Robotic device

Abstract

Robotic device (1) with - at least one controllable gripping device (2) for gripping an object (3) in space (R) with a gripper (20); - at least one sensor device (4) which is designed to detect the object (3) to be grasped and then to provide a set of object points in the form of three-dimensional data points in a first coordinate system (K1) as data; - at least one data processing device (5) which is designed to process data provided by the sensor device (4) and to calculate position parameters for the position and orientation of the object (3) in space (R) therefrom, wherein the processing of the data is carried out by means of an artificial neural network, and - at least one control device (6) for controlling the at least one gripping device (2), wherein the control device (6) is designed to generate control commands for controlling the gripping device (2) from the position parameters for the position and orientation of the object (3) calculated by the data processing device (5) and to transmit these control commands to the gripping device (2) in order to move to a target configuration of the gripping device (2) in space (R) for gripping the object (3) and to grip the object (3) there, characterized in that the data processing device (5) is designed to process the data detected by the sensor device (4) as a set of object points in the manner and sequence listed below in order to obtain the position parameters for the position and orientation of the object (3) transferred to the control device (6): - encoding the set of object points as three-dimensional data points in the first coordinate system (K1) into a data vector of ordered components; - Execution of a sequence of operators to successively reduce the data vector in its dimension, where the operators are each composed ◯ from a linear transformation and ◯ a subsequent non-linear function applied to each individual component of the previously generated data vector, so that each time a component of a new data vector is created; - linear transformation to a multidimensional data vector, the components of which contain predicted parameters for describing the spatial position of the target object in the first coordinate system (K1) used by the sensor device (4), and - Conversion of the predicted parameters for the object position into new position parameters which describe the spatial position of the target object in a second coordinate system (K2) used by the control device (6).

Description

The present invention relates to a robotic device having at least one controllable gripping device for gripping an object in space, in particular in space.A robot equipped with a gripping device, which is intended to grip an object that is geometrically known to it but unknown with regard to its object position, must first of all detect it with a sensor device, determine the object position and the object orientation from the sensor data, and then control its gripping device in such a way that it can reliably grip the object as a function of the determined object position and object orientation. This applies in particular if the object floats freely in space, for example in space. If the object appears in the sensor data in some way symmetrically, in particular rotationally symmetrically, and this is either due to an actual symmetry of the object or lack of the ability of the sensor device to distinguish different object orientations, the object orientation cannot be determined unambiguously. The orientation ambiguity which then results is unavoidable and has to be taken into account in the control of the gripping device in order to bring it into a suitable gripping position.The case of a limited distinguishability of the object orientation arises in particular in the case of discrete rotational symmetry of the object or if a deviation of the object shape and / or texturing from discrete rotational symmetry due to limited resolution or artifacts are not sufficiently reproduced in the data of imaging sensors.Regression methods are frequently used for estimating (predicting) the object position in six degrees of freedom (three degrees of freedom for the orientation and three degrees of freedom for the translation) from the data of imaging sensors, which regression methods implement continuous imaging of the sensor data to form unique position parameters. Continuity and uniqueness are enforced by the method used, for example by the use of artificial neural networks. In the case of limited distinguishability of the object orientation due to discrete rotational symmetry or corresponding orientation ambiguity due to limited resolution or artifacts in the data of imaging sensors, a steady and clear parameterization of the distinguishable object orientations is possible only by specially adapted orientation parameters. Such parameterization therefore increases the achievable accuracy of such regression methods for the object position within the distinguishable object orientations in such cases.Such a parameterization is given when using the class of the so-called "expanded orientation parameters". This type of parameter can be adapted to any discrete rotational symmetry or generally to a corresponding orientation ambiguity of an object in the sensor data, and also to the absence of rotational symmetry or orientation ambiguity, so that a clear and continuous parameterization always results within the distinguishable object orientations.Orientations of an object in a three-dimensional space are equivalent to rotations of the object starting from an arbitrarily selected reference orientation. The class of expanded orientation parameters is therefore equivalent to a corresponding class of expanded rotational parameters.DE 10 2018 105 544 A1 discloses a method for initializing a tracking algorithm for aviation and / or aerospace target objects, in which a three-dimensional point cloud of the target object is generated by means of an optical scan and a spatial position of the target object is determined iteratively on the basis of a 3D model of the target object. With the aid of an artificial neural network, a spatial position of the target object is first determined and a tracking algorithm is then initialized. In this case, firstly, a prediction of correspondences between a part of the scan data and a part of the model data is carried out by means of the artificial neural network, in that polygons are respectively generated from individual data points. A transformation between the model data and the scan data is then calculated by means of classic optimization. Finally, the quality of the predicted correspondence is evaluated in the form of a geometric match between the selected polygons. If the quality is sufficient, the transformation is used as initialization of the position tracker; if the quality is not sufficient, a new prediction is carried out. The computing effort and the computing duration are high, so that a fast and energy-hungry computer is required.It is the object of the present invention to improve a device of the generic type in such a way that it achieves satisfactory results with minimal computing effort and low energy requirements.This object is achieved by a device having the features of claim 1.A robotic device is provided with at least one controllable gripping device for gripping an object in space with a gripper, at least one sensor device which is designed to detect the object to be gripped and then to provide a set of object points as three-dimensional data points in a first coordinate system as data, at least one data processing device which is designed to process data provided by the sensor device and to calculate position parameters for the position and the orientation of the object in space therefrom, wherein the data are processed by means of an artificial neural network, and at least one control device for controlling the at least one gripping device, wherein the control device is designed to, in order to generate control commands for controlling the gripping device from the position parameters for the position and the orientation of the object calculated by the data processing device and to transmit these to the gripping device in order to approach a target configuration of the gripping device in space for gripping the object and to grip the object there. This robotic device is characterized in that the data processing device is designed to process the data recorded by the sensor device as a set of object points in the manner and sequence listed below in order to obtain the position parameters for the position and the orientation of the object that are passed to the control device:encoding the set of object points as three-dimensional data points in a first coordinate system into a data vector of ordered components;executing a sequence of operators to successively reduce the data vector in its dimension, wherein the operators are respectively composed◯ from a linear transformation and◯ a subsequent nonlinear function applied to each individual component of the previously generated data vector, so that in each case a component of a new data vector is again produced;linear transformation to a multidimensional data vector whose components contain predicted parameters for describing the spatial position of the target object in the first coordinate system used by the sensor device, andconverting the predicted parameters for the object position into new position parameters which describe the spatial position of the target object in a second coordinate system used by the control device.By a "target configuration of the gripping device in space" is meant the position and orientation of the gripping device, especially the gripper, which must be taken in order to be able to grip the object in the desired manner.The parameters for the object position are the position parameters for the position and the orientation of the object in space.The term "coding" is understood to mean a transformation of a set of points in space, described by their coordinates, into a numerical vector.The term "operator" is understood to mean a computing step which transforms a number vector into another number vector.The term "vector" or "numerical vector" is understood here to mean a numerical tuple.In the data processing device of the robotic device according to the invention, the neural network directly predicts the position parameters and thus a transformation between a model coordinate system and a coordinate system of the sensor device without the intermediate step of predicting scan model correspondences, thereby saving computer capacity. The position of an object in any coordinate system, for example as here in the coordinate system of the sensor device, is equivalent to a transformation from a model coordinate system, in which the model geometry is described, into a target coordinate system, that is here into the coordinate system of the sensor device: this transformation "moves" the object model into the sensor data world. This procedure running in the data processing device has the advantage that no recursion of computing steps is required until suitable position parameters are obtained which describe the spatial position of the target object in the second coordinate system used by the control device, as a result of which computing time is saved.Further preferred and advantageous design features of the robotic device according to the invention are the subject matter of the dependent claims 2 to 9.According to a preferred embodiment of the robotic device, it is provided that the data processing device is configured such that the predicted parameters for the object position (predicted object position parameters) are subjected to an iterative correction before the conversion into the new position parameters of the second coordinate system in order to reduce an error measure of the object position. For this purpose, a local optimization of the calculated predicted object position parameters takes place, wherein an iterative correction of these parameters is carried out on the basis of the calculated predicted object position parameters, so that an error measure of the object position is further reduced in each iteration step. This iteration is preferably terminated when no reduction in the error measure is achieved in this manner. The error measure of the object location is a function of the object location parameters, a geometric model of the object, and the object points as three-dimensional data points in the first coordinate system that takes the smallest possible value for location parameter values desired for the given model of the object and the given data points as target values.In the case of a moving object, it is advantageous to first track the object in order to include the dynamics of the movement in the prediction. For this purpose, it is preferably provided that the data processing device is configured such that the determination of the predicted parameters for the object position is carried out repeatedly in time sequences several times before the conversion into the new position parameters of the second coordinate system in order to track the moving object. The determination of the predicted parameters for the object position is carried out as a temporal sequence based on a temporal sequence of sets of object points as three-dimensional data points in the first coordinate system by calculating the first element of the sequence of the predicted parameters based on the first element of the sequence of the set of data points in the data processing device only by the neural network without optimization and without sequence and then correcting it by local optimization, and obtaining each further element of the sequence of the predicted parameters as correction of the preceding element by local optimization based on each further element of the sequence of the set of data points. The local optimization is ended for each element of the sequence of predicted parameters and the result is transferred at the latest when the next element of the sequence of the set of data points is available or the result is to be converted into the new position parameters of the second coordinate system and transferred to the control device.In a further preferred embodiment of the robotic device, it is provided that the data processing device is configured such that the parameters predicted by the data processing device for describing the spatial orientation of the target object, namely the orientation parameters, are selected in the first coordinate system used by the sensor device such that the parameterization is unique and continuous for each parameter value.Such parameterization is advantageous in particular when implementing the data processing device using artificial neural networks, because this continuity corresponds to the fundamental properties of parameter regression by artificial neural networks. Such parameterization therefore enables more accurate parameter regression with artificial neural networks than one with discontinuities.It is advantageous to realize a clear and continuous parameterization by using the class of the so-called "expanded orientation parameters" for the object orientation. This type of parameter can be adapted to any discrete rotational symmetry or generally to a corresponding orientation ambiguity of an object in the sensor data, and also to the absence of rotational symmetry or orientation ambiguity, so that a clear and continuous parameterization always results within the distinguishable object orientations. In contrast, conventional orientation parameters such as, for example, Euler angles, quaternions, axis-angle parameters are not unambiguous in the sensor data in the case of discrete rotational symmetry or generally corresponding orientation ambiguity of an object and are not continuous in the case of some parameter values.The use of expanded parameters causes each object view (set of object points) to have exactly one vector of parameter values associated therewith. It can be said that the multivalueness of object views with respect to orientation parameter values is eliminated when parameterizing with expanded orientation parameters. The use of expanded parameters also brings about the continuity of the parameterization.When parameterizing with expanded orientation parameters, the object orientation is described by the directions in space of two orthogonal unit vectors in the object model. These are parameterized by their Cartesian coordinates, i.e. by a six-dimensional parameterization. Discrete rotational balances and generally corresponding orientation ambiguities of an object in the sensor data are taken into account as follows: In the case of n-fold symmetry or ambiguity about an axis, a parameterized object vector, i.e. an object direction, is not fixedly anchored in the object orthogonally to this axis, but rather rotates n-fold of an object rotation about the axis of symmetry.It is also advantageous to perform a correction of the predicted expanded orientation parameters in order to compensate for deviations from the specification of the parameterization. These can arise from errors, including small errors, in the prediction of the parameters, which are unavoidable in practice. In this case, the two calculated predicted vectors are rotated in equal parts in opposite directions by the smallest possible angle in order to align them orthogonally with respect to one another, and the two vectors thus rotated are then normalized to unit vectors.Preferably, the data processing device is configured such that the multi-dimensional data vector generated by the linear transformation is a nine-dimensional data vector. Of these, three dimensions are reserved for the position parameter values of the object in space and six dimensions are reserved for the expanded orientation parameter values of the object in space.It is also advantageous if the data processing device is configured such that the new position parameters of the second coordinate system form a rotation matrix and a translation vector.Alternatively, it is advantageous if the data processing device is configured such that the new position parameters of the second coordinate system are formed by Euler angles and / or quaternions.The gripping device preferably has at least one robot gripping arm or it is designed as a robot gripping arm, wherein the robot gripping arm is provided with the gripper.It is also particularly advantageous if the sensor device has at least one LIDAR sensor or is designed as a LIDAR sensor. As a result, the sensor device can capture three-dimensional object points in space. This can also be achieved if the sensor device is formed by a depth camera or a stereo camera system.The invention is also directed to a space satellite formed as such a robotic device or equipped with such a robotic device, wherein the object to be gripped is a space object, for example a dentate or defective satellite or a piece of space scrap.A particularly suitable application of the robotic device according to the invention is an application in space, for example in order to grasp as object a space object, such as space scrap or a damaged satellite, for example, and to transfer from its current orbit into another position, for example in a Friedhofen sorbitol, or to bring it to a crash on a controlled path. For this purpose, a space satellite is designed as a robotic device according to the invention or a robotic device according to the invention is provided on a space satellite.The robotic device according to the invention enables a rapid position estimation of objects in space that are to be gripped by a gripping device with a relatively low computing effort. In a preferred application, these objects are satellites in an orbit that must be gripped by a service satellite for service purposes in an on-orbit service deployment. For this purpose, such an object must first be detected by means of a sensor device, for example a LIDAR. The data processing device of the robotic device forms from the acquired LIDAR data an estimate of the position of the object, with the result of which the gripping device is brought into a gripping position and gripping orientation in space. This is done by means of an artificial neural network, which is trained on simulated data. Preferably, a special parameterization is used for regression of the satellite orientation. The advantage of this special orientation parameterization is not limited to satellites, on-orbit serving or LIDAR data, but can also result in other cases of object position estimates.A prerequisite for a regression is uniqueness, that is, only one value of the parameters exists for each object view (given as the set of object points). In the case of parameter regression, therefore, there is a need for explicit parameterization. For regression, in particular with neural networks, continuity is advantageous, that is to say that, in the case of a small change in the object view, only a small change in the parameter values occurs and no jumps occur in the parameter values. The challenge is thus to simultaneously ensure a clear and a continuous parameterization of the object orientation for objects with discrete rotational symmetry, or generally formulated with corresponding orientation ambiguity in the sensor data. This is effected according to the invention by means of orientation parameterization by parameters from the class of expanded orientation parameters. Parameters from this class can be systematically adapted to each discrete rotational symmetry or generally to a corresponding orientation ambiguity, and also to the absence of any orientation ambiguity. Therefore, the term "class" is chosen for all variants of these parameters.Preferred exemplary embodiments of the invention with additional design details and further advantages are described and explained in more detail below with reference to the attached drawings.It shows: FIG. 1 shows a schematic arrangement of a robotic device with a gripping device for gripping an object in space, and FIG. 2 shows a schematic diagram of the principle of an orientation parameterization preferably used for grasping the object from the robotic device.FIG. 1 shows a schematic illustration of a robotic device 1 having a controllable gripping device 2 for gripping an object 3, also referred to as a "target object", in a space R. In this example, the robotic device 1 is a service satellite and the object 3 is a piece of space scrap floating in the space R to be caught by the gripping device 2.The robotic device 1, which is designed as a service satellite, has a satellite body 10, on which a robot gripping arm 22' of the gripping device 2, which is designed as an articulated arm 22, is mounted so as to be pivotable about at least one axis by means of a first joint 23. The articulated arm 22 has a first articulated arm rod 24 which is pivotably mounted on the satellite body 10 by means of the first joint 23, on the free end of which rod a second joint 25 is attached, which in turn is connected to a second articulated arm rod 26. The second joint 25 enables the second articulated arm rod 26 to be pivoted relative to the first articulated arm rod 24 about at least one axis. At the free end of the second articulated arm rod 26, a third articulated arm rod 28 is arranged such that it can be pivoted about at least one axis by means of a third joint 27. The third articulated arm rod 28 is designed to be rotatable about a longitudinal axis. A gripper 20 of the gripping device 2 is arranged at the free end of the third articulated arm rod 28, wherein the gripper 20 has two articulated gripping jaws 21, 21' which form a gripping tong 29.On the outside of the satellite body 10 there is arranged a sensor device 4 which is directed into the space R in the same direction as the gripping device 2 and is designed to detect the object 3 to be gripped. In the example shown, the sensor device 4 is provided with a LIDAR sensor 40. The LIDAR sensor 40 directed into the space R is configured to provide a set of object points as three-dimensional data points in a first coordinate system K 1 as data. The term "three-dimensional data points" does not mean that the data points themselves have a three-dimensional extent, but that the data points are described by three-dimensional coordinates in the first coordinate system K 1. The data thus obtained are sent from the LIDAR sensor 40 via a sensor data transmission device 42, for example a data line or a radio link, to a data processing device 5 provided in the satellite corpus 10. The data processing device 5 is designed to process the data transmitted by the sensor device 4 by means of an artificial neural network in such a way that position parameters for the position and the orientation of the object 3 in the space R are formed therefrom.The data processing device 5 is connected - wirelessly or by cable - to a control device 6 for data transmission provided in the satellite corpus 10 by a position parameter transmission device 50. The control device 6 serves for controlling actuators of the gripping device 2, not shown in FIG. 1, for which purpose the control device 6 is designed to generate control commands for controlling the gripping device 2 from the position parameters for the position and the orientation of the object 3 transmitted by the data processing device 5 and to transmit these to the gripping device 2 in order to approach a target configuration for the position and the orientation of the gripping device 2 in the space R for gripping the object 3 and to grip the object 3 there.The data processing device 5 is specifically designed to process the data acquired by the sensor device 4 as a set of object points of the object 3 in the three-dimensional space R in the manner and sequence listed below in order to obtain the position parameters for the position and the orientation of the object 3 in the space R which are transferred to the control device 6:encoding the set of object points in the three-dimensional space R into a data vector of ordered components;executing a sequence of operators to successively reduce the data vector in its dimension, wherein the operators are respectively composed◯ from a linear transformation and◯ a subsequent nonlinear function applied to each individual component of the previously generated data vector, so that in each case a component of a new data vector is again produced;linear transformation to a multidimensional data vector whose components contain parameters for describing the spatial position of the object 3 in the first coordinate system K 1 used by the sensor device, andconverting these parameters for the object position into new position parameters which describe the spatial position of the target object, i.e. of the object 3, in a second coordinate system K 2 used by the control device 6.The position of an object in any coordinate system, here for example in the first coordinate system K 1 of the sensor device, is equivalent to a transformation of the position data of the object from a model coordinate system in which the model geometry is described into this target coordinate system. This transformation "moves" the object model into the sensor data world.FIG. 2 schematically shows the principle of orientation parameterization used here for eliminating multi-significances and discontinuities of the position parameters in the case of rotational symmetry of the object in a simplified two-dimensional representation in the drawing surface of FIG. 2.In the first row a, a square body 100 is shown in multiple orientations (A, B, C, D) which it assumes during a counterclockwise rotation about a central axis x orthogonal to the plane of the drawing. In the first position A, the square body 100 stands upright, its base runs horizontally and the sides run vertically. In the second position B, the body 100 is rotated by approximately 30°. In the third position C, it is rotated by approximately 60° and in the fourth position D, it is rotated by 90°. In the fourth position D, the body 100 appears to a viewer who views only the outline of the body in the same position as at the beginning (position A) due to the symmetry of a square.In the second row b, an orientation or direction vector 101 assigned to the body 100 is shown, which rotates at the same angular speed as when the body 100 is rotated in the first row. This direction vector 101 corresponds to two orientation parameters of the position of the body 100 in space, namely the two coordinates of the direction vector. In the fourth position D, the direction vector 101 is rotated through 90° with respect to the first position. For the same view of the outline of the body 100 in the positions A and D, there are thus two different alignments of the direction vector 101 (positions A and D) during the parameterization of the second row. When the body 100 is rotated completely through 360°, this results in four alignments of the direction vector (90°, 180°, 270° and 360°), although the outline view of the body 100 is the same in all four alignments. This fourfold superiority of the two orientation parameters corresponds to fourfold symmetry of the square body 100. Each parameter value in each orientation parameterization is ambiguous here due to symmetry four times. However, it is necessary that there is only one parameter value for each view.In the third row c, the specific parameterization preferably used in the robotic device 1 according to the invention is schematically depicted. Here, although the direction vector 102 representing the position of the body 100 in space rotates in the same direction (counterclockwise) as the rotating body, the direction vector 102 thereby assumes four times the rotational speed with respect to the body 100. This has the result that after a rotation of the body through 90°, i.e. in position D, the direction vector 102 again points in the same direction as in position A. If the quadratic, i.e. four-fold symmetrical, body 100 assumes the same orientation in space with respect to the outline view of the body 100, the direction vector 102 always points in the same direction. According to this procedure according to the invention, the rotational speed is thus multiplied by the value n of n-fold symmetry.The orientation of the square is parameterized in both variants, the variant of row b and the variant of row c, not by the angle of the orientation or direction vector to any axis (for example by the polar angle), but by the two components of this vector.For objects with discrete rotational symmetry, such as the square shown in row a, or generally formulated with corresponding orientation ambiguity in the sensor data, the object location in each parameterization is ambiguous, i.e., each parameter value corresponds to multiple physically distinguishable orientations.In FIG. 2, the point is not uniqueness, but continuity of parameterization. The parameterization with the orientation or direction vector 101 or 102 is unique both in the variant of the row b and in the variant of the row c, because the parameterization is unique also in the middle row b, since the permitted parameter range is limited to the first quadrant. The parameterization in the middle row b is only not continuous, because a jump in direction takes place at the two edges of this region, at which the square passes through the same object view. This jump is avoided in the variant of the third row c. Therefore, the third variant of the row c is continuous. The third row c shows the principle of the expanded parameters.If the body had, for example, an outline in the form of an isosceles triangle, there would be triple symmetry and the direction vector would have to rotate at triple the speed of the body.The directions in the three-dimensional space of two orthogonal unit vectors in the object model are parameterized by their Cartesian coordinates in the first coordinate system used by the sensor device, that is to say in the present case a 6-dimensional parameterization.In the case of a discrete rotational symmetry of the target object or a corresponding ambiguity in its orientation in the object points generated by the sensor device, one of these unit vectors is selected orthogonally to the axis of symmetry or to the axis of the orientation ambiguity, while the other is selected either orthogonally to this axis or along this axis. The symmetry or ambiguity is taken into account as follows: In n-fold rotational symmetry or orientation ambiguity about this axis, the unit vector that is orthogonal to this axis is or both unit vectors that are orthogonal to this axis are not firmly anchored in the object, but rather rotate n-fold an object rotation about this axis.Reference numerals in the claims, the description and the drawings are used merely for better understanding of the invention and are not intended to limit the scope of protection.List of reference charactersReference numerals 1 robotic device 2 controllable gripping device 3 object 4 sensor device 5 data processing device 6 control device 10 satellite body 20 gripper 21 gripping jaw 21' gripping jaw 22 articulated arm 22' robot gripping arm 23 first joint 24 first articulated arm rod 25 second joint 26 second articulated arm rod 27 third joint 28 third articulated arm rod 29 gripping tongs 40 LIDAR sensor 42 sensor data transmission device 50 position parameter transmission device 100 object 101 direction vector 102 direction vector x center axis of 100 K1 first coordinate system K2 second coordinate system R spaceReferences included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Patent Literature citedDE 10 2018 105 544 A1

[0007]

Claims

Robotic device (1) having - at least one controllable gripping device (2) for gripping an object (3) in space (R) with a gripper (20); - at least one sensor device (4) which is designed to detect the object (3) to be gripped and then to provide a set of object points in the form of three-dimensional data points in a first coordinate system (K1) as data; - at least one data processing device (5) which is designed to process data provided by the sensor device (4) and to calculate position parameters for the position and the orientation of the object (3) in space (R) therefrom, wherein the data are processed by means of an artificial neural network, and - at least one control device (6) for controlling the at least one gripping device (2), wherein the control device (6) is designed to, in order to generate control commands for controlling the gripping device (2) from the position parameters for the position and the orientation of the object (3) calculated by the data processing device (5) and to transmit these to the gripping device (2) in order to approach a target configuration of the gripping device (2) in the space (R) for gripping the object (3) and to grip the object (3) there, characterized in that the data processing device (5) is designed to process the data recorded by the sensor device (4) as a set of object points in the manner and sequence listed below, in order to obtain the position parameters for the position and the orientation of the object (3) that are transferred to the control device (6): - coding the set of object points as three-dimensional data points in the first coordinate system (K1) into a data vector comprising ordered components; - executing a sequence of operators in order to successively reduce the data vector in its dimension, wherein the operators are each composed ◯ of a linear transformation and ◯ of a subsequent nonlinear function applied to each individual component of the previously generated data vector, so that in each case a component of a new data vector is again produced; linear transformation to a multidimensional data vector whose components contain predicted parameters for describing the spatial position of the target object in the first coordinate system (K1) used by the sensor device (4), and conversion of the predicted parameters for the object position into new position parameters which describe the spatial position of the target object in a second coordinate system (K2) used by the control device (6).Robotic device according to claim 1, characterised in that the data processing device (5) is designed such that the predicted parameters for the object position are subjected to an iterative correction before the conversion into the new position parameters of the second coordinate system (K2) in order to reduce an error measure of the object position.Robotic device according to claim 1 or 2, characterised in that the data processing device (5) is configured such that the determination of the predicted parameters for the object position is carried out repeatedly several times in time sequences before the conversion into the new position parameters of the second coordinate system (K2) in order to track a moving object (3).Robotic device according to one of the preceding claims, characterized in that the data processing device (5) is designed such that the parameters predicted by the data processing device for describing the spatial orientation of the target object in the first coordinate system used by the sensor device are selected such that the parameterization is unique and continuous for each parameter value.Robotic device according to one of the preceding claims, characterized in that the data processing device (5) is designed in such a way that the multidimensional data vector generated by the linear transformation is a nine-dimensional data vector.Robotic device according to one of the preceding claims, characterized in that the data processing device (5) is designed in such a way that the new position parameters of the second coordinate system (K2) form a rotation matrix and a translation vector.Robotic device according to one of claims 1 to 5, characterised in that the data processing device (5) is designed such that the new position parameters of the second coordinate system (K2) are formed by Euler angles and / or quaternions.Robotic device according to one of the preceding claims, characterized in that the gripping device (2) has at least one robot gripping arm (22') or is designed as a robot gripping arm (22') and is provided with the gripper (20).Robotic device according to one of the preceding claims, characterized in that the sensor device (4) has at least one LIDAR sensor (40) or is designed as a LIDAR sensor (40).Space satellite formed as or equipped with a robotic device (1) according to any of the preceding claims, wherein the object (3) to be gripped is a space object, for example a dedicated satellite or a piece of space scrap.

Images