METHOD AND DEVICE FOR COLLISION-FREE MOTION PLANNING OF A MANIPULATOR
Patent Information
- Application Number
- DE502019012378
- Authority / Receiving Office
- DE · DE
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2018-09-13
- Filing Date
- 2019-09-05
- Publication Date
- 2024-11-21
- Estimated Expiration
- 2039-09-05
Description
[0001] The present invention relates to a method for collision-free motion planning of a manipulator from a starting point to a target point and to a corresponding device comprising at least one manipulator and a control unit which executes the method.
[0002] Collision-free motion planning in robotics is the task of finding a motion path for a robot or manipulator from a starting point to an endpoint without colliding with an obstacle or with itself during the movement. A method for collision-free motion planning is known, for example, from DE 10 2010 007 458 A1.
[0003] DE 10 2010 007 458 A1 discloses a collision-free motion planning system for an industrial robot with a single robot arm (manipulator), such as those used in manufacturing plants. Industrial robots are machines used for handling, assembling, or processing workpieces. Typically, an industrial robot comprises at least one manipulator with multiple axes of motion and a programmable logic controller (PLC) that coordinates the robot's movements during operation by controlling its drives accordingly.
[0004] The use of robots in close proximity to humans requires motion planning methods capable of calculating collision-free movements with minimal delay. In particular, these planning methods must be adaptable to a constantly changing environment to account for moving obstacles. This necessitates replanning or adaptation of the plan during the robot's movement.
[0005] Continuous replanning of a robot's motion becomes difficult with established global path planning methods, such as those based on dynamic roadmaps (DRM), when the robot's dynamics cannot be adequately considered. These methods typically perform initial path planning without considering robot dynamics, only incorporating them during subsequent time parameterization of a path. The execution of the time-parameterized path is then monitored for potential collisions with moving obstacles, and the robot is safely stopped in the event of a collision. From this safe position, replanning is initiated, allowing the robot to adapt its path to the new conditions. However, moving obstacles in the environment often lead to frequent stops and consequently, slow and unnatural robot movements.
[0006] One way to more effectively account for the robot's dynamics is kinodynamic motion planning, as exemplified in "RRTx: Real-time Motion Planning / Replanning for Environments with Unpredictable Obstacles" (Michael Otte). However, kinodynamic motion planning is based on probabilistic methods, which have the disadvantage of not generating deterministic trajectories. While the required computation time can be limited, it is generally unpredictable. Deterministic methods, such as DRM-based approaches, are unsuitable for kinodynamic motion planning.
[0007] Furthermore, local reactive methods, such as artificial potential fields, are known as an alternative for changing environments. Instead of generating complete trajectories, these methods generate control impulses for the robot's movement. The control impulses "push" the robot away from obstacles, preferably towards a defined target. With these methods, there is no separation between planning and control; instead, the robot reacts directly to obstacles. This allows for particularly fast responses to moving obstacles. However, a problem is that these approaches, based on local optimization, can get stuck in local minima and are therefore of limited use for high-dimensional robots in complex environments.
[0008] ABRAHAM SÄNCHEZ ET AL, A Reactive Lazy PRM Approach for Nonholonomic Motion Planning, ADVANCES IN ARTIFICIAL INTELLIGENCE - IBERAMIA-SBIA 2006 LECTURE NOTES IN COMPUTER SCIENCE; LECTURE NOTES IN ARTIFICIAL INTELLIGENCE; LNCS, SPRINGER, BERLIN, DE, PAGE(S) 542 - 551, (20060101), ISBN 978-3-540-45462-5, describe the combination of a PRM planning approach with a reactive planner.
[0009] QUINLAN S ET AL, "Elastic bands: connecting path planning and control", PROCEEDINGS OF THE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION ATLANTA, MAY 2-6, 1993; [PROCEEDINGS OF THE INTERNATIONAL CONFERENCE ON ROBOTICS AND AUTOMATION], LOS ALAMITOS, IEEE COMP. SOC. PRESS, US, (19930502), vol. CONF. 10, doi:10.1109 / ROBOT.1993.291936, ISBN 978-0-8186-3450-5, also describe approaches to combining local (reactive) planning and global planning.
[0010] DIMITROV V ET AL. "Hierarchical Navigation Architecture and Robotic Arm Controller for a Sample Return Rover", 2013 IEEE INTERNATIONAL CONFERENCE ON SYSTEMS, MAN, AND CYBERNETICS describe a navigation system for a rover based on local and global planning, as well as a controller for controlling a manipulator (robot arm) of the rover.
[0011] It is therefore an object of the present invention to provide an optimized motion planning system for a robot in a changing environment, which avoids the aforementioned disadvantages. In particular, it is an object to provide a motion planning system that takes into account the dynamics of a robot and enables continuous replanning without stopping or reducing speed.
[0012] According to one aspect of the present invention, this problem is solved by a method for collision-free motion planning of a first manipulator of a robot from a starting point to a target point, comprising the steps: Determining a target path for the first manipulator to the target point using a global planner; continuously determining the movement of the first manipulator using a local planner based on the target path of the global planner; executing the movement by the first manipulator in parallel with the determination by the global and local planners. where the global planner performs a replanning at least once and the local planner determines the movement of the first manipulator based on the replanned target path.
[0013] The present invention thus combines repeated global planning with continuous local planning for the motion planning of a robot manipulator. "Continuous" in this context means that replanning takes place continuously, preferably at defined intervals. "Repeated," on the other hand, means that the global planner performs replanning at least once. In particular, the repeated determination of the target path is initiated without any apparent trigger; that is, a re-determination of the target path is initially performed regardless of whether the current target path is still valid or executable. In contrast to known methods, replanning is thus deliberately performed even when the environment has obviously not changed for the local planner. A replanning by the global planner therefore does not need to be "triggered" by a local planner.In this way, global planning can find a shorter target path if, for example, an obstacle disappears from the previously planned target path.
[0014] According to the invention, two different planners are provided, a local and a global one, which are executed in parallel, complement each other, and, in particular, take each other into account during their respective planning. In this context, "parallel" means temporally parallel, i.e., the planners are processed computationally in parallel, e.g., in threads.
[0015] The global planner generates a globally planned path, and the local planner, which can generate rapid control actions, avoids collisions with obstacles while taking the globally planned path into account. The robot's manipulator is thus guided along the globally planned path and simultaneously "pushed away" from obstacles. Both planners, global and local, are executed multiple times in a loop and preferentially utilize each other's results.
[0016] The combination of both planners has the advantage that, on the one hand, the problem of local minima can be avoided, since the local planner is "guided" by the global planner, and on the other hand, the global planning becomes more "dynamic" because the local planner optimizes the target path locally. In particular, the global planning can still be essentially a purely kinematic planning process, and the otherwise usual path smoothing and time parameterization are replaced by path following control using the local planner.
[0017] Through continuous planning with the local planner and supplementary replanning with the global planner, motion planning can be optimally adapted to a changing and complex environment, as the disadvantages of one method are offset by the advantages of the other. In particular, the robot's dynamics can be adequately considered during planning, enabling replanning without stopping or reducing the robot's speed, as well as seamless adaptation of movement to changing environments. This allows for anticipatory avoidance of moving obstacles and the use of shorter paths if obstacles are removed. The task mentioned at the outset is thus completely solved.
[0018] In a preferred configuration, the global planner continuously determines the current target path. Because the global planner plans continuously rather than repeatedly, a complex, dynamic environment can be reliably considered, as a new target path is constantly being determined. Even with continuous determination, the planner preferably executes a new plan initially without any apparent motivation, i.e., regardless of the validity of the current target path.
[0019] In a further embodiment, the global planner is a deterministic planner, specifically based on a dynamic roadmap (DRM). Deterministic planners have the advantage over probabilistic methods of providing a deterministic trajectory, i.e., a unique and reproducible solution within a defined timeframe. With probabilistic methods, the required computation time is generally also limitable, but regularly unpredictable. Since combining the global planner with a local planner does not impose any requirements on the type of global planner, using a deterministic planner advantageously enables rapid deterministic replanning of collision-free movements in dynamically changing environments. Determining the target path by the global planner based on a dynamic roadmap (DRM) is particularly preferred.Global planning with a DRM-based planner is particularly advantageous because it enables fast and runtime deterministic global path planning.
[0020] In a further refinement, the global planner queries the local planner for a result during replanning and incorporates this into determining the target path. The global planner can thus consider results from the parallel local planning at any interval. Specifically, the global planner can query the local planner for a predicted, current starting state and use it as the new starting state during replanning. Replanning therefore does not need to begin from a rest position, as is generally the case. In other words, the robot's dynamics are taken into account, so the robot does not have to stop to perform replanning. This refinement thus contributes to smoother and more natural robot movement.
[0021] In a further implementation, the local planner is based on model predictive control (MPC), and the motion is determined through optimization by solving a dynamic optimization problem. Optimization-based methods are particularly well-suited for motion planning to find paths that enable smooth and natural movement of a robot or its manipulator. The paths are deterministic and therefore reproducible and can be determined in a finite computation time. At the same time, a planner based on model predictive control can comprehensively account for a multitude of variable parameters and thus, in particular, the robot's dynamics.
[0022] In a preferred embodiment, the dynamic optimization problem includes a cost functional that weights deviations of the manipulator from the target path as well as deviations of a current path parameter from a target parameter defined by the target path. By considering deviations from a target path and a current path parameter during planning, the dynamic optimization problem advantageously links the local planner with the global planner. The target path is thus an integral part of the dynamic optimization problem, and the global planner "guides" the local planner, which typically operates with shorter cycle times.
[0023] In a further embodiment, the dynamic optimization problem also includes path dynamics and path input constraints, where the path dynamics are based in particular on the target path of the global planner. By additionally considering path dynamics, especially the path dynamics of the target path, the local planner can further improve the robot's natural movement.
[0024] In a further embodiment, the dynamic optimization problem also includes at least one inequality constraint for a collision distance. Additional inequality constraints can define a minimum distance to obstacles, which must be considered when solving the optimization problem. This helps avoid collisions with obstacles. At the same time, the manipulator does not have to strictly adhere to the predetermined path but can deviate from it depending on the situation, for example, to reach the goal faster. This makes the robot's movement more natural in relation to the changing environment.
[0025] In a further embodiment, the execution of the determination by the global planner can be configured via a defined condition. The execution of the global planner can thus be linked to conditions, enabling the planning to react to special features and exceptional cases in global planning, preferably automatically. In this way, more robust motion planning can be achieved.
[0026] It is particularly preferred that the defined condition allows the global planner to set an unconditional, no, or conditional execution of the determination. This gradation into unconditional, no, or conditional execution has proven advantageous because these settings, preferably dynamically selectable, allow for an appropriate response to the significant special cases that can occur during global planning.
[0027] Furthermore, conditional execution is preferably dependent on whether a determined target path and / or a determined movement is valid within a defined planning horizon, and on whether the distance between a current state and a target state has decreased during a last sampling step. The failure to find a valid path or the lack of progress in planning are special cases that can be addressed by conditional execution. This makes motion planning more robust.
[0028] In a further embodiment, the movement of the first manipulator is paused if planning by the global planner fails. Alternatively, the movement of the first manipulator is continued solely based on the local planner if planning by the global planner fails. Pausing or continuing planning exclusively via the local planner are two alternative approaches that can be applied in the inventive combination of two planners in response to special cases of the global planner. This further increases the robustness of the motion planning.
[0029] In a further implementation, the global and / or local planner partially predicts the movement of an obstacle and / or artificially enlarges a representation of an obstacle. By artificially enlarging obstacles or predicting their movement, planners are encouraged to replan in advance to avoid collisions. This approach can, in particular, increase error tolerance to adequately account for uncertainties in estimation.
[0030] In a further embodiment, the method also includes motion planning for at least one second manipulator. This additional motion planning occurs in parallel and is coordinated with the collision-free motion planning of the first manipulator. The method allows for the simple execution of motion planning for multiple manipulators, for example, on a two-arm robot. For instance, the global planner can have its own DRM (Directional Resource Model) for each manipulator to determine a coordinated target path for both. The local planner then works on this coordinated target path to execute the movement. Since the same algorithms can be used for this purpose, the movement of multiple manipulators can be easily implemented.
[0031] It is understood that the features mentioned above and those to be explained below can be used not only in the combinations specified, but also in other combinations or on their own, without leaving the scope of the present invention.
[0032] Exemplary embodiments of the invention are shown in the drawing and are explained in more detail in the following description. They show: Fig. 1 a schematic representation of an embodiment of the new device, Fig. 2 a schematic representation of an embodiment of the new method, Fig. 3 a preferred embodiment of the new method, Fig. 4 a formal representation of an optimization problem of a local planner, and Fig. 5 an exemplary motion planning according to an embodiment of the new method in the case of a moving obstacle.
[0033] In the Fig. 1 An embodiment of the new device in its entirety is designated by reference numeral 10.
[0034] The device 10 is an industrial robot 12. Industrial robots are universal, programmable machines for handling, assembling, or machining workpieces, with at least one manipulator 14 that enables physical interaction with the environment and performs the robot's mechanical work. The manipulator 14 can include axes 16 with drive units such as motors, gearboxes, angle encoders (not shown here), and joints 18 for the mechanical coupling of the axes 16. The drive units are controlled by a control unit 20, which can be designed as a programmable computing unit and coordinates the movement of the robot and the manipulator.
[0035] In the exemplary embodiment according to Fig. 1 The robot 12 is a stationary robot with a base 22 on which a manipulator 14 is movably mounted. The manipulator 14 comprises, by way of example, three axes 16 and three joints 18. At the end of the manipulator 14 is an end effector 24, which is shown here by way of example as a gripper. The control unit 20 coordinates the movement of the robot 12 and the manipulator 14 and can be integrated into the robot 12 or connected to it as an external device.
[0036] It is understood that a stationary robot represents only one possible embodiment of a robot to which the method according to the invention can be applied. Alternatively, mobile robots, robots with multiple manipulators, or robots with a modular control unit divided into a multitude of components are conceivable. Likewise, the method is not limited to industrial robots but can also be applied to other automated machines.
[0037] Controlling the robot includes not only controlling the drives but also collision-free motion planning, i.e., searching for a motion path for the robot or its manipulator from a starting point 26 to an endpoint 28 without colliding with itself or an obstacle. Obstacles can be static 30 or moving 32. The latter means that motion planning must be performed not just once, but multiple times if an originally planned path becomes invalid due to an obstacle 32. Collision-free motion planning for the robot's manipulator in dynamic environments is the subject of the method according to the invention, which is described below with reference to the Fig. 2 will be explained in more detail.
[0038] Fig. 2 A schematic representation shows a method according to an embodiment of the present invention.
[0039] The method comprises three processes, which can preferably be executed in parallel. The first and second processes are planning processes. The third process involves the actual control of the robot. In this context, "parallel" means that the processes are executed in parallel from a computational perspective, for example, in synchronous or asynchronous threads.
[0040] The first process is a repeated motion planning with a global planner 34, and the second process is a continuous motion planning with a local planner 36. The third process executes the plan generated by the first and second planners 34 and 36 by controlling the robot's drives.
[0041] A global planner 34 within the meaning of the present invention is a planner that determines a target path from a defined starting point to a defined starting point. Kinematic path planning is sufficient for this purpose. In other words, the global planner does not need to perform path smoothing or time parameterization to determine an actual trajectory for the movement, since this is achieved by path following control using the local planner 36.
[0042] The result of each iteration of the global planner, provided a result can be found, is a kinematic path plan, i.e., a possible target path to the destination point. In a preferred embodiment, the starting point or a starting state can vary in each iteration based on the results of the local planner 36.
[0043] A local planner 36 within the meaning of the present invention is a planner that operates reactively; that is, instead of planning an entire path, the local planner 36 directly generates control impulses in response to its environment. The local planner 36 thus initiates movements that can "push" the robot away from obstacles to avoid collisions, with the overall movement preferably being "driven" toward a target point. The method can include a prediction of the dynamics over a horizon and allow for the consideration of constraints.
[0044] The combination according to the invention provides that both planners influence the third process, i.e., the control of the drives. Preferably, the global and local planners 34, 36 are not executed independently of each other, but rather in mutual exchange. The local planner 36 can build upon the results 40 of the global planner 34, and the global planner 34 can consider the results 42 of the local planner 36 when replanning. This exchange is indicated here by the arrow 44 and is realized through the implementation of the respective planners 34, 36. The exchange 44 is thus preferably an integral part of the respective planner.
[0045] The global planner 34 plans repeatedly, i.e., it performs at least one replanning during the movement. The local planner 36 plans continuously, i.e., the planning takes place continuously, in particular in a fixed cycle T 2. In a preferred embodiment, the global planning also takes place continuously, preferably also in a fixed cycle T 1. The cycle T 1 can be many times larger than the cycle T 2, for example by a factor of 10 to 30, in particular 20, and depends essentially on the "worst-case" computation times of the individual steps.
[0046] A replanning by global planner 34 occurs, in particular, without any external trigger; that is, global planner 34 carries out at least one replanning process that has not been triggered externally. In other words, the global replanning is initially independent of whether an old solution is still permissible or not.
[0047] In principle, the method according to the invention is not limited to a specific implementation of the respective planners. Other global planners that can repeatedly generate a target path, or other local planners that can use the target path of a global planner as a reference variable, are conceivable for combination in accordance with the present invention.
[0048] Regarding the Fig. 3 The following section explains in more detail a preferred global planner, a preferred local planner, and their combination. The same reference symbols refer to the same parts as before.
[0049] The procedure involves planning with a global planner 34 and a local planner 36. In this embodiment, the planners 34 and 36 are supplemented by a modeler 38, which includes an environment model 46 and can provide a distance map 47 for the local planner 36.
[0050] The global motion planning of the global planner 34 is based in the exemplary implementation according to Fig. 3 on a dynamic roadmap (DRM). Dynamic roadmaps are a planning approach for changing environments where frequent replanning is necessary. A DRM comprises a graph G = (V, E) in a configuration space C of the robot, as well as mapping rules. The mapping rules map the rasterized workspace W to nodes V and edges E in the graph G. They are used during the planning phase to quickly determine valid portions of the graph for specific obstacles.
[0051] The DRM-based planning approach basically comprises two phases: the preparatory phase and the planning phase.
[0052] In a preliminary phase, each of the n nodes is first assigned a random configuration q(u), where u ∈ V. The configuration is only checked for self-collision, as an empty environment is assumed in this phase. Subsequently, the k nn nearest neighbors v ∈ V of each node are determined using a distance metric d(u, v) = d(q(u), q(v))) to generate the edges (u, v) ∈ E of the graph. The choice of distance metric is an important factor influencing the quality of the roadmap.
[0053] The mappings φv : W → V and φe : W → E specify, for each voxel w ∈ W, which nodes and edges of the graph G are invalid if the voxel is occupied by an obstacle. All configurations and movements corresponding to the nodes and edges in the graph are rasterized to generate the inverse mapping rules φe --1< : V → W and φv --1< : E → W. Using these inverse mapping rules, a mapping of the DRM can be performed easily.
[0054] During the planning phase, DRM is used to find a valid path between a start configuration qs and a target configuration qG. First, the start and target are connected to their k nn nearest neighbors in the graph. This is the only step that requires collision checking during the planning phase. Next, all voxels that intersect obstacles are identified, and the mapping rules φv and φe are used to invalidate all vertices and edges of the graph that would lead to a collision. Finally, the shortest valid path between the start and target is sought. The result corresponds to a target path as defined by the invention.
[0055] DRM-based planners can achieve planning times of less than 100 milliseconds on standard hardware. Furthermore, parallel implementation on GPUs allows the roadmap size to scale up to one million nodes.
[0056] Planning using DRM is performed repeatedly, meaning a target path is determined multiple times at different points in time. Preferably, global planning also takes place continuously at a fixed interval. Fig. 3 For example, a global replanning is performed every 200 milliseconds. The global planner 34 thus operates in a continuous loop, generating a current target path until either the goal is reached or the planning is to be aborted. The latter occurs if the predicted trajectory is invalid within a global sampling time t ∈ [0, Δt global ], or if the maximum permissible execution time has expired.
[0057] In a preferred embodiment, the execution of the continuous determination by the global planner is also adjustable via a defined condition. Thus, the execution of the global planner 34 can be influenced by one or more guidelines. A distinction is preferably made between general planning guidelines 48 and guidelines for the error case 50.
[0058] Planning Guideline 48 allows you to specify when the global planner should be invoked to determine the current target path. A defined condition can specify unconditional, no, or conditional execution of the global planner. Unconditional execution performs global planning in every cycle, while if the defined condition specifies no execution, only local planning is used. In contrast, conditional execution allows you to determine whether or not the planner should be executed based on various criteria. One criterion, for example, could be whether the predicted trajectory is not valid for the entire horizon. Another or complementary criterion could be whether the distance between the current state and the target state has decreased during the last cycle, i.e., whether the local planner has made sufficient progress toward the target.
[0059] The guidelines for error case 50, on the other hand, can intervene if the global planner 34 cannot find a valid path, either because there is no solution or because the dynamic roadmaps are incomplete. In this case, the appropriate behavior can be defined via the guidelines for error case 50. For example, the robot can enter a waiting state where the current state is defined as the target path. Alternatively, the existing target path can be retained.
[0060] The local planner 36 then attempts either to "maintain" the current position or to proceed along the previously valid path. Alternatively, a theoretical, direct path can be set as the new path by the policy for error case 50, thus requiring the local planner to reach the destination via the most direct route, regardless of any obstacles. In this case, the local planner is essentially on its own and unaffected by global planning.
[0061] The planning guidelines 48 and the guidelines for the error case 50 thus allow the motion planning to be flexibly adapted to the respective scenario, thereby increasing the application possibilities of the new procedure.
[0062] The target paths, determined here using the DRM-based planner, serve as input for the local planner 36, as indicated by arrow 52. The local planner 36 operates continuously, for example, as shown here, with a cycle time of 10 milliseconds.
[0063] In this embodiment, the local planner 36 is based on model predictive control and is also referred to as a Model Predictive Controller (MPC). Model predictive control determines the actual control input for the manipulator and is responsible, among other things, for collision avoidance and path tracking. For this purpose, an optimization problem is solved in each cycle, and the first part of the optimal solution is used as the control input. The target path p(s) with s ∈ [0, s ∈ G ], which is generated by the global planner 34, serves as a guide for the local planner 36, thus avoiding the problem of local minima. An optimization problem can be, as in Fig. 4 be presented, formulated.
[0064] Fig. 4 shows a formal representation of an optimization problem of a preferred local planner 36.
[0065] The dynamic optimization problem 54 takes as its starting point the system model, which describes the dynamics 56 of the moving system, i.e., here the manipulator 14, with the states 58 and control variables 60. The dynamics 56 can be simplified as integrator 1 = u with the states x = q and the control variables q̇ or as a dual integrator q̈ = u with the conditions x [ q T< , q̈ T< and the control variables u = q be accepted.
[0066] If the calculated trajectory is to be executed directly, the double integrator is preferable, as it allows the minimum and maximum acceleration to be taken into account.
[0067] A cost functional 62 weights the states 58 and control variables 60. The task of optimization is to find the optimal control function from the typically infinite number of control functions that minimizes the cost functional 62. The states 58 may be subject to state constraints 64, and the control variables 60 to control variable constraints 66.
[0068] To couple the local planner 36 to the global planner 34, the cost functional 62 additionally weights a deviation of the manipulator 14 from the target path as well as a deviation of a current path parameter s, i.e. the current position and / or the current state, from a target parameter s G specified by the target path.
[0069] The deviation between the current path parameter s and the target parameter sG is preferably not weighted quadratically, since the dimension of the gradient depends on the distance to the target. Instead, cost functional 62 uses the absolute value and switches to quadratic weighting when the absolute value is less than a defined limit Δs. In this way, oscillations near the target can be avoided.
[0070] Furthermore, the dynamic optimization problem 54 can include path dynamics 68 and path entry constraints 70. The path dynamics 68 are based on the target path of the global planner 34. Figuratively speaking, the local planner follows the globally planned path as if suspended by a spring.
[0071] The dynamic optimization problem 54 can also include at least one inequality constraint 72, which allows obstacles to be taken into account. The inequality constraints 72 consider a distance to self-collisions and / or a distance to collisions with obstacles 30, 32.
[0072] To calculate the self-collision distance, the robot can be approximated by spheres with a center point and radius. To calculate the obstacle collision distance, the environment can be approximated by a raster map, with the required distances being obtained through the Euclidean distance transformation of the raster map. The smallest distance to either self-collision and / or collision with an obstacle is considered. The obstacle rasterization and the Euclidean distance transformation are preferably calculated in advance (Modeler 38) so that the real-time capability of the optimization is not affected.
[0073] To account for rapidly moving obstacles, the planner can implement obstacle movement estimation and short-term prediction. Since both dynamic roadmaps and optimization-based planners rely on a rasterized representation of the workspace W, there is no information about when an object will be in a specific position. However, additional voxels can be blocked along the direction of an obstacle's movement and / or an obstacle's representation can be artificially enlarged, allowing the planner to react to obstacles early. This can increase the robustness of the process.
[0074] Fig. 5 Finally, a schematic representation illustrates a possible movement in railway planning with a combined global and local planner.
[0075] Fig. 5 Figure 1 shows a two-dimensional configuration space C with an obstacle that defines an invalid region C obs. The globally planned target path p(s) 74 is used as a guide for the local optimization-based planner, which determines the actual movement in the form of the path x(t) 76.
[0076] In the Fig. 5 The images, shown side by side, illustrate three consecutive points in time from left to right.
[0077] In the first illustration, the local planner 36 generates a path 76a that deviates from the target path 74a to enable faster progress. In the second illustration, after the obstacle has moved, the local planner 36 generates path 76b to avoid a collision with the obstacle. As soon as a new global path 74b is calculated, the local planner 36 automatically adjusts the movement to the new target path 74b, as shown in the last illustration. The global planner 34 can, as shown here, request an intermediate result in the form of a predicted position 78 from the local planner 36 and take it into account when determining the new target path 74b of the manipulator 14.
[0078] It is understood that the implementations presented here are only an example of the method and device according to the invention. Other embodiments may also be covered by the subject matter of the invention. In principle, the scope of protection of the present application is defined by the claims and not limited by the description or the features shown in the figures.
Claims
1. A method for collision-free motion planning of a first manipulator (14) of a robot (12) from a starting point (26) to a destination point (28), having the steps: - determining a target path of said first manipulator (14) to the destination point (28) with a global planner (34); - continuously determining a movement of the first manipulator (14) with a local planner (36) based on a current target path of the global planner; - performing the movement by the first manipulator (14) in parallel with the determination by said global planner and said local planner (34, 36) characterized in that the global planner carries out replanning at least once and the local planner performs the movement of the first manipulator (14) based on the replanned target path.
2. The method according to claim 1, wherein the global planner (34) continuously determines the current target path.
3. The method according to one of claims 1 or 2, wherein the global planner (34) is a deterministic planner, in particular based on a dynamic roadmap (DRM).
4. The method according to any one of claims 1 to 3, wherein the global planner (34) inquires a result of the local planner (36) during re-planning and takes it into account when determining the target path.
5. The method according to any one of claims 1 to 4, wherein the local planner (36) is based on model predictive control (MPC) and the determination of the movement is based on optimization by solving a dynamic optimization problem (54).
6. The method according to claim 5, wherein the dynamic optimization problem (54) comprises a cost function (62) which weights a deviation of the manipulator (14) from the target path and a deviation of a current path parameter from a target parameter specified by the target path.
7. The method according to claim 6, wherein the dynamic optimization problem (54) further comprises path dynamics (68) and path input constraints (70), and the path dynamics (68) is based in particular on the target path of the global planner (34).
8. The method according to any one of claims 5 to 7, wherein the dynamic optimization problem (54) further comprises at least one inequality constraint (72) for a distance to collisions.
9. The method according to any one of claims 1 to 8, wherein the execution of the determination by the global planner is adjustable via a defined condition, in particular wherein an unconditional, no or conditional execution of the determination by the global planner (34) can be set via the defined condition.
10. The method according to claim 8, wherein the conditional execution depends on whether a determined target path and / or a determined movement is valid in a defined planning horizon and whether a distance between a current state and a target state has been reduced during a last determination cycle of the local planner (36).
11. The method according to any one of claims 1 to 10, wherein moving the first manipulator (14) pauses if planning by the global planner (34) fails.
12. The method according to any one of claims 1 to 10, wherein moving the first manipulator (14) is continued exclusively based on the local planner (36) if planning by the global planner (34) fails.
13. The method according to one of claims 1 to 12, wherein the global and / or local planner (34, 36) partially predicts a movement of an obstacle and / or artificially enlarges a representation of an obstacle (30, 32).
14. The method according to any one of claims 1 to 13, wherein the method further comprises additional motion planning for at least a second manipulator of the robot (12) and the additional motion planning is carried out in parallel and coordinated with the collision-free motion planning of the first manipulator (14).
15. An apparatus (10) comprising a robot (12) having at least one manipulator (14) and a control unit (20), the control unit (20) being adapted to perform the method according to any one of claims 1 to 14.
