,v,iv) which depends on the action sequence β that led to state s, ĉβ is the associated compressed cost and γβ the associated resource status, v∈custom-character is the state that precedes state s in action sequence β and iv is the respective predecessor label.


In other words, the construction of such a multi-dimensional label l fore each state s is an extension of the team-model state space custom-character to a higher-dimensional, infinitely large label space custom-character, in which each label l∈custom-charactercustom-character of state s instantiates one possible continuous resource configuration γ and transitions between the labels are described by their predecessor relations. custom-character denotes the set of instantiated, i.e., feasible, labels at state s and custom-character=custom-charactercustom-charactercustom-characterdenotes the set of all feasible labels.


It is possible to model a resource constraint as a proposition πi, e.g., πi:=(γβ,i>0). Whether or not πi is true would, in the state space custom-character, depend on a full action sequence β. However, in label space custom-character, πi is either true or false for each element of the label space custom-character since it is possible to associate a single label l∈custom-character with a specific γl,iβ,i as its second component. In a preferred embodiment, the resource constraints are indeed modeled in this way and denote the corresponding set of resource constraint propositions with Πγ.


The actual algorithm which is illustrated in FIG. 5 starts with an initialization in step 2010. A set of temporary labels Lt,v is initialized as Lt,v={0,γ0, Ø, Ø} for each initial state v∈custom-character. For each other state s∈custom-character\custom-character, a set of temporary labels Lt,s is initialized as Lt,s=Ø. Furthermore, for each state s∈custom-character a set of permanent labels Lp,s is initialized as Lp,s=Ø.


In the following step 2020, it is checked whether the set of temporary labels Lt,s is empty for each state s. If this is the case, no final state f is reachable and the algorithm stops with an error indication in step 2021, which may result in a controlling agents 11, 12, 13 accordingly, e.g. by transitioning the control system 10 into a safe state.


If, however, it is determined that the set of temporary labels Lt,s is not empty for at least one state s, the method proceeds to step 2030. In step 2030, the compressed cost vector compressed cost vector ĉβ is computed according to equation (1) and the compressed team cost {circumflex over (κ)}(ĉβ(l)) is computed according to equation (2). This is possible since each label l specifies its predecessor label, and the action sequence β leading to label l can be reconstructed. Then, a minimizing state s* and a minimizing label l* from the corresponding set of temporary labels Lt,s* is determined such that they minimize the compressed team cost {circumflex over (κ)}, i.e.







(


s
*

,

l
*


)

=


argmin


s


S



,

l


L

t
,
s









κ
^



(


c
^

β

(
l
)


)


.






In the next step 2040, the minimizing label l* is removed from the set of temporary labels Lt,s, corresponding to the minimizing state s*, i.e. Lt,s*←Lt,s*\{l*} and added to the corresponding set of permanent labels Lp,s*, i.e. Lp,s*←Lp,s*∪, {l*}.


In the following step 2050, it is checked whether the minimizing state s* is an accepting state. If this is the case, the method continues with step 2060, if not, it continues with step 2080.


In step 2060, a final label lfin is set to label l*. As outlined above, the corresponding final action sequence βfin is reconstructed iteratively from the predecessor labels. The final action sequence βfin is the selected action sequence β with minimal compressed team costs {circumflex over (κ)} and hence minimal team costs κ. This concludes the algorithm.


In step 2080, a set V of all neighboring states v of minimizing state s*and a corresponding set Lv of corresponding neighboring labels is determined. For example, V may be determined by intitializing V=Ø, Lv=Ø,exploring each state v=(rv,qv,sv)∈custom-character and adding state v to set V if and only if there is an action a that links the minimizing state s* to state V, i.e. a=(s*,v)∈custom-character. If state v∈custom-character is added to set V, the corresponding new costs ĉnew are computed depending on action a via








c
^

new

=

{





(





||


c
^

1

(
l
)



,



c
^

3

(
l
)




||










||


c
^

2

(
l
)



,



c
^

3

(
l
)




||
1







0



)





if





a


ζ








c
^


(
l
)


+


(

0
,
0
,

C


(
a
)



)

T




otherwise



.






Similarly, corresponding new resources γnew are computed depending on action a via







γ
new

=

{





(




γ
global

(
l
)







γ

0
,

r
v






)





if





a


ζ







γ

(
l
)


+

Γ


(
a
)





otherwise



.






In this formula, γglobal(l) denotes the part of the resources γ(l) that is global, i.e. independent of the agent, and γ0,rv denote the initial resources of agent rv. A corresponding new label lv=(ĉnew, γnew, s*, is*) is generated, with is*=card(Lp,s*). This corresponding new label lv is then added to the set of neighboring labels Lv. After exploration of all states v is completed, the method continues with step 2090.


In the next step 2090, it is checked it is checked for each neighboring label l∈Lv whether the corresponding new resource status γnew satisfies all constraints. For this purpose, an extended transition function Δ:custom-character×custom-characterM→{custom-character,⊥} which is an extension of the transition function δ of the nondeterministic finite automaton custom-character is defined as Δ:(a=((rs, qs, ss), (rt, qt, st)), γ)custom-character(λ(ss)∪Πγ)custom-characterδ(qs, qt). The action al associated with neighboring label l is determined and it is checked whether Δ(alnew) is true. If it is not true, the method branches back to step 2020. If it is true, however, it is also checked whether the neighboring label l is non-dominated in the Pareto sense.


For ease of notation, an operator <P denotes a “less than”-relation in the Pareto sense, i.e. (a1, . . . , an)T<P(b1, . . . , bn)T⇔a≠bΛai≤bi∀i∈{1, . . . , n}. An operator ≤P relaxes this relation and also allows a=b. A label is non-dominated in the Pareto sense if there does not exist another label custom-character in either the set of temporary labels Lt,s or the set of permanent label Lp,s at the same state v such that (ĉ(custom-character),−γ(custom-character))≤P(l),−γ(l)).


If it is found that no such label custom-character exists, it is deemed that the neighboring label l is non-dominated in the Pareto sense and the method continues with step 2100. If, however, such a label custom-character exists, the method skips back to step 2020.


In step 2100, all labels custom-character which are dominated by any neighboring label l∈Lv where said neighboring label found to satisfy all constraints and be non-dominated in the Pareto sense by another label are removed from the set of temporary labels Lt,v at the same state v, i.e., Lt,v←Lt,v\{custom-character∈Lt,v:l<Pcustom-character}.


Next, in step 2110, all said aforementioned neighboring labels l are added to the set or temporary labels Lt,v, i.t. Lt,v←Lt,v ∪{l}. The method then continues with step 2020.

Claims
  • 1. A method for controlling a plurality of agents to complete a mission, comprising: deriving a decomposition set of decomposition states in a set of possible states of an automaton, wherein the automaton characterizes the mission; andderiving a sequence of actions to be e carried out by the plurality of agents depending on the decomposition set, where each of the actions is to be carried out by at most one of the plurality of agents.
  • 2. The method according to claim 1, further comprising: controlling the plurality of agents in accordance with the derived sequence of actions.
  • 3. The method according to claim 1, further comprising: generating the decomposition set by exploring an essential sequence of an accepting run through one or more candidate decomposition states.
  • 4. The method according to claim 3, further comprising: adding the one or more candidate decomposition state to the decomposition set depending on whether a complementary sequence to the explored essential sequence around the respective one or more candidate decomposition state is accepting.
  • 5. The method according to claim 4, wherein in which the decomposition set includes all those states in the set of possible states of the automaton, for which the complementary sequence to the explored essential sequence around the respective state is accepting.
  • 6. The method according to claim 1, further comprising: generating a team model based on the automaton that characterizes the mission and based on automata that each characterize the capabilities of one of the plurality of agents.
  • 7. The method according to claim 6, wherein the team model comprises a set of actionsthat comprises switch transitions which change the acting agent from one of the plurality of agents to another one of the plurality of agents.
  • 8. The method according to claim 7, wherein the switch transitions are configured to each change the acting agent from one of the plurality of agents to a next one of the plurality of agents.
  • 9. The method according to claim 8, wherein the switch transitions are configured such as to act only if the automaton that characterizes the mission is in a decomposition state.
  • 10. The method according to claim 6, further comprising: deriving the sequence of actions to be carried out by the plurality of agents by a label-setting algorithm in which each state of a set of states of the team model is associated with labels that are characterized by a sequence of action leading to the respective state.
  • 11. The method according to claim 10, further comprising: constructing a reachable set of temporary labels for each state and a set of permanent labels.
  • 12. The method according to claim 11, further comprising: constructing, for each selected label, a set of consecutive labels by extending an action sequence associated with the selected label by all available actions and adding the resulting labels to the reachable set of temporary labels.
  • 13. The method according to claim 12, wherein each label comprises at least one component that characterizes a cost under the corresponding sequence of actions.
  • 14. The method according to claim 13, wherein the derived sequence of actions to be carried out by the plurality of agents is the one out of all actions that satisfy a characterization of the mission that minimizes a team cost which depends on the component that characterizes the cost.
  • 15. The method according to claim 14, wherein only actions resulting in Pareto-optimal labels at their target state are added to the reachable set of temporary labels.
  • 16. The method according to claim 13, wherein the component that characterizes the cost under the corresponding sequence of actions is depending on costs associated with each of these actions with one component each for each one of the agents.
  • 17. The method according to claim 16, wherein the component that characterizes the cost under the corresponding sequence of actions is stored in memory by way of a data structure that comprises at least one component that characterizes costs associated with a selected one of the agents and at least one component that characterizes the costs associated with a group of agents that precede the selected one of the agents.
  • 18. The method according to claim 17, wherein each label comprises at least one component that characterizes a resource status at the respective state under the corresponding sequence of actions.
  • 19. The method according to claim 18, wherein the characterization of the mission comprises an inequality constraint that restricts the at least one component that characterizes a resource status to a predefined region.
  • 20. A non-transitory machine-readable storage medium on which is stored a computer program for controlling a plurality of agents to complete a mission, the computer program, when executed by a computer, causing the computer to perform: deriving a decomposition set of decomposition states in a set of possible states of an automaton, wherein the automaton characterizes the mission;deriving a sequence of actions to be e carried out by the plurality of agents depending on the decomposition set, where each of the actions is to be carried out by at most one of the plurality of agents; andcontrolling the plurality of agents in accordance with the derived sequence of actions.
  • 21. A system for controlling a plurality of agents to complete a mission, which is configured to: derive a decomposition set of decomposition states in a set of possible states of an automaton, wherein the automaton characterizes the mission;derive a sequence of actions to be e carried out by the plurality of agents depending on the decomposition set, where each of the actions is to be carried out by at most one of the plurality of agents; andcontrol the plurality of agents in accordance with the derived sequence of actions.
  • 22. The system according to claim 21, further comprising at least one of the plurality of agents.
Priority Claims (1)
Number Date Country Kind
102017215311.3 Sep 2017 DE national