Predicting likelihood of collision of robot on trajectory, planning trajectory of robot, operating robot, and training data processing for determining distance values
By using machine learning methods to predict the probability of collisions on the robot's trajectory, the problem of inaccurate prediction in existing technologies is solved, and the accuracy and safety of trajectory planning and operation are improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-20
- Publication Date
- 2026-04-10
AI Technical Summary
Existing technologies struggle to accurately predict the likelihood of a robot colliding with its trajectory, resulting in inaccurate trajectory planning and operation.
By employing at least part of machine learning-based data processing, the collision probability is predicted by determining the boundary distance of the virtual space region filled by the trajectory of the robot model from the starting pose to the ending pose. The machine learning model is used to train the data processing to represent the virtual space region, thereby predicting the collision between the robot and the object.
It enables more accurate prediction of collision probability on the robot's trajectory, improves the accuracy and safety of trajectory planning, and simplifies the robot operation process.
Smart Images
Figure FT_1 
Figure FT_2
Abstract
Description
Technical Field
[0001] The present invention relates to a method for predicting the probability of a robot collision on a trajectory, a method for planning a robot trajectory and / or manipulating a robot based on the predicted robot collision probability, a method for training at least partly machine learning-based data processing for determining distance values depending on the distance between an object and the boundary of a virtual space region filled by a trajectory traversed by a robot model (Abbild) from a starting pose to an ending pose, and a system or computer program or computer program product for performing at least one of the methods. Background Technology
[0002] By predicting the likelihood of a robot colliding with its trajectory, it is possible to improve the robot's trajectory planning and / or operation. Summary of the Invention
[0003] Accordingly, the object of the present invention may be to improve the prediction of the probability of collision of a robot on a trajectory, the trajectory planning of a robot, the operation of a robot, and / or the data processing for determining a distance value that depends on the distance between the object and the boundary of a virtual space region filled by the trajectory traversed by the robot model from the starting pose to the ending pose.
[0004] The object of the invention is achieved by a method having the features described in claims 1, 11, 12, or 13. Claims 14 and 15 protect a system, computer program, or computer program product for performing the methods described herein. Dependent claims relate to advantageous extensions.
[0005] According to one embodiment of the present invention, a method for predicting the probability of a robot colliding with a trajectory or on a trajectory includes the following steps: - Provides the robot's starting pose; - Provide the robot's end pose, which in a preferred embodiment is the same as the start pose, particularly by providing the start pose, or in a particularly preferred embodiment the end pose may be different from the start pose; - Provide at least one object to be checked for the probability of collision with the robot, which is referred to herein, without loss of generality, as the first object; and - based on - The provided starting pose - The provided endpoint pose, and - The first object provided, Through data processing based at least partially on machine learning, a distance value is determined, which, without loss of generality, is referred to as the first distance value and depends on... - First object and - Boundary of a virtual space region The distance between these virtual spatial regions is filled, traversed, or swept (overall) by the robot's model, particularly its data model (virtually or simulatedly), along a preferably pre-defined trajectory from the starting pose to the ending pose, or is defined therein, particularly as the "swept volume". Specifically, based on the first distance value, the probability of the robot colliding with the trajectory is predicted, in particular whether (it is very likely) the robot will collide, especially with the first object, or whether (it is very likely) the robot will not collide, especially with the first object.
[0006] The embodiments of the present invention are based on the idea that virtual spatial regions or "sweep volumes" are implicitly or indirectly represented by data processing based at least partially on machine learning. These virtual spatial regions or "sweep volumes" are filled, traversed, occupied, or defined by a robot model, particularly a data model (virtually or simulatedly), traversing a trajectory from a starting pose to an ending pose. The likelihood of collisions between the robot and objects on this trajectory is then predicted using these virtual spatial regions or "sweep volumes." Advantageously, different starting and ending poses, as well as correspondingly different trajectories and virtual spatial regions, are represented by the same, correspondingly trained data processing, and the likelihood of collisions with different objects can be examined.
[0007] Accordingly, in an advantageous extension, the method includes the following steps: - Provide one or more other objects to check for the possibility of collision with the robot; and - (each) using at least part of machine learning-based data processing, based on the provided starting pose, the provided ending pose, and the other provided objects, another distance value is determined, preferably in the same manner as the first distance value which depends on the distance between the first object and the boundary, depending on the distance between the other objects which are defined in the same manner as the distance between the first object and the boundary and the boundary of the virtual space region filled by the trajectory driven by the robot's model from the starting pose to the ending pose; Here, the probability of a robot colliding with its trajectory is also predicted based on one or more other distance values, in particular whether a collision with the robot (in particular) is (very likely) to occur, in particular with the corresponding other object, or whether a collision with the robot (in particular) is (very likely) not to occur, in particular with the (corresponding) other first object.
[0008] Preferably, the first distance value and one or more other distance values can be determined sequentially or, particularly, at least partially in parallel on a GPU or similar device. In one embodiment, the data processing, at least partially based on machine learning, can be distributed across multiple computing devices.
[0009] In one implementation, the object to be inspected (the first or the other) comprises a spatial point, preferably two-dimensional or three-dimensional, preferably within the robot's, particularly Cartesian or three-dimensional, workspace, and may be such a spatial point. In one implementation, the distance between the object and the boundary is a shortest distance, and / or an Euclidean distance, and / or a one-dimensional or scalar distance. In one implementation, predicting the probability of a collision with the robot on its trajectory includes predicting whether (it is very likely) a collision will occur between the robot and, in particular, one or more of the objects, or whether (it is very likely) no collision will occur between the robot and, in particular, the one or more objects.
[0010] The determined distance value, in one embodiment and in an extended embodiment, includes a signed and / or continuous or quasi-continuous value, which in particular may include a distance, in one embodiment a directed distance or its (numerical) value, and in another embodiment it may be a distance, in an extended embodiment a directed distance or its (numerical) value. In an extended embodiment, if the distance is lower than a preset limit value, preferably lower than a preset minimum (safe) distance, particularly when the distance is negative, then based on such a distance value, it is predicted that the robot is likely to collide on the trajectory.
[0011] Thus, in one implementation, the probability of a collision can be predicted with particular accuracy and / or reliability.
[0012] Additionally or alternatively, the determined distance value includes, in one extension, a discrete value, and in a particularly preferred extension, a binary value; in one embodiment, the value indicates whether a collision is likely to occur or exist.
[0013] Therefore, in one implementation, data processing can be trained more effectively.
[0014] In one embodiment, the determined distance value is a one-dimensional value, preferably the aforementioned discrete or (quasi-)continuous, particularly signed value. This advantageously reduces computational costs. In another embodiment, the determined distance value is a multi-dimensional value, preferably including the aforementioned discrete and (quasi-)continuous values, particularly signed values. This, in one embodiment, can improve training.
[0015] In one embodiment, the robot includes a robotic arm and / or a mobile base, which may in particular be constituted therefrom. In one embodiment, the robot, particularly its robotic arm and / or its mobile base, has three or more, preferably at least six, and in one embodiment at least seven (motion) degrees of freedom or (motion) axes, particularly translational and / or rotational degrees of freedom or axes, especially joints. The present invention is particularly suitable for such robots due to their complex and / or varied motion possibilities and / or use of boundary conditions, particularly the surrounding environment. In one embodiment, coordinates describing or indicating the robot's (motion) degrees of freedom or (motion) axes are... q The configuration space of the robot. The surface points of the robot or its model in the robot's Cartesian or Euclidean workspace. x With coordinates q Or the robot's configuration space and workspace can be transformed into each other in a known manner through (forward or reverse) transformations. x = T(q) In one implementation, the robot's workspace is a Cartesian or Euclidean space, preferably a three-dimensional space, in which the outer contour of the robot or its model can be located or moved; the robot's configuration space is a space whose dimensions correspond to the robot's (motion) degrees of freedom, wherein each pose of the robot is uniquely determined by a point.
[0016] In one embodiment, the data processing, at least in part based on machine learning, includes at least one artificial neural network, preferably at least one deep neural network, and may in particular be constituted therefrom. Through such data processing, distance values can be determined particularly accurately, simply, quickly, and / or reliably. Accordingly, in one embodiment, the machine learning includes training data processing, preferably training one or more artificial neural networks, which may particularly include supervised or unsupervised machine learning.
[0017] In one implementation, the trajectory from the starting pose to the ending pose is a pre-defined, linear trajectory, preferably linear, preferably within the robot's workspace or particularly preferably within its configuration space. The robot model fills a virtual space region by (virtually or simulatedly) traversing this trajectory. q s Represents the starting point pose, q g Indicates the final pose. t [0, 1] represents the trajectory parameters, then the virtual space region can be represented, for example, by... SV ( q s , q g )= t [0, 1] V ((1-t) q s + t q g This can be described using the term ). Thus, it is particularly simple to pre-set a trajectory and / or virtually or simulate driving along that trajectory, or to define a virtual space accordingly.
[0018] In one implementation, the robot's trajectory from its starting pose to its ending pose is preferably a single or multiple curved, pre-defined trajectory, preferably curved, within the robot's configuration space or particularly preferably within its workspace. In a particularly advantageous extension, this is a spline trajectory, through which the robot model fills a virtual space region by (virtually or simulatedly) traversing the trajectory. The curved trajectory can also have straight sections. This allows for particularly flexible pre-setting of the trajectory or corresponding determination of the virtual space to suit the specific circumstances. Particularly advantageously, the (pre-defined) spline trajectory or other defined curved trajectory within the workspace enables (more)smoother movement of the robot's (end-effector) flange.
[0019] In one implementation, to determine the distance value, one or more, preferably a fixed (preset) number of, intermediate poses of the robot are provided, particularly between the starting pose and the ending pose and / or on the trajectory, and the distance value is also determined based on the provided intermediate poses by means of data processing. Thus, in one implementation, the data processing can be particularly advantageously trained, especially for faster machine learning and / or to provide better results.
[0020] In one implementation, one or more objects to be examined for the likelihood of collision with the robot, particularly a first object and / or one or more other objects, are provided based on the robot's environment as detected, particularly by sensing technology, and in one implementation by means of at least one camera, particularly at least one depth or 3D camera, preferably as a point cloud or the like. Thus, the likelihood of collision between the robot and the environment or obstacles can be predicted particularly advantageously, especially simply, reliably, and / or accurately.
[0021] In one implementation, the robot's workspace provides the possibility of collision between one or more objects to be inspected, particularly a first object, and / or one or more other objects and the robot; in an extended embodiment, this is provided as spatial points. p i 3This allows for the particularly advantageous, and especially simple, reliable, and / or accurate, prediction of the likelihood of a robot colliding with its environment or obstacles.
[0022] In one implementation, in having f Configuration space of a robot with one degree of freedom or (motion) axis f Provides the starting pose q s f and / or endpoint pose q g f Therefore, it is particularly easy to pre-set a trajectory, and / or virtually or simulate driving through a trajectory, or to define a virtual space relatively simply.
[0023] When determining distance values using data processing, in one implementation, additional one-dimensional or multi-dimensional output values are determined; in an advantageous extension, kinematic parameters for the transformation between the robot's workspace and configuration space are determined; in a particularly advantageous implementation, the positions of one or more reference points of the robot in the workspace are determined, in one implementation being the positions of joints (points). Thus, in one implementation, data processing can be particularly advantageous for training, especially for faster machine learning and / or providing better results, particularly by (implicitly) (better) (together) learning this transformation between configuration space and workspace, which affects the determination of distance values when one or more objects are provided in the workspace and when start and end poses are provided in the configuration space.
[0024] According to one embodiment of the present invention, a method for planning the trajectory of a robot includes the following steps: - Providing the robot's trajectory, in one implementation, is achieved through corresponding (initial) trajectory planning, such as based on robot-specific reference points, particularly TCP, robot guidance tools, etc., with preset target positions and / or orientations, and / or based on preset quality criteria such as energy, time, path length, etc.; and - Predict the probability of a robot colliding with the trajectory based on the methods described in this paper; The starting and ending poses, and in one extended scheme, the intermediate poses used to predict the probability of collision are provided based on the provided trajectory. In one embodiment, the provided trajectory has the starting and ending poses (and, if necessary, the intermediate poses), and in one extended scheme, includes a trajectory that fills a virtual space region by driving over it, particularly as a sub-trajectory or trajectory segment. The provided trajectory is evaluated based on the prediction. In one embodiment, if a robot collision is predicted to occur (most likely), particularly with one or more objects provided to the data processing, the provided trajectory is modified or replaced with another alternative trajectory. For the modified or alternative trajectory, the probability of robot collision is preferably predicted again according to the methods described herein, and this process is repeated until a termination criterion is met, such as reaching a preset maximum number of attempts, or it is predicted that a robot collision, particularly with one or more objects provided to the data processing, will not occur (most likely) on the modified or alternative trajectory.
[0025] According to one embodiment of the invention, the probability of a robot colliding with a trajectory is predicted according to the method described herein. In an extended embodiment, the robot's trajectory is planned according to the method described herein, and the robot is operated based on the prediction, particularly the planned trajectory, in particular by having the robot drive over the planned trajectory, or when a collision with the robot is predicted to occur (most likely), particularly with one or more objects provided for data processing, a corresponding avoidance response is executed, such as stopping the robot, performing collision avoidance movements, replanning the trajectory, etc.
[0026] According to one embodiment of the present invention, a method for training at least partially machine learning-based data processing based on a provided endpoint pose, a provided object, the data processing being used to determine a distance value preferably as described herein, the distance value preferably depending on the distance between the object and the boundary of a virtual space region filled by a pre-defined trajectory traversed by a robot model from a starting pose to an endpoint pose, the method comprising the following steps: - Provides multiple (training) datasets, including starting and ending poses, and in one extended scheme, intermediate poses; wherein, for each of these datasets... - Determine the virtual space region filled by the robot model traversing a preferred pre-defined trajectory from the starting pose to the ending pose of the corresponding dataset, and in an extended scheme, also passing through intermediate poses; - Provide multiple objects, specifically at least one object within the virtual space region, and / or at least one object outside the virtual space region, and / or at least one object whose distance to the robot model and / or the boundary of the virtual space region is less than half the maximum extension of the robot model and / or the virtual space region; in other words, near the robot model or near the boundary of the virtual space region; and provide separate [details for each of these objects]. - Determine a distance value, which preferably depends on the distance between the corresponding object and the boundary of the corresponding virtual space region in the manner described herein or in the same manner as when predicting the probability of collision; and - The data processing is trained based on the provided dataset, the provided objects, and the distance values determined therefor, preferably so that the data processing maps the provided dataset and objects to the determined distance values, or assigns the determined distance values to the provided dataset and objects.
[0027] According to one embodiment of the invention, a system is designed, in particular, by hardware and / or software techniques, especially programming techniques, to perform the methods described herein.
[0028] According to one embodiment of the present invention, the system includes: - A device used to provide the starting pose of a robot; - A device used to provide the final pose of a robot; - In one extended scheme, a means for providing at least one intermediate pose; - A means for providing a first object to be checked for the probability of collision with the robot, and, in an extended embodiment, a means for providing another object to be checked for the probability of collision with the robot; and - At least partially based on machine learning data processing, for determining a first distance value based on a provided starting pose, a provided ending pose (and, if necessary, an intermediate pose), a provided first object, and, in an extended scheme, other corresponding distance values for (correspondingly) other objects, the first distance value depending on the distance between the first object and the boundary of a virtual space region filled by a trajectory traversed by the robot's model from the starting pose to the ending pose, and, in an extended scheme, through the intermediate pose; and - A device for predicting the probability of a robot colliding with a trajectory based on a first distance value and, if necessary, one or more other distance values.
[0029] Additionally or alternatively, according to one embodiment of the present invention, the above system includes: - A means for providing multiple datasets of starting pose and ending pose, and in an extended scheme, at least one intermediate pose; - A device for performing the following operations on these datasets respectively. - Determine the virtual space region filled by the trajectory traversed by the robot model from the starting pose to the ending pose in the corresponding dataset, and in an extended scheme, also passing through intermediate poses, and - Provide multiple objects, specifically at least one object within the virtual space region, and / or at least one object outside the virtual space region, and / or at least one object whose distance to the robot and / or the boundary of the virtual space region is less than half the maximum extension of the robot and / or the virtual space region, and - Determine a distance value for each of these objects, which depends on the distance between the corresponding object and the boundary of the corresponding virtual space region; and - A device for training data processing based on a provided dataset, provided objects, and distance values determined for this purpose.
[0030] Additionally or alternatively, according to one embodiment of the present invention, the above system includes: - A device used to provide the robot's trajectory; - A device for predicting the probability of a robot colliding with a trajectory according to the methods described herein; and - A device for evaluating the provided trajectory based on the prediction.
[0031] Additionally or alternatively, according to one embodiment of the present invention, the above system includes: - A device for predicting the probability of a robot colliding with a trajectory according to the methods described herein; and - A device used to operate the robot based on this prediction.
[0032] The systems and / or apparatuses of this invention can be constructed using hardware and / or software technologies, and in particular, include: at least one processing unit, preferably connected to a storage system and / or a bus system via data or signal connections, particularly a digital processing unit, particularly a microprocessor unit (CPU), graphics card (GPU), etc.; and / or one or more programs or program modules. The processing unit may be designed to execute instructions implemented as a program stored in the storage system; acquire input signals from the data bus; and / or send output signals to the data bus. The storage system may have one or more, particularly different, storage media, particularly optical, magnetic, solid-state, and / or other non-volatile media. The program may be designed to implement or execute the methods described herein, enabling the processing unit to perform the steps of these methods, thereby particularly being able to predict the probability of a robot's collision, plan the robot's trajectory, operate the robot, or train data processing at least partially based on machine learning. In one embodiment, the computer program product may include, particularly may be, a computer-readable, non-volatile storage medium for storing programs or instructions, or on which programs or instructions are stored. In one implementation, the program or these instructions are executed by a system or controller, particularly a computer or an array of computers, such that the system or controller, particularly the one or more computers, performs the methods described herein or one or more steps thereof, or the program or instructions are designed for this purpose.
[0033] In one implementation, one or more, particularly all, steps of the method are implemented entirely or partially by a computer, or one or more, particularly all, steps of the method are executed entirely or partially automatically, particularly by the system or its apparatus.
[0034] In one implementation, the system includes a robot.
[0035] In one implementation, at least partly machine learning-based data processing maps the provided start pose, the provided end pose, and the provided (corresponding) object, as well as the provided intermediate poses if necessary, to corresponding distance values and additional output values if necessary. In particular, determining the distance values and additional output values by means of at least partly machine learning-based data processing and based on the provided start pose, the provided end pose, and the provided object, as well as one or more provided intermediate poses if necessary, may include, and in particular, such mapping.
[0036] As already mentioned, the starting pose and the ending pose can also coincide or be the same. Therefore, the trajectory from such a starting pose to such an ending pose is, in the sense of this invention, a zero motion of the robot (model), and the virtual space region filled by the robot model traversing the trajectory from the starting pose to the ending pose is the virtual space region filled by the robot model in that starting / ending pose.
[0037] The provision in the sense of this invention may particularly include, and especially may be: generating; and / or transmitting; and / or receiving or obtaining; and / or invoking, particularly loading from memory or intermediate memory; and / or being ready, in one embodiment being ready in memory or intermediate memory, preferably for (further) use in the methods described herein. Attached Figure Description
[0038] Further advantages and features are given in the dependent claims and embodiments. This section schematically illustrates: Figure 1 A system according to an embodiment of the present invention; and Figure 2 This is a method according to one embodiment of the present invention. Detailed Implementation
[0039] Figure 1 A system according to an embodiment of the present invention is shown, comprising a robot and a controller 30. For clarity, the robot consists of a proximal segment 10 and a distal segment 12, wherein the proximal segment 10 is rotatably mounted in a rotational joint 11, and the distal segment 12 is rotatably mounted on the proximal segment 10 via a rotational joint 13. A model of this robot is shown in... Figure 1 The configuration space of the robot is represented by two spheres covering joints 11 and 13 and two cylinders covering limbs 10 and 12. Accordingly, the configuration space of the robot is represented by its joint coordinates. q = [ q 1, q 2] T The workspace is formed by the opening of a Cartesian space, in which, for clarity, a first object or spatial point is again exemplarily shown in the Cartesian space. p 1 and another object or spatial point p 2.
[0040] Now, in order to plan a route that allows a robot to pass through two spatial points without collision... p 1. p The trajectory of 2, trained with an artificial neural network 31, or using data processing trained in this way.
[0041] For training purposes, a set of point pose datasets and a set of endpoint pose datasets are provided multiple times. q s,i , q g,i}( Figure 2 Step S10), and for each dataset, determine the virtual space region filled by the robot model traveling along a trajectory from the starting pose to the ending pose of the corresponding dataset. Figure 2 Step S20).
[0042] For the purpose of explanation, Figure 1 The text exemplifies how shadows are used to represent such virtual spatial regions, for example...
[0043] - A virtual space region, for: Figure 1 The starting pose is shown in solid lines, with two limbs 10 and 12 extending horizontally; the ending pose is shown, with the distal limb 12 pivoting upwards by 90° while the proximal limb 10 remains horizontally extended; and the straight trajectory is shown. q 1 = 0°, q 2= t 90°] t [0, 1]; - A virtual space region, for: taking the above-mentioned endpoint pose as the starting pose, wherein the distal limb 12 pivots upward by 90° while the proximal limb 10 extends horizontally; the endpoint pose, wherein the proximal limb 10 pivots upward by 45° when there is no relative movement between the distal limb 12 and the proximal limb; and a straight trajectory [ q 1= t 45° q 2=90°], t [0, 1]; and - A virtual space region, for: taking the above-mentioned endpoint pose as the starting pose, wherein the proximal limb 10 pivots upward 45° from the horizontal line, and the distal limb 12 pivots upward 90° relative to the distal limb 12 from the extension direction or the horizontal direction; the endpoint pose, wherein the proximal limb 10 further pivots upward 45°, and the distal limb 12 pivots or rotates -90° relative to the proximal limb 10; and a straight line trajectory [ q 1= t 45° q 2 = (1 - t ) 90°] t [0, 1].
[0044] Provide multiple objects in the form of spatial points, and assign a distance value to each of these objects. Figure 2 (Step S30) The distance value includes a signed continuous value of the distance between the corresponding object and the boundary of the corresponding virtual space region and / or a binary value indicating whether the corresponding object is located inside the virtual space region.
[0045] For the purpose of explanation, Figure 1 Chun Chun exemplarily illustrates two such spatial points. p 1. p 2.
[0046] During the provision of more datasets and objects and the determination of distance values for them, and / or after the provision of multiple datasets and objects and the determination of distance values for them, the data processing is trained based on these provided datasets and objects and the distance values determined for them as follows: 31 Figure 2 Step S40): The data processing maps the start pose, end pose, and spatial point to distance values: [ q s , q g , p ] →{ dist , flag},in, dist Is it a signed distance, a binary value? flag ={collision; no collision}. In one variant, one of these two distance value components can also be omitted.
[0047] After data processing 31 was trained in this way ( Figure 2 Step S50: “Y”), now perform path planning. To this end, an initial trajectory is provided in step S100, and in step S110, the initial trajectory is subdivided into shorter sub-trajectories. For each of these sub-trajectories, the probability of collision of the robot on that sub-trajectory is predicted by means of the trained data processing 31. Figure 2 Step S120), here or for this purpose, provides the start and end points of the corresponding sub-trajectories as the start pose and end pose, respectively. Figure 2 Step S112), and provide, for example, the spatial points described above. p 1. p 2 as the first object or other objects ( Figure 2 Step S114), and using the trained data processing, determine the corresponding first distance value or other distance value ( Figure 2 Step S116), and if there is a signed distance distLess than zero or less than a pre-defined positive minimum distance between the boundary and a spatial point located outside the spatial region, or flag = "collision", then it is predicted that the robot is very likely to collide with the corresponding spatial point.
[0048] If, for at least one sub-trajectory prediction, the robot is likely to collide with one of the spatial points ( Figure 2 Step S130: If "Y" is selected, the initially provided trajectory is evaluated as unsuitable and modified or replaced with another available trajectory. Figure 2 Step S140), and repeat steps S110 to 130 or 140 until a trajectory is found for which it is predicted that the robot is unlikely to collide with any spatial points (for all its sub-trajectories). Figure 2 Step S130: “N”).
[0049] Subsequently, the robot drove across the trajectory. Figure 2 (Step S150). Also for illustrative purposes, in... Figure 1 The example illustrates a collision-free trajectory along which the robot passes two spatial points. p 1, p 2. By having a model of the robot traverse the trajectory, the model fills in... Figure 1 The virtual spatial region SV is represented by shaded areas. It can be seen that there are two spatial points... p 1. p 2 are positively distanced from the boundary G of the spatial region SV, or in other words, they are located outside the spatial region SV.
[0050] This embodiment illustrates that by utilizing data processing 31 trained on different datasets and spatial points based on the starting and ending poses, distance values can be determined for different (sub)trajectories and potential obstacles with different spatial point forms. This allows for simple, fast, and reliable checking for collisions on different trajectories.
[0051] It needs to be emphasized again that this embodiment is for illustrative purposes only and is therefore greatly simplified. In practice, there are often more complex robots, and / or datasets using significantly more spatial points and start-point and end-point poses for training data processing. It is preferable to use start-point and end-point poses distributed throughout the robot's entire configuration space, as well as spatial points located near the boundaries of virtual spatial regions of the (corresponding) trajectories between the (corresponding) start-point and end-point poses, in addition to those of the robot (model). For clarity, Figure 1 Two spatial points p 1. p2 is used not only to illustrate training but also to illustrate the planned trajectory. In addition, the aforementioned subdivision of the trajectory and checking for collisions in the sub-trajectories can be omitted, and / or the start pose and end pose of one or more poses or trajectory points to be checked can be completely identical (übereinstimmen), so that the virtual space region corresponds to the virtual space filled by the robot model under that start / end pose.
[0052] In this disclosure, "having X" is not generally intended to be an exhaustive list, but rather a shorthand for "having at least one X," and also includes "having two or more X" and "having Y in addition to X." Although exemplary embodiments have been set forth in the foregoing description, it should be noted that many variations may exist.
[0053] For example, trajectory planning and / or training can be performed (separately) on separate computing devices, or the planned trajectory can be provided to the controller, and / or the trained data processing can be provided to the trajectory planner.
[0054] Furthermore, it should be noted that this exemplary embodiment is merely an example and should not be construed as limiting the scope of protection, application, or construction. Rather, the foregoing description provides guidance for those skilled in the art to make modifications to at least one exemplary embodiment, wherein various changes, particularly concerning the function and arrangement of the components, can be implemented without departing from the scope of the invention, for example, as can be obtained according to the claims and their equivalent combinations of features.
[0055] List of reference numerals
[0056] 10 Proximal side robotic limbs
[0057] 11. Rotate the joint
[0058] 12 distal robotic segments
[0059] 13. Rotate the joint
[0060] 30 Controllers
[0061] 31 Data Processing
[0062] p 1. Spatial point (first object)
[0063] p 2. Spatial point (another object)
[0064] SV Virtual Space Area
[0065] The boundary of the G virtual space region
[0066] q1, q2 Joint coordinates.
Claims
1. A method for predicting the probability of a robot colliding on a trajectory, wherein the method includes the following steps: - Provide the starting pose of the robot as described in (S112); - Provide (S112) the final pose of the robot; - Provide (S114) a first object to be checked for the possibility of collision with the robot; as well as - Using at least part of machine learning-based data processing, based on the provided starting pose, the provided ending pose and the provided first object, determine (S116) a first distance value, the first distance value depending on the distance between the first object and the boundary of a virtual space region filled by the trajectory traversed by the model of the robot from the starting pose to the ending pose. Specifically, the probability of the robot colliding with the trajectory is predicted (S120) based on the first distance value.
2. The method according to claim 1, characterized in that, Includes the following steps: - Provide (S114) at least one other object to be checked for the possibility of collision with the robot; as well as - Using the data processing at least in part based on machine learning, based on the provided starting pose, the provided ending pose and the provided other objects, determine (S116) other distance values, which depend on the distance between the other objects and the boundary of the virtual space region filled by the trajectory driven by the model of the robot from the starting pose to the ending pose. Furthermore, the probability of the robot colliding with the trajectory is predicted (S120) based on the other distance values.
3. The method according to any one of the preceding claims, characterized in that, The determined distance values include discrete, and in particular binary, values.
4. The method according to any one of the preceding claims, characterized in that, The determined distance values include signed values.
5. The method according to any one of the preceding claims, characterized in that, The trajectory of the robot from the starting pose to the ending pose is a straight line.
6. The method according to any one of the preceding claims, characterized in that, The trajectory of the robot from the starting pose to the ending pose is a curved trajectory.
7. The method according to any one of the preceding claims, characterized in that, To determine the distance value, at least one intermediate pose of the robot is provided, and the distance value is also determined based on the provided intermediate pose by means of the data processing.
8. The method according to any one of the preceding claims, characterized in that, Based on the robot's environment detected, particularly by sensing technology, and / or providing at least one object in the robot's workspace to check the possibility of collision with the robot, and / or providing the starting pose and / or ending pose in the robot's configuration space.
9. The method according to any one of the preceding claims, characterized in that, When determining distance values using the data processing, additional output values are determined, particularly kinematic parameters for the transformation between the robot's workspace and configuration space.
10. The method according to any one of the preceding claims, characterized in that, The data processing is used to train, in particular, to be trained, according to the method described in claim 13.
11. A method for planning the trajectory of a robot, wherein, The method includes the following steps: - Provide the trajectory of the robot as described in (S100); and - The method according to any one of the preceding claims predicts (S120) the probability of collision of the robot on the trajectory; Specifically, the starting pose and ending pose are provided based on the provided trajectory, and the provided trajectory is evaluated based on the prediction (S140).
12. A method for operating a robot, wherein, The method according to any one of the preceding claims is used to predict the probability of the robot colliding on the trajectory, and the robot is operated based on the prediction (S150).
13. A method for training at least partially machine learning-based data processing based on a provided starting pose, a provided ending pose, and a provided object, the data processing being used to determine distance values depending on the distance between the object and the boundary of a virtual space region filled by a trajectory traversed by a robot model from the starting pose to the ending pose, particularly for predicting the probability of collision of the robot on the trajectory according to any one of the preceding claims, wherein, The method includes the following steps: - Provide multiple datasets for the starting pose and the ending pose as described in (S10); wherein, for each dataset... - Determine (S20) the virtual space region filled by the trajectory of the robot model traversing from the starting pose to the ending pose of the corresponding dataset; - Provide (S30) a plurality of objects, specifically at least one object within the virtual space region, and / or at least one object outside the virtual space region, and / or at least one object whose distance to the boundary of the robot and / or the virtual space region is less than half the maximum extension of the robot and / or the virtual space region; and provide for each of the plurality of objects respectively - Determine (S30) a distance value that depends on the distance between the corresponding object and the boundary of the corresponding virtual space region; and - The data processing is trained (S40) based on the provided dataset, the provided objects, and the distance values determined for this purpose.
14. A system designed to perform the method according to any one of the preceding claims, and / or comprising: - A device used to provide the starting pose of a robot; - A device for providing the final pose of the robot; - A means for providing a first object to check the possibility of collision with the robot; as well as - At least partly machine learning-based data processing for determining a first distance value based on a provided start pose, a provided end pose, and a provided first object, the first distance value depending on the distance between the first object and the boundary of a virtual space region filled by a trajectory traversed by the robot's model from the start pose to the end pose; as well as - A means for predicting the probability of the robot colliding with the trajectory based on the first distance value; And / or include: - A means for providing multiple datasets of the starting pose and the ending pose; - A means for defining a virtual space region for each of the plurality of datasets, providing a plurality of objects, and determining a distance value for each of the plurality of objects, wherein the virtual space region is filled by a trajectory traversed by a model of the robot from a starting pose to an ending pose of the corresponding dataset; the plurality of objects are in particular at least one object within the virtual space region and / or at least one object outside the virtual space region and / or at least one object whose distance to the robot and / or the boundary of the virtual space region is less than half the maximum extension of the robot and / or the virtual space region; the distance value depends on the distance between the corresponding object and the boundary of the corresponding virtual space region; and - A means for training the data processing based on the provided dataset, the provided objects, and the distance values determined therefor.
15. A computer program or computer program product, wherein, The computer program or computer program product comprises instructions, particularly stored on a computer-readable and / or non-volatile storage medium, which, when executed by one or more computers or the system according to claim 14, cause the one or more computers or the system to perform the method according to any one of claims 1 to 13.