GB2606752A

GB2606752A - Robot fleet management method and system using a graph neural network

Info

Publication number: GB2606752A
Application number: GB2107204.6A
Authority: GB
Inventors: Colin Hoy Michael; Panagiotis Geromichalos Dimitrios; Singh Rahul
Original assignee: Continental Automotive GmbH
Current assignee: Continental Automotive GmbH
Priority date: 2021-05-20
Filing date: 2021-05-20
Publication date: 2022-11-23
Also published as: WO2022242966A1; GB202107204D0; DE112022002704T5

Abstract

A robot management system 10 for controlling a fleet of robots in an environment 14. The environment 14 includes a plurality of locations 16 represented by nodes 22. The nodes are connected by at least one route, represented by connecting edges 24. The robot management system utilizes a graph neural network (GNN) having a main graph auto encoder network which deals with the system state as a whole and further comprises an auxiliary graph auto encoder network for each route constraint. Executing the GNN results in the generation of a single-step command 50 for each robot 12. The commands are transmitted to each robot 12 causing the robot(s) to move about an environment 14. The robot management system may use the GNN to find an optimal route and/or task allocation based upon the route constraints. When multiple routes are available for each robot then each route may be scored based on route properties, such as the time taken to complete each route, with the optimal route generally being the route with the shortest journey time.

Description

DESCRIPTION

Robot fleet management method and system using a graph neural network

TECHNICAL FIELD

The invention relates to a computer implemented robot management method and a robot management system. The invention further relates to a graph neural network adapted for use in said method and system.

BACKGROUND

Fleet management systems usually incorporate functions like task allocation, constraint satisfaction and cooperative path planning, and are in widespread use across many logistic industries. These are traditionally solved using a combination of meta-heuristic search algorithms (e.g., Tabu search, simulated annealing, deconfliction based planning) and graph search algorithms (e.g., A* planning). For complicated, multifaceted problems, additional engineering work is typically needed to develop and tune application specific metaheuristics algorithms.

Another idea is the use of reinforcement learning to train a neural network to learn policies that are directly optimized for each problem type. For example, a solution is to use multi-agent RL, and have each robot independently learn a navigation policy that causes optimized cooperative behavior to emerge.

Steve Paul et al., "Learning to Solve Multi-Robot Task Allocation with a Covariant-Attention based Neural Architecture", ICLR 2021, discloses a graph neural network architecture over which reinforcement learning can be performed to obtain online policies for multi-robot task allocation problems.

US 10 474 141 B2 discloses a method and an apparatus for controlling movement of transporting devices. Transport containers are stored in stacks arranged in a facility having pathways arranged in a grid-like structure above the stacks. The transporting devices are configured to operate on the grid-like structure.

US 10 089 586 B2 discloses a job management system for processing job requests in an automated physical environment, such as a factory, hospital, order processing facility or office building, wherein the job requests are handled by a fleet of autonomously -navigating mobile robots.

SUMMARY OF THE INVENTION

It is the object of the invention to improve robot fleet management, especially in highly detailed environment maps.

The invention provides a computer implemented robot management method for operating a robot management system that controls at least one robot in an environment, the method comprising: a) providing an input map of the environment, the input map having a plurality of nodes and at least one connecting edge that connects two nodes, each node indicating a location that the robot is able to visit within the environment, each connecting edge indicating a route that the robot is able to take between connected nodes through the environment; b) collating the current state of the robot management system, the current state of the robot management system including at least one node state of each node, at least one edge state of each edge, and at least one robot state of each robot; c) compiling a list of route constraints, wherein the route constraints relate to information indicative for properties of a route that the robot is able to take to a node; d) initializing a graph neural network system, wherein the graph neural network system is configured for determining one single-step command for each robot and the graph neural network system comprises a main graph autoencoder network and at least one auxiliary graph autoencoder network for each route constraint of the list of route constraints, the main graph autoencoder network being initialized based on the current state of the robot management system obtained in step b), wherein each separate auxiliary graph autoencoder is initialized based on the respective route constraint; e) executing the graph neural network system, so as to determine a single-step command for each robot, the single-step command being configured to control a next action of the corresponding robot; and f) transmitting each single-step command obtained in step e) to the corresponding robot, in order to cause the robot to perform its next action according to the single-step command.

Preferably, step c), step d) or step e) are performed multiple times. Preferably, steps c) and d), steps c) and e), or steps d) and e) are performed multiple times. Preferably, steps c), d), and e) are performed multiple times. Preferably, in case of multiple performance of the steps, an intermediate state is propagated from one iteration to the next iteration. This intermediate state is handled similarly within the message passing graph neural network.

Preferably, before step e), a hard attention mask is generated and step e) is performed only on a subgraph of the graph neural network system, wherein the subgraph is determined by the hard attention mask.

Preferably, the hard attention mask is generated by randomly selecting one of the route constraints; determining a subgraph of the graph neural network system to which the route constraint relates; updating the internal neural state of the graph neural network system and/or updating the single-step command at each node for the robot.

Preferably, the subgraph is determined such that the deviation of the subgraph from the shortest route is below a predetermined threshold.

Preferably, it is randomly selected, whether in step e), the whole graph neural network system or a portion of the graph neural network system is executed.

Preferably, in step e), each single-step command is determined after the graph neural network system was iteratively executed for a predetermined number of iterations.

Preferably, the graph neural network system is configured as a message passing graph neural network system. Preferably, the main graph autoencoder network includes a first fully connected neural network for each node Preferably, each auxiliary graph autoencoder network includes a second fully connected neural network that is associated with a particular route constraint.

Preferably, in step e) the node state is passed to the first and second fully connected networks and the route constraint is only passed to the associated second fully connected neural network.

Preferably, the main graph autoencoder network includes a first aggregator for each node, and the output of any, some, or all of the first and second fully connected neural networks is passed to the first aggregator.

Preferably, the graph neural network system comprises a second aggregator for each node that is associated with that node, wherein the second aggregator of any node receives the output of any, some, or all of the first aggregators of the unassociated other nodes.

Preferably, the main graph autoencoder network includes a third fully connected neural network for each node.

Preferably, each auxiliary graph autoencoder includes a fourth fully connected neural network associated with a particular route constraint.

Preferably, for each node the third fully connected network receives the output of the first aggregator and/or the second aggregator that is associated with that node. Preferably, for each node each fourth fully connected neural network receives the output of the first aggregator and/or the second aggregator that is associated with that node.

Preferably, the main graph autoencoder network includes a third aggregator for each node, wherein for each node the output of any, some, or all of the third and fourth fully connected neural networks is passed to the third aggregator of that node, and this output is passed on as output of the graph neural network system.

Preferably, in addition the output of any, some, or all of the fourth fully connected networks of that node may be passed on as additional output.

Preferably the node state includes a pick-up indicator, that indicates that a package is to be picked up from that node. Preferably, the node state includes a drop-off indicator, that indicates that a package is to be dropped off at that node. Preferably, the node state includes a battery charger indicator, that indicates that a battery charger is available at that node. Preferably, the note state includes a capacity indicator, that indicates whether or how much robots or packages that node can still accommodate.

Preferably, the edge state includes an edge capacity indicator, that indicates whether or how much robots are able to travel along that edge, preferably without delay or congestion. Preferably, the edge state includes a distance indicator, that is indicative of the physical distance between the two nodes connected by that edge.

Preferably, the robot state includes a carrying capacity indicator, that is indicative whether or how much freight that robot is able to carry. Preferably, the robot state includes a battery status indicator, that is indicative of the charge status of the robot batter, in particular whether the robot needs to go to a charging station, i.e., a node with a battery charger.

The invention further provides a robot management system for controlling a plurality of robots in an environment, the environment comprising a plurality of locations that are connected by at least one route, the robot management system comprising at least one robot that is configured to received commands generated by a computer system, and a computer system that is configured to perform one, some, or all of the steps of the method previously described.

The invention further provides a computer program, a computer readable storage medium or a data carrier signal comprising instructions which, upon execution by a computer system, cause the computer system to carry out one, some or all of the steps of the previously described method.

The invention further provides a computer implemented graph neural network system configured for controlling robots in an environment, the environment comprising a plurality of locations represented by nodes that are connected by at least one route that is represented by at least one connecting edge, the environment further comprising route constraints relating to information indicative for properties of a route that the robot is able to take to a node, the graph neural network system comprising a main graph autoencoder network for each node characterized by an auxiliary graph autoencoder network for each route constraint.

Preferably, the graph neural network system is configured such that the node state is passable to the first and second fully connected networks and the route constraint is only passable to the associated second fully connected neural network.

Preferably, the main graph autoencoder network includes a first aggregator for each node, and the output of any, some, or all of the first and second fully connected neural networks is passable to the first aggregator.

Preferably, the graph neural network system comprises a second aggregator for each node that is associated with that node, wherein the second aggregator of any node is configured to receive the output of any, some, or all of the first aggregators of the unassociated other nodes.

Preferably, the third fully connected network is configured to, for each node, receive the output of the first aggregator and/or the second aggregator that is associated with that node. Preferably, each fourth fully connected neural network is configured to, for each node, receive the output of the first aggregator and/or the second aggregator that is associated with that node.

Preferably, the main graph autoencoder network includes a third aggregator for each node, wherein for each node the output of any, some, or all of the third and fourth fully connected neural networks is passable to the third aggregator of that node, and this output is able to be passed on as output of the graph neural network system.

Preferably, in addition the output of any, some, or all of the fourth fully connected networks of that node are passable on as additional output.

The invention aims to coordinate robots on highly detailed world maps (described as graphs with many nodes and edges) or environments. For such applications, there is only little work applying reinforcement learning with neural networks.

Traditional search algorithms either have long computation times (e.g., for general search algorithms), or require much engineering effort (e.g., for hand-tuned heuristics). Multi-agent RL approaches require a separate neural network instance for each agent, and thus have a limited capacity for each individual network to accept input information. For example, for large graphs, it would be computationally infeasible for every agent to analyze the entire graph individually. The solution to this is only to provide coarse, high level information to the neural network inputs, which makes certain types of applications difficult.

The instant invention proposes a new graph neural network architecture for solving arbitrary task allocation, constraint satisfaction and cooperative routing problem(s). By designing a novel network structure, we can accept an arbitrary and variable number of route constraints as input. This allows us to use a single neural network comprised of subnetworks to make decisions for the entire fleet, which has better computational scalability.

The solution described has in general the following steps.

Initially the current state of the system is collated. This includes the state of each graph node, which may include information about the state of robots currently positioned at each graph nodes. Some input data may be computed for each node, such as, whether there are pickups or drop-offs at the node, whether there is a battery charger at the node, what the robot carrying capacity is (if there is a robot on the node), or what the battery status is Of there is a robot on the node).

Furthermore, a list of route constraints is compiled. Route constraints are typically relating to a particular route. For instance, a route constraint can define where the robot goes next, in order to arrive at its goal node. As an example, two robots that have picked up items, each require route constraints to ensure that the items are routed to the correct locations.

Not all constraints, however, are route constraints. Those are referred to as non-route constraints. For example, if a robot requires use of a charger, it is not relevant which charger is used; in general, any charger can be used.

Route constraints that are calculated as input date for each node include, for example, the distance to the robot, the distance to the goal node, the scoring of the route (if multiple routes are available). Scoring can be done based on any property of the route. Mostly, this will be the time the robot takes for completing the route, including possible detours due to congestion, for example.

The system may have additional constraints, such as the robots have to be in a charged state.

After all the system information about the nodes, edges, and robots is compiled, the state information is presented to the input of a graph autoencoder network. For each route constraint, a separate auxiliary graph autoencoder to initialized.

Now, the initialized graph autoencoder network is executed. Basically, any graph neural network structure can be used, although the description will focus on an example of a message passing graph neural network.

The graph autoencoder network was trained for the particular deployment environment or world map using reinforcement learning. Since, in general the underlying system model is known, the network may be trained in a simulator using a range of techniques such as 0-learning, stochastic value gradients, or AlphaGo style policy improvement.

During execution the graph autoencoder network generates/determines a single-step command for each node and robot. The single-step commands are then sent to the corresponding robots.

The steps so far can be referred to "node level attention", since each node, robot and route attribute is considered and as a result, each node and robot are given a result from the graph autoencoder network.

In addition to these steps, one may make use of an additional network mechanism which is referred to as "route level attention".

The steps described above have run the network once so far, however, here the network may run multiple times, each time potentially only computing a portion or subgraph of the neural network, e.g., by using a hard attention mask.

In order to generate the attention mask, the invention proposes to randomly select one of the route constraints. Next, the subgraph "near" the shortest route is isolated. Thus, a small deviation between the isolated subgraph and the shortest route is allowed Then, the graph neural network is executed on the isolated subgraph and the output of the graph neural network is updated, which can include values for the internal neural state and/or interim actions for the robot at each node.

For each iteration, it can be randomly decided whether to execute the neural network over the whole graph or only over a portion of the graph, i.e., a subgraph, as defined above.

Once some number of predetermined iterations are complete, the actions for the robot at each node can be finalized.

BRIEF DESCRIPTION OF THE DRAWINGS

Embodiments of the invention are described in more detail with reference to the accompanying schematic drawings.

Fig. 1 depicts an environment in which robots are navigating; Fig. 2 depicts an autoencoder; Fig. 3 depicts an updated environment; and Fig. 4 depicts a block diagram of a graph neural network.

DETAILED DESCRIPTION OF EMBODIMENT

Fig. 1 schematically depicts a robot management system 10. The robot management system 10 is configured to control a plurality of robots 12 in an environment 14. The environment 14 comprises a plurality of locations 16. The robots 12 are capable of travelling between two locations 16 via a path 18.

The environment 14 is abstracted into a machine-readable format. The environment 14 is represented as a route graph 20 that includes a plurality of nodes 22 and at least one connecting edge 24. Each node 22 represents one of the physical locations 16 in the physical environment 14. Each connecting edge 24 connects two nodes 22 and represents a physical path 18, that one of the robots 12 can take between two physical locations 16.

It should be noted that while the connecting edge 24 is represented as a straight line, the corresponding physical path 18 need not be straight and can further include curves or ramps or other non-straight elements.

Each robot 12 has associated with it a robot state. The robot state contains information indicative of the respective operational state of the robot 12. The robot state can include information about the current battery charge, remaining battery charge, estimated charge consumption, the carrying capacity, the current robot position (in physical space and/or in graph space), and any other property that describes the operational situation of the robot 12.

Furthermore, each node 22 has associated with it a node state, which can include information about the location 16 that the corresponding node 22 represents. The information may include, whether a package 26 is to be picked up from the node 22, whether or which of the packages 26 is to be dropped off at the node 22, whether the node 22 has a charging capability, i.e. the location 16 has a charging dock 28, and whether the node 22 is capable of accepting a robot 12 (i.e. whether the location 16 has enough space to accommodate a robot 12).

Each of the connecting edges 24 has associated with it an edge state, which contains information relating to properties of the specific connecting edge 24, i.e. the path 18 which is represented by the connecting edge 24. This information may include the capacity of the connecting edge 24 (for example, how many robots 12 can occupy the connecting edge 24 at the same time), the distance between the two nodes 22 that are connected by the specific connecting edge 24, or other properties of the connecting edge 24 as well as the associated path 18.

Fig. 1 depicts an example for a current state of the robot management system 10. Therein, a first robot 30 should go to the charging dock 28. A second robot 32 and a third robot 34 have each picked up a package 26 and should deliver it.

The second robot 32 has picked up the package 26 from a first pickup 36 and the third robot 34 has picked a package 26 from a second pick up 38.

In this example, the first robot 30 and the third robot 34 are still at their respective starting nodes 40. The second robot 32 has already moved from its starting node 40 to an intermediate node 42. None of the robots 30, 32, 34 are at their respective goal nodes 44.

The first robot 30 has as its goal node 44 the charging dock 28. The second robot 32 has as its goal node 44 a first drop-off 46. The third robot 34 has as its goal node 44 a second drop-off 48.

The previously described configuration of the robot management system 10 is known as the system state and includes the various robot, node, and edge states.

Subsequently, an embodiment of a robot management method is described.

The robot management method operates the robot management system 10 such that each robot 12 receives a single-step command 50, which instructs the respective robot 12 what action it should perform next.

As an initial step, the system state (as exemplified in Fig. 1) is collated. As a next step, a list of route constraints is compiled. The route constraints relate to information that is indicative for properties of a route that each robot 12 is able to take towards its goal node 44. The route may be composed of one or more paths 18. The route constraints may include the total distance to travel to the goal node, the (estimated) time to arrival, the route capacity, or any other information that is relevant to the journey of the respective robot 12 along the route.

Referring now to Fig. 2, a graph neural network system 52 is initialized based on the current state of the robot management system 10 and the list of route constraints. As depicted in Fig. 2, the graph neural network system 52 comprises a main graph auto encoder network 54. The graph neural network system 52 further comprises a separate auxiliary graph auto encoder network 56 for each route constraint within the list of route constraints.

A plurality of first routes 58 that are associated with the third robot 34 initialize a first of the auxiliary graph auto encoder networks 56. The plurality of first routes 58 preferably include all possible routes for the third robot 34 from its current node 22 to its goal node 44, wherein the first routes 58 are selected such that they deviate from the shortest possible route by not more than a predetermined deviation. In this case, the third robot 34 can either go directly from the second pick up 38 to the second drop-off 48 or can go on a detour 60 via another node 22.

A second of the auxiliary graph auto encoder networks 56 is initialized based on a plurality of second routes 62, which are routes that the second robot 32 is able to take from its current node 22 to its goal node 44. The second routes 62 are selected in the same manner as the first routes 58 so that they do not deviate by more than a predetermined amount from the shortest route.

In the same manner, another auxiliary graph auto encoder network 56 may be initialized for the first robot 30 (not shown here).

The graph neural network system 52 has been trained by methods known per se in the art. The training on the environment 14 is preferably done by simulator and can use a variety of techniques, such as cue-learning, stochastic value gradients, or policy improvement.

After the graph neural network system 52 is initialized, it is executed and results in a set of single-step commands 50 that are associated with each node 22 and corresponding robot 12.

As depicted in Fig. 3, the first robot 30 receives an instruction to move north to the next node 22, i.e. to the node 22 that has the charging dock 28.

The second robot 32 has received an instruction to also go north to the next node 22 in order to avoid collision with the first robot 30 while still moving closer to the first drop-off 46. The third robot 34 is instructed to move north towards its goal node 44 which is the second drop-off 48.

For the purposes of illustration it is assumed that each node 22 and each connecting edge 24 can only accommodate a single robot 12 at a time. Here, the graph neural network system 52 has managed the movement of the robots 12 such that each robot 12 is closer to its goal node 44, while simultaneously avoiding that any connecting edge 24 and any node 22 is occupied by two robots 12 at the same time.

Referring to Fig. 4, an example of the configuration of the graph neural network system 52 is described in more detail.

The graph neural network system 52 is configured as message passing graph neural network system. Each node state 54 and each route constraint information 66 is input as initial information (far left in Fig. 4).

The main graph auto encoder network 54 comprises a first fully connected neural network 68 or first FCN 68 (hereinafter, "fully connected neural network" is referred to as "FCN"). Furthermore, each auxiliary graph auto encoder network 56 comprises a second FCN 70 for each route constraint information 66.

The node state 64 is passed to both the first and second FCNs 68, 70. The route constraint information 66 is only passed to the second FCN 70 of its corresponding auxiliary graph auto encoder network 56.

The main graph auto encoder network 54 further comprises a first aggregator 72 associated with each node state 64. The output of the first and second FCNs 68, 70 is passed on to the first aggregator 72. Furthermore, the output of the second FCN 70 is additionally passed on without aggregation.

The graph neural network system 52 further comprises a second aggregator 74 for each node state 64. The second aggregator 74 of a specific node 22 is connected to the first aggregator 72 of the other nodes 22.

The main graph auto encoder network 54 comprises a third FCN 76. The auxiliary graph auto encoder network 56 comprises a fourth FCN 78 for each route constraint information 66.

The third FCN 76 receives the output from the first and second aggregators 72, 74. The fourth FCN 78 receives the output of the second aggregator 74 and the output of their corresponding second FCN 70. The fourth FCN 78 of one route constraint information 66 does not receive the output from the second FCN 70 of any other route constraint information 66.

Finally, the main graph auto encoder network 54 comprises a third aggregator 80 that receives the output from the third and fourth FCNs 76, 78.

The graph neural network system 52 has a first output 82, which corresponds to the output of the third aggregator 80.

Furthermore, the graph neural network system 52 has a second output 84 for each route constraint information 66. The first output 82 corresponds to the output of the third aggregator 80 and the second output 84 corresponds to the output of the fourth FCN 78 of the corresponding route constraint information 66.

The graph neural network system 52 with this architecture is able to generate the single-step commands 50 for each robot 12 based on the system state and an arbitrary number of route constraints.

In order to improve scalability and manage an arbitrary number of route constraints, the invention provides a robot management system 10 that utilizes a graph neural network system 52. The graph neural network system 52 has a main graph auto encoder network 54, which deals with the system state as a whole and further comprises an auxiliary graph auto encoder network 56 for each route constraint of the robot management system 10, in order to generate commands 50 for the robots 12. The robots 12 move about an environment 14 based on the single-step commands 50.

REFERENCE SIGNS

robot management system 12 robot 14 environment 16 location 18 path route graph 22 node 24 connecting edge 26 package 28 charging dock first robot 32 second robot 34 third robot 36 first pickup 38 second pickup starting node 42 intermediate node 44 goal node 46 first drop-off 48 second drop-off single-step command 52 graph neural network system 54 main graph auto encoder network 56 auxiliary graph auto encoder network 58 first route detour 62 second route 64 node state 66 constraint info 68 first FCN second FCN 72 first aggregator 74 second aggregator 76 third FCN 78 fourth FCN third aggregator 82 first output 84 second output

Claims

CLAIMS1. A computer implemented robot management method for operating a robot management system (10) that controls at least one robot (12) in an environment (14), the method comprising: a) providing an input map of the environment (14), the input map having a plurality of nodes (22) and at least one connecting edge (24) that connects two nodes (22), each node (22) indicating a location (16) that the robot (12) is able to visit within the environment (14), each connecting edge (24) indicating a route that the robot (12) is able to take between connected nodes (22) through the environment (14); b) collating a current state of the robot management system (10), the current state of the robot management system (10) including at least one node state (64) of each node (22), at least one edge state of each edge (24), and at least one robot state of each robot (12); c) compiling a list of route constraints, wherein the route constraints relate to information indicative for properties of a route that the robot (12) is able to take to a node (22); d) initializing a graph neural network system (52), wherein the graph neural network system (52) is configured for determining one single-step command (50) for each robot (12) and the graph neural network system (52) comprises a main graph autoencoder network (54) and at least one auxiliary graph autoencoder network (56) for each route constraint of the list of route constraints, the main graph autoencoder network (54) being initialized based on the current state of the robot management system (10) obtained in step b), wherein each separate auxiliary graph autoencoder (56) is initialized based on the respective route constraint; e) executing the graph neural network system (52), so as to determine a single-step command (50) for each robot (12), the single-step command (50) being configured to control a next action of the corresponding robot (12); and f) transmitting each single-step command (50) obtained in step e) to the corresponding robot, in order to cause the robot (12) to perform the next action according to the single-step command (50).
2. The method according to claim 1, characterized in that any combination of the steps c), d) and e) is performed multiple times.
3. The method according to any of the preceding claims, characterized in that, before step e), a hard attention mask is generated, and step e) is performed only on a subgraph of the graph neural network system (52), wherein the subgraph is determined by the hard attention mask.
4. The method according to claim 3, characterized in that the hard attention mask is generated by randomly selecting one of the route constraints; determining a subgraph of the graph neural network system (52) to which the route constraint relates; updating the internal neural state of the graph neural network system (52) and/or updating the single-step command (50) at each node (22) for the robot (12).
5. The method according to claim 4, characterized in that the subgraph is determined such that the deviation of the subgraph from the shortest route is below a predetermined threshold.
6. The method according to any of the preceding claims, characterized in that it is randomly selected, whether in step e), the whole graph neural network system (52) or a portion of the graph neural network system (52) is executed.
7. The method according to any of the preceding claims, characterized in that, in step e), each single-step command (50) is determined after the graph neural network system (52) was iteratively executed for a predetermined number of iterations.
8. The method according to any of the preceding claims, characterized in that the graph neural network system (52) is configured as a message passing graph neural network system (52), wherein the main graph autoencoder network (54) includes a first fully connected neural network for each node (22), wherein each auxiliary graph autoencoder network (56) includes a second fully connected neural network that is associated with a particular route constraint, wherein in step e) the node (22) state is passed to the first and second fully connected networks and the route constraint is only passed to the associated second fully connected neural network.
9. The method according to claim 8, characterized in that the main graph autoencoder network includes a first aggregator (72) for each node (22), and the output of any, some, or all of the first and second fully connected neural networks is passed to the first aggregator (72).
10. The method according to claim 9, characterized in that the graph neural network system (52) comprises a second aggregator (74) for each node (22) that is associated with that node (22), wherein the second aggregator (74) of any node (22) receives the output of any, some, or all of the first aggregators (72) of the unassociated other nodes (22).
11. The method according to claim 10, characterized in that the main graph autoencoder network (54) includes a third fully connected neural network for each node (22), wherein each auxiliary graph autoencoder (56) includes a fourth fully connected neural network associated with a particular route constraint, wherein for each node (22) the third fully connected network receives the output of the first aggregator (72) and/or the second aggregator (74) that is associated with that node (22), wherein for each node (22) each fourth fully connected neural network receives the output of the first aggregator (72) and/or the second aggregator (74) that is associated with that node (22).
12. The method according to claim 11, characterized in that the main graph autoencoder network (54) includes a third aggregator (80) for each node (22), wherein for each node (22) the output of any, some, or all of the third and fourth fully connected neural networks is passed to the third aggregator (80) of that node (22), and this output is passed on as output of the graph neural network system (52).
13. A robot management system (10) for controlling a plurality of robots in an environment (14), the environment (14) comprising a plurality of locations (16) that are connected by at least one route, the robot management system (10) comprising at least one robot (12) that is configured to received commands generated by a computer system, characterized by a computer system that is configured to perform one, some, or all of the steps of the method according to any of the claims 1 to 12.
14. A computer program, a computer readable storage medium or a data carrier signal comprising instructions which, upon execution by a computer system, cause the computer system to carry out one, some or all of the steps of the method according to any of the claims 1 to 12.
15. A computer implemented graph neural network system (52) configured for controlling robots in an environment (14) , the environment (14) comprising a plurality of locations (16) represented by nodes (22) that are connected by at least one route that is represented by at least one connecting edge (24) , the environment (14) further comprising route constraints relating to information indicative for properties of a route that the robot (12) is able to take to a node (22), the graph neural network system (52) comprising a main graph autoencoder network (54) for each node characterized by an auxiliary graph autoencoder network (56) for each route constraint.