Method for controlling a robotic device using a conditionally reversible neural network model

Abstract

The various aspects involve methods (100, 500), including: controlling a robotic device based on corresponding initial input parameter values ​​of a plurality of input parameters (206), and determining corresponding initial output parameter values ​​of one or more output parameters (210); determining a plurality of Shapley values ​​by inputting the initial input parameter values ​​as inputs and the initial output parameter values ​​(one or more) as conditions into a conditionally reversible neural network model, wherein each Shapley represents the contribution of the corresponding input parameter to one or more output parameters; determining a target output parameter value for each output parameter; determining a plurality of target input parameter values ​​by inputting the plurality of Shapley values ​​conditioned on the target output parameter values ​​into the inverse path of the conditionally reversible neural network model; and controlling the robotic device based on the plurality of target input parameter values.

Description

Existing technology

[0001] For various technical (e.g., physical or chemical) processes, such as manufacturing processes and / or machining processes (e.g., drilling, milling, heat treatment, etc.), it may be desirable to determine input parameters (e.g., process temperature, process time, vacuum or gas atmosphere, etc.) for controlling the technical process based on given output parameters of the technical process in order to achieve given output parameters (e.g., one or more properties of the process workpiece, such as hardness, strength, thermal conductivity, electrical conductivity, density, microstructure, macrostructure, chemical composition, etc.).

[0002] To describe the relationship between the input and output parameters of a technical process, a (black-box) Bayesian optimization model can be used, which is generated by first learning a probabilistic model using measurement data of initial quantities, and then optimizing the probabilistic model using an acquisition function for selecting new measurement points (e.g., a trade-off between exploration and exploitation).

[0003] However, Bayesian optimization models have several drawbacks: the model does not allow determining whether a particular measurement point is not optimal. Furthermore, the acquisition function must be optimized, and therefore suffers from the common challenges of nonlinear optimization. For example, optimizing discrete systems using such methods would be very challenging because discrete optimization lacks continuity, which in turn challenges the application of gradient-based optimization. In practice, the output can simply be a binary variable (pass / fail) representing whether the system operates as expected, such as in quality assessment / auditing at the end of a production line. Moreover, Bayesian optimization does not support instance-based (e.g., context-aware) optimization. Therefore, Bayesian optimization allows determining how to make the system perform optimally, but does not allow determining what (e.g., minimal) changes must be applied to the technical system to handle existing faults. Summary of the Invention

[0004] This disclosure relates to a method for controlling a robotic device that allows for (e.g., real-time) explanation of system failures, such as determining which input parameters contribute to the failures and to what extent. Furthermore, the method allows for generating input parameters for controlling the robotic device that overcome system failures. Additionally, the method allows for determining these input parameters such that only small (e.g., minimal) changes to the technical system are required to prevent system failures.

[0005] For example, this is achieved by using a conditionally reversible neural network model that is configured (e.g., trained) to map input parameters to Shapley values ​​based on output parameters (by conditioned on the output parameters) and vice versa. In this way, input parameters can be generated by a single pass through the conditionally reversible neural network model (as opposed to iterative methods requiring multiple iterations), thereby reducing the computational cost required for control.

[0006] Illustratively, this method allows, in general, the computationally efficient generation of alternative input parameters (as control inputs) in the event of a given system failure or undesirable output parameters (one or more) (as observations) to optimize a given system toward a more desired operating range (e.g., a manufacturing pipeline process chain).

[0007] Various aspects relate to a method for controlling a robotic device, the method comprising: controlling the robotic device based on a corresponding initial input parameter value for each of a plurality of input parameters; and determining (e.g., detecting) a corresponding initial output parameter value for each of one or more (e.g., physical) output parameters generated by the control; and determining (e.g., predicting, estimating) a plurality of Shapley values ​​by inputting the corresponding initial input parameter values ​​for each of the plurality of input parameters as inputs and the corresponding initial output parameter values ​​for each of the one or more output parameters as conditional inputs to a conditionally invertible neural network model, wherein each of the plurality of Shapley values ​​corresponds to a pair of input parameters in the plurality of input parameters. The input parameters are associated and represent the contribution of the corresponding input parameter to one or more output parameters; a corresponding target output parameter value is determined for each of the one or more output parameters, wherein the corresponding target output parameter value of at least one of the one or more output parameters is different from the corresponding initial output parameter value of at least one output parameter; a corresponding target input parameter value is determined for each of the multiple input parameters by inputting multiple Shapley values ​​conditioned on the corresponding target output parameter value of each of the one or more output parameters into the inverse path of the conditional invertible neural network model; and the robot device is controlled according to the corresponding target input parameter value of each of the multiple input parameters.

[0008] Various examples are described below.

[0009] Example 1 is a method for controlling robotic equipment as described above.

[0010] In Example 2, the subject of Example 1 can optionally include multiple input parameters, each of which is a physical parameter, a chemical parameter, or a control parameter.

[0011] Therefore, input parameters can be any parameters used directly or indirectly to control the robotic device. As an illustrative example, the robotic device could be a drilling machine controlled to drill a hole. In this example, the input parameters could be physical parameters (such as coolant temperature, rotational speed, etc.) or control parameters (such as an applied voltage used to generate the desired rotational speed). Each output parameter could be a physical parameter, such as, in the example above, the smoothness of the drilled hole, the diameter of the drilled hole, etc.

[0012] In Example 3, the method of Example 1 or 2 may also optionally include: for at least one (e.g., each) of one or more predefined output parameters: determining whether the corresponding initial output parameter value of the at least one predefined output parameter is within a predefined range; and if it is determined that the corresponding initial output parameter value of the at least one predefined output parameter is not within a predefined range (e.g., thus detecting an anomaly), determining that the corresponding target output parameter value of the at least one predefined output parameter is within a predefined range.

[0013] Illustratively, this allows for the detection of, for example, anomalies. For instance, at least one (e.g., each) of one or more output parameters can be continuously monitored, and if at least one predefined output parameter is detected to be different from the expectation (e.g., outside the predefined range), the value of that at least one predefined output parameter is changed (to a target output parameter value) to meet the expectation (e.g., within the predefined range), and this value is then fed into the inverse path of the conditionally reversible neural network model to determine a new input parameter value that ensures at least one predefined output parameter meets the expectation.

[0014] In Example 4, the method of any one of Examples 1 to 3 may also optionally include: adapting the plurality of Shapley values ​​before inputting them into the inverse path of the conditional invertible neural network model, wherein adapting the plurality of Shapley values ​​includes changing one or more of the plurality of Shapley values ​​to zero.

[0015] This allows for a limitation on the number of input parameters that can be changed to achieve the target output parameter value, since a Shapley value changed to zero indicates that the corresponding input parameter does not contribute to one or more output parameters.

[0016] In Example 5, the subject of Example 4 can optionally include one or more Shapley values ​​that are a predefined number of minimum Shapley values; or one or more Shapley values ​​that are all Shapley values ​​that are equal to or less than a threshold among a plurality of Shapley values.

[0017] The larger the Shapley value, the smaller the value of the corresponding input parameter that must be changed to achieve a change in one or more output parameters. Therefore, when these adapted Shapley values ​​are input into the inverse path of a conditionally invertible neural network model, only the input parameters associated with the larger Shapley values ​​are adapted to achieve the target output parameter values. Since the changes to these input parameters associated with larger Shapley values ​​are smaller than the changes made to other input parameters, the total change to multiple input parameters is reduced (e.g., minimized) while still achieving the target output parameter values.

[0018] In Example 6, the method of any one of Examples 1 to 5 may also optionally include: adapting the plurality of Shapley values ​​before inputting them into the inverse path of the conditional invertible neural network model, wherein adapting the plurality of Shapley values ​​includes changing all of the plurality of Shapley values ​​except the largest Shapley value to zero.

[0019] In this case, only one input parameter value needs to be changed to achieve the target output parameter value, thereby increasing the effectiveness of the robot device.

[0020] Example 7 is a robot device controller configured to perform any of the methods in Examples 1 through 6.

[0021] Example 8 is a technical system comprising: a robot device controller according to Example 7; a robot device; and one or more sensors configured to acquire a corresponding initial output parameter value for each of one or more output parameters.

[0022] Example 9 is a computer program that includes instructions that, when executed by a computer, cause the computer to perform a method according to any one of Examples 1 to 6.

[0023] Example 10 is a computer-readable medium including instructions that, when executed by a computer, cause the computer to perform a method according to any one of Examples 1 to 6. Attached Figure Description

[0024] In the accompanying drawings, similar reference numerals generally refer to the same parts throughout the different views. The drawings are not necessarily drawn to scale; instead, the focus is generally on illustrating the principles of the invention. In the following description, various aspects are described with reference to the following drawings, wherein: Figure 1 A flowchart illustrating a method for controlling a robotic device based on various aspects (using a conditionally reversible neural network model) is shown. Figure 2 An exemplary system is shown that can execute the method thereon, depending on various aspects; Figure 3The architecture of a conditionally reversible neural network model based on various aspects is shown; Figure 4 This illustrates the process of adapting input parameters to a reversible neural network model based on various usage conditions; and Figure 5 A flowchart illustrating a method for managing system failures is shown. Detailed Implementation

[0025] The following detailed description is taken with reference to the accompanying drawings, which illustrate by way of illustration specific details and aspects of the present disclosure in which the invention may be practiced. Other aspects may be utilized, and structural, logical, and electrical changes may be made, without departing from the scope of the invention. The various aspects of the present disclosure are not necessarily mutually exclusive, as some aspects of the present disclosure can be combined with one or more other aspects of the present disclosure to form new aspects.

[0026] The various examples will be described in more detail below.

[0027] Figure 1 A flowchart of a method 100 for controlling a robotic device according to various aspects is shown.

[0028] Method 100 may (in 102) include controlling the robotic device based on a corresponding initial input parameter value for each of a plurality of input parameters, and determining (e.g., detecting) a corresponding initial output parameter value for each of one or more (e.g., physical) output parameters generated by the control.

[0029] Method 100 may (in 104) include determining (e.g., predicting, estimating) multiple Shapley values ​​by inputting a corresponding initial input parameter value for each of a plurality of input parameters as input and a corresponding initial output parameter value for each of one or more output parameters as conditional inputs to a conditionally invertible neural network model. Each of the multiple Shapley values ​​may be associated with a corresponding input parameter among the plurality of input parameters. Each of the multiple Shapley values ​​may represent the contribution of the corresponding input parameter to one or more output parameters.

[0030] Method 100 may (in 106) include determining a corresponding target output parameter value for each of one or more output parameters. Depending on various aspects, the corresponding target output parameter value for at least one of the one or more output parameters may differ from the corresponding initial output parameter value for at least one output parameter.

[0031] Method 100 may (in 108) include determining the corresponding target input parameter value for each of the multiple input parameters by inputting multiple Shapley values, conditional on the corresponding target output parameter value of each of the one or more output parameters, into the inverse path of the conditional invertible neural network model.

[0032] Method 100 may (in 110) include controlling the robot device according to the corresponding target input parameter value of each of the plurality of input parameters.

[0033] In the following text, see references Figure 2 The technical system 200 shown further details various aspects of method 100. Technical system 200 may include robotic device 202. Robotic device 202 may be configured to perform technical processes. Robotic device 202 may be any kind of (computer-controlled) technical equipment configured to perform technical processes, such as robots (e.g., manufacturing robots, maintenance robots, household robots, medical robots, craft tools (e.g., drills), etc.), vehicles (e.g., at least partially automated (e.g., autonomous) vehicles), household appliances, production machines, personal assistants, access control systems, etc., and any other type of robotic device.

[0034] Depending on the aspects, a technical process can be a physical or chemical process, such as, for example, a manufacturing process (e.g., manufacturing a product or intermediate product), a machining process (e.g., machining a workpiece), a control process (e.g., moving a robotic arm), a setup process (e.g., calibrating a measuring device), and so on.

[0035] To control the robot device 202, the technical system 200 may include a robot device controller (hereinafter referred to as "controller") 204. The controller 204 may be configured to control the operation of the robot device 202 (according to a control program). The term "controller" can be understood as any type of logical implementation entity, which may include, for example, circuitry and / or a processor, firmware, or a combination thereof, capable of executing software stored in a storage medium, and capable of providing instructions, such as providing instructions to an actuator in this example. The controller may be configured to control the operation of the system 200 (e.g., the robot device 202) for example through program code (e.g., software).

[0036] The controller 204 can be configured to control the robot device 202 based on a plurality of input parameters 206. Therefore, the controller 204 can control the robot device 202 based on the corresponding input parameter value of each of the plurality of input parameters 206(n). The plurality of input parameters 206 may include a number of N input parameters, where N is any integer equal to or greater than 2.

[0037] The input parameters used herein may be physical parameters, chemical parameters, control parameters, or any other kind of parameters describing the technical system 200.

[0038] One or more sensors 208 can be used to acquire the results of the technical process. The one or more sensors 208 can be configured to acquire the corresponding output parameter value of each of the one or more output parameters 210(m). The one or more output parameters 210 may include M output parameters, where M is any integer equal to or greater than 1.

[0039] As used herein, output parameters can be parameters that describe the output of a technical process. Output parameters can be attributes of a product, workpiece, captured images, or other results. Output parameters can have application-specific quality standards. Output parameters can be component-related parameters, such as dimensions or layer thickness, or material-related parameters, such as hardness, thermal conductivity, electrical conductivity, density, chemical composition, etc.

[0040] As an example, the robotic device 202 could be a drilling machine controlled to drill a hole. Therefore, the technical process could be drilling a hole. Input parameter 206(n) could be, for example, physical parameters such as the temperature of the coolant used during drilling, the rotational speed of the drill bit, etc., or it could be control parameters such as the voltage applied to a unit of the robotic device 202 (e.g., an actuator configured to rotate the drill bit according to an applied voltage). Output parameter 210(m) could be, for example, the smoothness of the drilled hole, the diameter of the drilled hole, or any other property of the drilled hole.

[0041] As another example, the technical process can be a production or manufacturing process performed in a processing chamber for manufacturing a workpiece. In this example, input parameter 206(n) can be, for example, the temperature of the workpiece during the technical process, the voltage applied to the actuator, or chemical parameters, the concentration or (partial) pressure of the processing gas in the processing chamber. The input parameter can also be another parameter representing the technical process, such as the opening of a valve configured to regulate the inlet of the processing gas. Thus, the valve opening can indicate the (partial) pressure of the processing gas.

[0042] It is important to understand that these are examples for illustration, and the technical process can be any other kind of technical process, such as another production or machining process (e.g., milling, heat treatment), the technical process of a vehicle as a robotic device (where the input or output parameters are, for example, braking force, speed, etc.), and so on.

[0043] The corresponding output parameter value of each of the one or more output parameters 210(m) can be acquired during the execution of the technical process (i.e., in-situ) and / or after the execution of the technical process (i.e., ex-situ) (using one or more sensors 208).

[0044] Therefore, in an example where the technical process is a production or manufacturing process performed in a processing chamber (i.e., in situ), the output parameter acquired during this process could be the temperature of the workpiece during manufacturing. As an ex-situ example, the technical process could be hardening the workpiece in an oven with temperature as the input parameter, and the output parameter could be the hardness of the workpiece at room temperature after the hardening process.

[0045] Controller 204 can be configured to implement a conditionally reversible neural network model 212. Controller 204 can be configured to determine, for at least one (e.g., each) of one or more output parameters 210, whether the at least one output parameter 210(m*) satisfies a predefined requirement, such as being within a predefined range. It should be understood that a predefined range can also be provided by combining a predefined value with (acceptable) tolerances. Controller 204 can be configured to adapt at least one output parameter 210(m*) to satisfy the predefined requirement (e.g., to be within a predefined range) if at least one output parameter 210(m*) does not satisfy the predefined requirement. Controller 204 can be configured to use the conditionally reversible neural network model 212 to determine a corresponding target input parameter value for each of a plurality of input parameters 206(n), which, when controlling the robotic device 202 according to the target input parameter value, is expected to produce an adapted output parameter value for at least one output parameter 210(m*).

[0046] Figure 3 The architecture of a conditionally reversible neural network model 212 based on various aspects is shown.

[0047] For example, the architecture of the conditionally invertible neural network model 212 is described in L. Ardizzone et al.'s "GUIDED IMAGE GENERATION WITH CONDITIONALINVERTIBLE NEURAL NETWORKS", arXiv:1907.02392,2019 (hereinafter referred to as reference [1]).

[0048] The following provides a general overview of the architecture of the conditionally invertible neural network model 212. For further details, please refer to the aforementioned reference [1].

[0049] The conditionally reversible neural network model 212 may include five (deep) neural networks. , , , and The conditionally reversible neural network model 212 may include a unidirectional (i.e., non-reversible) path 302. Other (upper and lower) paths of the conditionally reversible neural network model 212 are reversible. Therefore, illustratively, when inverted, the output and conditions can be input into the conditionally reversible neural network model 212 (the inverted path) to determine the (new) input. Illustratively, the upper and lower paths allow flow in either direction. This is achieved through the use of neural networks. , , and This is achieved through an invertible functional relationship between the outputs, such as element-wise multiplication. summation element by element .

[0050] In the forward non-reverse path, the corresponding input parameter value of each of the multiple input parameters 206(n) is input (as input), and the corresponding output parameter value of each of the one or more output parameters 210(m) is input as a condition to the conditional reversible neural network model 212. Depending on various aspects, the conditional reversible neural network model 212 is configured to generate a corresponding Shapley value 304(n) for each of the multiple input parameters 206(n) as output. Therefore, the conditional reversible neural network model 212 can output multiple Shapley values ​​304. (including N Shapley values).

[0051] The input (i.e., the corresponding input parameter value for each of the multiple input parameters 206(n)) can be divided into two parts. and The result on the upper path can be determined according to the following equation. and results on the lower path : These results They can then be concatenated to determine multiple Shapley values ​​of 304.

[0052] When in the opposite direction (i.e., in the opposite direction), the two parts and It can be determined by the following equation: .

[0053] For further details regarding the architecture of the conditionally reversible neural network model 212, please refer to the aforementioned reference [1].

[0054] Shapley values ​​have been introduced for multiplayer game scenarios, where a Shapley value is determined for each of the multiple players, and this Shapley value represents that player's contribution to the result achieved by all players together. In the case of technical system 200, multiple input parameters 206 are players, and one or more output parameters 210 are the consequences of the interaction of all those input parameters 206. Therefore, each Shapley value 304(n) associated with a corresponding input parameter 206(n) represents the contribution of the corresponding input parameter 206(n) to achieving one or more output parameters 210.

[0055] Determining the Shapley value is an NP-hard problem that is difficult to scale in real-world applications. The standard formulas for calculating the Shapley value are either the exponential formula (navigating the space of all player subsets) or the combinatorial formula (sorting the space of all players): Exponential formula: in It is the collection of all players (also known as features or attributes in the model). It is a subset of players (excluding the target player) ),and It is a value / set function that returns the game reward when playing with a given set of players (e.g., the performance of a machine learning model when only a subset of specific features is used as input).

[0056] Combination formula: in It is the set of all possible player sortings, and In sorting In the middle of The set of all players preceding the current player.

[0057] In either case, the precise calculation of the Shapley value is only feasible for a very small number of system parameters (about 10 parameters) and cannot be scaled to meet real-world needs.

[0058] Depending on the aspects, the conditional invertible neural network model 212 may be, or may have been, trained to predict (e.g., estimate) the Shapley value 304 in response to inputs of multiple input parameter values ​​as inputs and one or more output parameter values ​​as conditions. In particular, the conditional invertible neural network model 212 may have been trained using the following loss function: in These are model parameters. It is the model input (i.e., multiple input parameters 206). This is the expected output (i.e., with a Shapley value of 304). It is a subset of features (and The weights of the encoded subsets, as given in the original Shapley formula. Is only a subset of features The model output, and It is the baseline value (the nominal prediction of the model without using features). It is the sum of the interpreters' attributions for a given set of players / features.

[0059] For example, the use of this loss function to train a neural network is described in N. Jethani et al., “FASTSHAP: REAL-TIME Sapley VALUE ESTIMATION”, arXiv: 2107.07436, 2022 (hereinafter referred to as Reference [2]). It has been found that this loss function can be used to train a conditionally invertible neural network model 212 because the dimensions of the input and output are the same (because there is exactly one Sapley value for each input parameter). Using the loss function above allows for a reduction in the computational cost of training the conditionally invertible neural network model 212. Using this loss function, the conditionally invertible neural network model 212 can be trained to predict Sapley values ​​without having to pre-compute them. For further details on how to use the loss function above, see the aforementioned Reference [2].

[0060] Figure 4 The process flow is shown to adapt the corresponding input parameter value 402(n) of each of the multiple input parameters 206(n) to the reversible neural network model 212 according to the usage conditions of various aspects.

[0061] As described in detail herein, controller 204 can (in 102) control robot device 202 based on a plurality of initial input parameter values ​​402 (each of the plurality of input parameters 206(n) has a corresponding initial input parameter value 402(n)). And can (using one or more sensors 208) detect one or more initial output parameter values ​​404 (each of the one or more output parameters 210(m) has a corresponding initial output parameter value 404(m)) (generated by the plurality of input parameter values ​​402).

[0062] In 104, multiple initial input parameter values ​​402 can be used as inputs, and one or more initial output parameter values ​​404 can be used as conditional inputs into the conditional reversible neural network model 212 (the forward path) to determine multiple Shapley values ​​304.

[0063] Then, one or more target output parameter values ​​408 can be determined using the corresponding target output parameter 408(m) for each of the one or more output parameters 210(m). Depending on various aspects, the target output parameter value 408(m*) of at least one output parameter 210(m*) may differ from the initial parameter value 404(m*) of at least one output parameter 210(m*).

[0064] In 108, multiple Shapley values ​​304 (or their adapted versions) (as outputs) and one or more target output parameter values ​​408 (as conditions) can be input into the inverse path of the conditional invertible neural network model 212 to determine multiple target input parameter values ​​410 (each of the multiple input parameters 206(n) has a corresponding target input parameter value 410(n)).

[0065] Depending on the circumstances, one or more Shaple values ​​from a plurality of Shaple values ​​304 can be adapted before being input into the inverse path of the conditional reversible neural network model 212.

[0066] For example, one or more Shapley values ​​can be adapted by changing their values ​​to zero. As described in detail herein, each Shapley value represents the contribution of the corresponding input parameter to one or more output parameters. Therefore, when its value is changed to zero, the input parameter is considered to have no contribution to one or more output parameters. This results in the input parameter values ​​of the corresponding input parameters not changing when multiple target input parameter values ​​410 are determined. Illustratively, the target input parameter values ​​of the input parameters corresponding to the Shapley values ​​then correspond to the initial input parameter values. This allows the adaptation of multiple input parameters 206 to be restricted to their finite subset.

[0067] Depending on the aspects, one or more Shapley values ​​are a predefined number of minimum Shapley values ​​(thus only the input parameters that contribute the most to one or more output parameters 210 are adapted, and therefore require a small change in their (one or more) values).

[0068] Depending on the circumstances, only the largest Shapley value can be used, and all other Shapley values ​​can be changed to zero. This allows one or more target output parameter values ​​of 408 to be achieved by changing only one input parameter.

[0069] Depending on various aspects, one or more target output parameter values ​​408 may be determined (e.g., only if at least one of the initial output parameter values ​​is determined to be non-compliant with a predefined requirement (e.g., indicating a system failure).

[0070] Figure 5 A flowchart of a method 500 for managing system failures based on various aspects is shown.

[0071] In 102, the technology system 200 can be controlled according to multiple initial input parameter values ​​402, and the output can be monitored (one or more output parameter values ​​404 can be determined).

[0072] Method 500 may include (in 502) determining, for at least one (e.g., each) output parameter 210 (m*), whether the initial output parameter value 404 (m*) of at least one output parameter 210 (m*) satisfies a predefined requirement (e.g., within a predefined range).

[0073] If it is determined that the initial output parameter value 404 (m*) of at least one output parameter 210 (m*) satisfies the predefined requirements (e.g., within the predefined range), the method 500 may proceed to step 102 (monitoring technology system 200).

[0074] If it is determined that the initial output parameter value 404(m*) of at least one output parameter 210(m*) does not meet the predefined requirements (e.g., is not within the predefined range), method 500 may include (in 106) determining one or more target output parameter values ​​408 by adapting the output parameter value of at least one output parameter 210(m*) to meet the predefined requirements (e.g., within the predefined range).

[0075] In 104, multiple Shapley values ​​304 can be determined as described in detail herein. Optionally, multiple Shapley values ​​304 can be used to explain one or more observed initial output parameter values ​​404 (e.g., by explaining the corresponding contribution of each input parameter 206(n)).

[0076] In 108, multiple Shapley values ​​304 (or multiple adapted Shapley values) and one or more target output parameter values ​​408 can be input into the inverse path of the conditional invertible neural network model 212 to determine multiple target input parameter values ​​410. Then (in 110) the robot device 202 can be controlled based on the multiple target input parameter values ​​410, and the output can be monitored again (in 102).

[0077] Illustratively, method 500 allows for the detection of anomalies (e.g., system failures) and the determination of adaptations to input parameters so that the anomalies are overcome. Illustratively, methods 100 and 500 allow the technical system 200 to be brought back to a desired (e.g., predefined) operating range.

[0078] The methods 100 and 500 described in this paper allow the use of any kind of input parameters because the conditional invertible neural network model 212 does not impose any data type restrictions (making the data type of the input parameters continuous, categorical, binary, etc.).

Claims

1. A method (100, 500) for controlling a robotic device, the method (100, 500) comprising: The robot device (202) is controlled (102) according to the corresponding initial input parameter value of each of the multiple input parameters (206), and the corresponding initial output parameter value of each of the one or more output parameters (210) generated by the control is determined; Multiple Shapley values ​​(304) are determined (104) by taking the corresponding initial input parameter value of each of the multiple input parameters (206) as input and taking the corresponding initial output parameter value of each of the one or more output parameters (210) as conditional input to a conditional invertible neural network model (212), wherein each of the multiple Shapley values ​​(304) is associated with a corresponding input parameter in the multiple input parameters (206) and represents the contribution of the corresponding input parameter to the one or more output parameters (210); Determine (106) a corresponding target output parameter value for each of the one or more output parameters (210), wherein the corresponding target output parameter value of at least one of the one or more output parameters (210) is different from the corresponding initial output parameter value of the at least one output parameter; By inputting the plurality of Shapley values ​​(304) conditioned on the corresponding target output parameter value of each of the one or more output parameters into the inverse path of the conditional invertible neural network model (212), the corresponding target input parameter value of each of the plurality of input parameters (206) is determined (108); as well as The robot device (202) is controlled (110) according to the corresponding target input parameter value of each of the plurality of input parameters.

2. The method according to claim 1 (100, 500). Each of the plurality of input parameters (206) is a physical parameter, a chemical parameter, or a control parameter.

3. The method (100, 500) according to claim 1 or 2, further comprising: For at least one of the one or more predefined output parameters: Determine whether the corresponding initial output parameter value of the at least one predefined output parameter is within the predefined range (502); as well as If it is determined that the corresponding initial output parameter value of the at least one predefined output parameter is not within the predefined range, then it is determined that the corresponding target output parameter value of the at least one predefined output parameter is within the predefined range.

4. The method (100, 500) according to any one of claims 1 to 3, further comprising: Before inputting the plurality of Shapley values ​​(304) into the inverse path of the conditional reversible neural network model (212), the plurality of Shapley values ​​(304) are adapted, wherein adapting the plurality of Shapley values ​​(304) includes changing one or more of the plurality of Shapley values ​​(304) to zero.

5. The method according to claim 4 (100, 500). The one or more Shapley values ​​mentioned therein are a predefined number of minimum Shapley values; or The one or more Shapley values ​​mentioned therein are all Shapley values ​​that are equal to or less than the threshold among the plurality of Shapley values ​​(304).

6. The method (100, 500) according to any one of claims 1 to 5, further comprising: Before inputting the plurality of Shapley values ​​(304) into the inverse path of the conditional reversible neural network model (212), the plurality of Shapley values ​​(304) are adapted, wherein adapting the plurality of Shapley values ​​(304) includes changing all Shapley values ​​(304) except the largest Shapley value to zero.

7. A robot device controller (204) configured to perform the method (100, 500) according to any one of claims 1 to 6.

8. A technical system (200), comprising: The robot device controller (204) according to claim 7; Robotic equipment (202); as well as One or more sensors (208) are configured to acquire the corresponding initial output parameter value of each of the one or more output parameters (210).

9. A computer program comprising instructions that, when executed by a computer, cause the computer to perform the method (100, 500) according to any one of claims 1 to 6.

10. A computer-readable medium comprising instructions that, when executed by a computer, cause the computer to perform the method (100, 500) according to any one of claims 1 to 6.

Images