Low-delay real-time obstacle avoidance motion control method and mobile robot
By combining iterative optimization and analytical geometric calculation, the problems of massive point cloud constraints and dynamic obstacle handling in robot navigation are solved, achieving low-latency and efficient obstacle avoidance control.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN MAXVISION TECH
- Filing Date
- 2026-06-15
- Publication Date
- 2026-07-14
AI Technical Summary
Existing robot navigation and obstacle avoidance methods suffer from severe computational delays when dealing with massive point cloud constraints and are unable to effectively cope with dynamic obstacles, resulting in insufficient robustness.
A low-latency, real-time obstacle avoidance motion control method is adopted. Through alternating iterative optimization of point cloud preprocessing, predicted point cloud generation, distance calculation, and trajectory optimization, combined with neural network batch inference and analytical geometry calculation, the optimal obstacle avoidance trajectory is generated.
It significantly reduces the computational burden, improves trajectory predictability and robustness in dynamic scenarios, and achieves a better balance between safety and task efficiency in the output trajectory.
Smart Images
Figure CN122387150A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of obstacle avoidance control technology for mobile robots, and more specifically, it relates to a low-latency real-time obstacle avoidance motion control method and a mobile robot. Background Technology
[0002] In the field of robot navigation, time-domain-based motion planning and control methods are a common and effective approach. The core idea is to predict future trajectories based on the current state within the time domain. Specifically, within a control cycle, the controller first establishes a motion model of the system, then constructs an objective function and constraints that include tracking error, control energy, and obstacle avoidance requirements. Next, it solves an optimization problem within a finite time domain, finally executing only the first control command and repeating the process at the next time step.
[0003] However, such methods encounter serious problems when dealing with obstacle avoidance: point clouds generated by sensors such as lidar usually contain tens of thousands of points, each of which needs to meet a safe distance constraint, resulting in an explosion in the number of constraints (e.g., 1000 points × 10 prediction times = 10000 constraints), making the solution of optimization problems very slow and unable to guarantee real-time performance.
[0004] Furthermore, traditional methods typically assume that obstacles are stationary, making them ineffective in dealing with dynamic obstacles. Perception and control are optimized independently in stages, which can easily lead to local suboptimal solutions and result in insufficient robustness. Summary of the Invention
[0005] The purpose of this application is to provide a low-latency real-time obstacle avoidance motion control method and a mobile robot to solve the technical problems of high latency and insufficient robustness in the existing technology of robot navigation and obstacle avoidance.
[0006] To achieve the above objectives, the technical solution adopted in this application is: to provide a low-latency real-time obstacle avoidance motion control method, comprising the following steps:
[0007] Point cloud preprocessing steps: Obtain the obstacle point cloud of the robot's surrounding environment, and generate velocity information for each obstacle point based on the point cloud;
[0008] Predicted point cloud generation steps: Obtain the robot's predicted trajectory, and generate a corresponding predicted point cloud sequence in the robot coordinate system for each time step in the prediction time domain based on the position and velocity information of the obstacle points and the robot's predicted pose.
[0009] Distance calculation steps: Infer the distance weight of each point relative to each edge of the robot's outer contour by batch reasoning of all points in the predicted point cloud sequence at once, and calculate the pushing direction of each point according to the distance weight;
[0010] Trajectory optimization steps: Based on the distance weights and the pushing direction, a multi-objective optimization problem is constructed, and the optimal obstacle avoidance trajectory of the robot in the prediction time domain is obtained by solving the optimization problem;
[0011] Iterative control steps: drive the predicted point cloud generation step, distance calculation step and trajectory optimization step to perform alternating iterative optimization until the convergence condition is met.
[0012] This application also provides a mobile robot, including:
[0013] The robot itself;
[0014] Sensor systems are used to collect environmental data;
[0015] Actuators, used to drive the robot's movement; and
[0016] A processor for processing the low-latency real-time obstacle avoidance motion control method described above, wherein the processor receives data from the sensor system, generates control commands, and outputs them to the actuator.
[0017] The beneficial effects of the low-latency real-time obstacle avoidance motion control method and mobile robot provided in this application are as follows: Compared with the prior art, the combination of batch inference and push direction analysis calculation makes the obstacle avoidance problem, which was originally impossible to solve in real time due to the constraints of massive point clouds, feasible, significantly reducing the computational burden in a single cycle; the spatiotemporal coupled prediction point cloud sequence generation method enables the robot to consider its own motion and obstacle motion at the same time, improving the trajectory predictability in dynamic scenes; the alternating iteration mechanism allows distance calculation and trajectory optimization to feed back to each other and gradually refine the process, effectively avoiding local suboptimal decisions caused by perception errors or model simplification, and the final output trajectory achieves a better balance between safety and task efficiency. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This is a flowchart illustrating the low-latency real-time obstacle avoidance motion control method provided in the first embodiment of this application.
[0020] Figure 2 This is a schematic diagram of the predicted point cloud generation steps provided in the first embodiment of this application.
[0021] Figure 3This is a schematic diagram illustrating trajectory optimization through distance calculation, as provided in the first embodiment of this application.
[0022] Figure 4 This is a schematic diagram of the reaction velocity based on the change in the pushing direction, provided for the first embodiment of this application.
[0023] Figure 5 This is a schematic diagram of a mobile robot provided in the second embodiment of this application. Detailed Implementation
[0024] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.
[0025] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0026] The low-latency real-time obstacle avoidance motion control method provided in the first embodiment of this application will now be described.
[0027] Please refer to the following: Figure 1 The low-latency real-time obstacle avoidance motion control method includes:
[0028] Point cloud preprocessing steps: Obtain the obstacle point cloud of the robot's surrounding environment, and generate velocity information for each obstacle point based on the point cloud.
[0029] Understandably, in the point cloud preprocessing step, the robot acquires point cloud data of the surrounding environment using LiDAR and depth cameras. After coordinate system alignment and height filtering, a 2D point cloud is generated. Simultaneously, velocity information is estimated for each point as needed for subsequent steps.
[0030] Predicted point cloud generation steps: Obtain the robot's predicted trajectory, and generate a corresponding predicted point cloud sequence in the robot coordinate system for each time step in the prediction time domain based on the position and velocity information of the obstacle points and the robot's predicted pose.
[0031] Understandably, in the predicted point cloud generation step, based on the robot's current predicted trajectory (e.g., obtained from the previous round of optimization results or the initial path), and combined with the position and velocity of each obstacle point, a point cloud sequence in the robot's coordinate system is generated for each time step in the prediction time domain (e.g., 1 second in the future, a total of 20 time steps).
[0032] Distance calculation steps: All points in the predicted point cloud sequence are inferred in batches at once to obtain the distance weight of each point relative to each edge of the robot's outer contour, and the pushing direction of each point is calculated based on the distance weight.
[0033] Understandably, in the distance calculation step, the point clouds from all the aforementioned time steps are concatenated into a long list and fed into a pre-trained distance calculation neural network all at once. This network outputs the distance weights of each point relative to each edge of the robot, and then calculates the pushing direction of each point using an analytical formula.
[0034] It is worth noting that the distance weight refers to the degree of membership of each point in the obstacle point cloud relative to the boundary lines of the robot's outer contour. Its value represents the proximity of the point to the corresponding boundary line, and all weights have been normalized. The push direction is a two-dimensional vector calculated geometrically based on the distance weight, the robot's current orientation angle, and the normal direction components of each boundary line. It is used to indicate the optimal motion trend for the robot to avoid the obstacle point.
[0035] Trajectory optimization steps: Based on the distance weight and the pushing direction, a multi-objective optimization problem is constructed, and the optimal obstacle avoidance trajectory of the robot in the prediction time domain is obtained by solving the optimization problem.
[0036] Understandably, in the trajectory optimization step, a multi-objective optimization problem is constructed based on the obtained distance weights and push direction. The objective function includes navigation deviation along the reference path, collision penalty with obstacles, trajectory smoothness, etc. By solving this optimization problem, the optimal trajectory of the robot in the prediction time domain is obtained.
[0037] Iterative control steps: drive the predicted point cloud generation step, distance calculation step and trajectory optimization step to perform alternating iterative optimization until the convergence condition is met.
[0038] Understandably, in the iterative control steps, the three steps of predictive point cloud generation, distance calculation, and trajectory optimization are executed alternately, typically converging in 2-3 rounds. During the iteration process, the change in the pushing direction between adjacent rounds is used to counteract the true velocity of the dynamic obstacle and correct the velocity estimate in the point cloud preprocessing. After convergence, the first control command (linear velocity and angular velocity) of the optimal trajectory is sent to the robot actuator.
[0039] The low-latency real-time obstacle avoidance motion control method provided in the first embodiment of this application designs a real-time obstacle avoidance control framework with tight coupling of perception, prediction, calculation, optimization, and iteration. Its core idea is to replace the discrete and repetitive geometric constraint solving in traditional models with single-pass neural network batch inference, and to directly derive the derivation direction from the distance weights output by the network using analytical geometric relationships, thereby significantly simplifying the calculation process. Simultaneously, it introduces spatiotemporally coupled point cloud sequence generation based on robot predicted pose and obstacle motion information, enabling the planning layer to perceive dynamic environmental changes in advance. This framework uses alternating iterations to allow the perception module and optimization module to mutually correct each other, fundamentally changing the inefficient mode of traditional methods where perception and control are separated and information is transmitted in only one direction.
[0040] The low-latency real-time obstacle avoidance motion control method provided in this application has the following advantages compared with the prior art: By combining batch inference with push direction analysis calculation, the obstacle avoidance problem that was originally impossible to solve in real time due to massive point cloud constraints becomes feasible, significantly reducing the computational burden in a single cycle; the spatiotemporally coupled prediction point cloud sequence generation method enables the robot to consider its own motion and obstacle motion simultaneously, improving trajectory predictability in dynamic scenarios; the alternating iteration mechanism allows distance calculation and trajectory optimization to feed back to each other and gradually refine the process, effectively avoiding local suboptimal decisions caused by perception errors or model simplification, and the final output trajectory achieves a better balance between safety and task efficiency.
[0041] It must be noted that the low-latency real-time obstacle avoidance motion control method provided in the first embodiment of this application is not simply a neural network superimposed on the existing framework, but a systematic reconstruction of the obstacle avoidance planning paradigm: it reduces the dimensionality of the neural network from an end-to-end black box to a weight generator, and then uses the inherent geometric structure of the robot to complete the direction resolution, which not only preserves physical interpretability, but also reduces the learning burden of the network; at the same time, it explicitly incorporates the influence of the robot's own motion on perception into the prediction point cloud generation process, so that the perception module can dynamically adapt to changes in the robot's posture.
[0042] Specifically, the velocity information generated for each obstacle point in the point cloud preprocessing step includes:
[0043] Using the fixed beam index of the lidar, beam points with the same index in adjacent frames are paired, and the velocity of each point is calculated by position difference in the global coordinate system.
[0044] The error vector between the predicted obstacle position in the previous cycle and the actual observed position in the current cycle is fed back into the current velocity estimate according to the gain coefficient.
[0045] Understandably, the velocity is first obtained directly by differentiating the global coordinates of beam points with the same index in adjacent frames, reducing the computational complexity to a linear relationship with the number of points. Then, a feedback mechanism for prediction-observation error is introduced, adding the deviation between the predicted position and the actual observation from the previous cycle to the current velocity estimate as a gain. This allows the originally fixed uniform velocity model to dynamically adapt to the non-uniform behavior of obstacles, such as sudden turns or acceleration / deceleration. The closed-loop error correction further endows the velocity estimation with self-learning capability, enabling the system to continuously correct model biases and maintain a low position prediction error even in scenarios such as pedestrians suddenly turning.
[0046] Specifically, the point cloud preprocessing step further includes:
[0047] Each point in the point cloud is weighted by confidence based on semantic information, with points corresponding to dynamic obstacles given high weight, points corresponding to static obstacles given medium weight, and points corresponding to uncertain regions given low weight.
[0048] Weighted downsampling is performed based on the confidence level weights, prioritizing the retention of points with higher weights.
[0049] It is understood that the terms high weight, medium weight, and low weight are relative, with high weight > medium weight > low weight. Dynamic obstacles refer to movable obstacles such as people, vehicles, and dogs, regardless of whether the obstacle is currently in motion; the same applies to static obstacles. Preferably, the determination of dynamic and static obstacles is based on visual classification detection.
[0050] By introducing semantic awareness, point clouds are divided into dynamic obstacles, static obstacles, and uncertain regions, and different confidence weights are assigned. Then, weighted downsampling is performed based on these weights, so that high-weight, high-risk points are preferentially retained in the downsampling process.
[0051] In this way, the differentiated allocation of perceived value is realized in embedded systems with limited computing power, which not only ensures the detection rate of key obstacles, but also avoids the risk of missed detection caused by global uniform downsampling.
[0052] Specifically, the point cloud preprocessing step further includes:
[0053] The threat level of each obstacle point is calculated based on the distance of the obstacle point relative to the robot, the magnitude of its velocity, and the angle between the velocity direction and the position vector.
[0054] The sampling quota for different regions of the point cloud is dynamically adjusted based on the threat level, so that points with high threat levels receive a higher sampling probability.
[0055] It is understandable that the closer the obstacle is to the robot, the greater its speed, and the smaller the angle between its speed direction and its position vector, the greater the threat level, and vice versa.
[0056] The key is that the introduction of threat index can complement the semantic weighting mentioned above: avoid key dynamic targets being mistakenly discarded in downsampling, such as a rolling trash can, but trash cans are usually static obstacles; dynamically compensate for potential sudden risks, such as a child standing still, but there is a potential risk of rushing over at any time, or a corner area, but there is a risk of face-to-face collision at the corner, so the corner area can also be defined as a dynamic obstacle.
[0057] For details, please refer to Figure 2 The predicted point cloud generation step specifically includes:
[0058] For the i-th obstacle point, based on its current global coordinates and velocity vector Predict its global coordinates at the h-th prediction step. ,in , where h = 0, 1...H, and Δt is the time interval for each prediction step;
[0059] Obtain the robot's predicted pose at the h-th prediction step, including position T. h and rotation matrix R h ;
[0060] Through the spacetime coupling coordinate transformation formula The global predicted position of the obstacle is transformed to the robot coordinate system to generate the predicted point cloud sequence; wherein Let h be the coordinates of the obstacle point in the robot coordinate system at step h. This is the transpose of the rotation matrix;
[0061] The uncertainty of each obstacle's motion is estimated based on the magnitude of the error vector, and the safe distance to the corresponding obstacle is dynamically adjusted based on the uncertainty.
[0062] Understandably, the spatiotemporal coupling coordinate transformation formula is used to transform the global coordinates to the robot's coordinate system at the corresponding moment, thereby generating a time-step-by-time independent predicted point cloud sequence. This step explicitly distinguishes between the active motion of obstacles and the passive motion of the robot itself, enabling the predicted point cloud to dynamically adjust with changes in the robot's posture, avoiding the prediction bias caused by traditional methods that use the current moment's point cloud to approximate all future moments.
[0063] On the one hand, the spatiotemporally coupled predicted point cloud sequence enables subsequent distance calculation and trajectory planning to accurately perceive the evolution trend of dynamic scenes, significantly improving the ability to avoid non-stationary obstacles such as moving pedestrians and vehicles. On the other hand, the magnitude of the error vector in the aforementioned point cloud preprocessing step estimates the prediction uncertainty, and the safety distance is dynamically adjusted based on this uncertainty, so that the robot automatically increases the safety margin when the confidence in future predictions decreases, thereby enhancing robustness.
[0064] For details, please refer to Figure 3 The distance calculation step specifically includes:
[0065] The point clouds from all time steps in the generated predicted point cloud sequence are concatenated into a total point list, and the distance weights of each point relative to each edge of the robot are calculated and output all at once. , where k=1…l, and l is the total number of robot boundary lines;
[0066] According to the distance weight Robot orientation angle θ h and the normal direction component of the robot boundary line (a k ,b k The direction of pushback at each point is calculated analytically using the following formula. , ):
[0067] ,
[0068] .
[0069] Understandably, by stitching together the point clouds of all time steps into a single list, the neural network can complete the forward inference of all points at once, avoiding the overhead of multiple startups and the cumulative latency of single inference caused by calling the network step by step. Furthermore, the network only outputs the distance weights of each point relative to each edge of the robot, and then, based on the robot's current orientation angle and the normal direction components of each edge, it solves the pushing direction using analytical geometric formulas, rather than having the network directly output the direction vector.
[0070] Batch inference increases the total computational cost almost linearly with the prediction time step, only increasing the matrix size in practical applications without additional kernel startup costs. Meanwhile, the analytical calculation of the push-out direction reduces the network output dimension from l+2 to l, significantly reducing the number of parameters and inference time, while providing clearer physical meaning. By cleverly utilizing prior geometric knowledge of the robot's outer contour, the traditional black-box output direction is improved to lightweight weighted perception. Directional information is then recovered through a fixed geometric mapping, preserving the speed of end-to-end methods while providing interpretability. Furthermore, it lays the mathematical foundation for differentiability.
[0071] For details, please refer to Figure 3 and Figure 4 The iterative control step further includes:
[0072] After each round of distance calculation is completed, for each obstacle point, record the pushing direction f obtained from the analytical calculation in this round. (k) The direction of the push in the previous round, f (k−1) , and the corresponding robot pose;
[0073] Based on the spatial rate of change of the pushing direction with respect to the obstacle's position, the following relationship is established:
[0074]
[0075] Where K is a coefficient matrix determined by the robot's geometric parameters and current pose, v obs Let be the velocity of the obstacle to be solved, and Δt be the iteration time interval;
[0076] The velocity v of the obstacle is obtained by solving the equation using the least squares method. obs The estimated value;
[0077] Using this estimate to the velocity vector The corrected velocity vector is then used in the next iteration of the predicted point cloud generation step.
[0078] It is understandable that the change in the pushing direction is caused by the robot's own movement and the obstacle's active movement. By using the known robot movement information, the obstacle's movement component can be extracted from the total change, thus achieving reverse perception of the environment's dynamics by inferring from obstacle avoidance decisions.
[0079] Unlike the traditional unidirectional information flow mechanism of perception-planning, this application uses the intermediate geometric quantities (push-back direction) generated during the iteration process as the observation source, which are then used to update the motion state of the perception layer, adopting an evolutionary mechanism of perception-planning-re-perception. Simultaneously, by pre-determining the coefficient matrix using the differentiable structure of the analytical push-back direction and the robot's inherent geometric parameters, the inversion problem is reduced to a lightweight least-squares solution, achieving adaptive evolution of dynamic perception without significantly increasing the computational burden.
[0080] For details, please refer to Figure 3 In the trajectory optimization step, the variable safety distance d is... h As an optimization variable, its value range is (d min d max ), and set a safe distance reward item in the total cost function. Where η is the reward weight; the variable safety distance d h With the corrected obstacle velocity v obsDynamic adjustment: when v obs When pointing at the robot, reduce the reward weight η so that d h tending towards d min When v obs When moving away from the robot, increase η to make d h tending towards d max .
[0081] It is understandable that in the trajectory optimization step, the motion direction output by the perception layer directly changes the behavioral preferences in the optimization target, enabling the safe distance strategy to contract or expand synchronously with the approach of the threat, rather than passively relying on the constraints of the collision penalty term. That is, based on the aforementioned velocity inversion and combined with policy adjustment, the motion intention of dynamic obstacles is injected into the cost function as prior knowledge. This not only improves the obstacle avoidance success rate in narrow dynamic scenarios but also avoids ineffective collision penalties or excessive conservatism caused by a fixed safe distance.
[0082] In addition, the optimization problem can also include at least one of navigation deviation cost, collision penalty cost, and trajectory smoothing constraint. By simultaneously optimizing the trajectory through multi-objective optimization, the cost can be reduced.
[0083] Specifically, the iterative control step includes a multi-resolution variable step-size prediction time domain, wherein the time step size of the recent prediction time step is smaller than the time step size of the distant prediction time step.
[0084] Understandably, this avoids the dilemma of fixed step size leading to an excessively short prediction time domain for distant targets while ensuring near-term safety, or coarse near-term discretization when pursuing a long time domain. Instead, it adaptively allocates discretization precision according to task requirements.
[0085] The mobile robot 100 provided in the second embodiment of this application will now be described.
[0086] Please see Figure 5 A mobile robot 100, comprising:
[0087] Robot body 101;
[0088] Sensor system 102 is used to collect environmental data;
[0089] Actuator 103, used to drive the robot's movement; and
[0090] A processor for processing the low-latency real-time obstacle avoidance motion control method provided in the first embodiment of this application, wherein the processor receives data from the sensor system 102, generates control commands, and outputs them to the actuator 103.
[0091] Specifically, the actuator 103 includes a differential drive chassis, the sensor system 102 includes sensors such as lidar and depth camera, the processor generates control commands according to a low-latency real-time obstacle avoidance motion control method, and the actuator 103 includes a motor, which drives the differential drive chassis to move, thereby realizing real-time obstacle avoidance navigation.
[0092] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A low-latency real-time obstacle avoidance motion control method, characterized in that, Includes the following steps: Point cloud preprocessing steps: Obtain the obstacle point cloud of the robot's surrounding environment, and generate velocity information for each obstacle point based on the point cloud; Predicted point cloud generation steps: Obtain the robot's predicted trajectory, and generate a corresponding predicted point cloud sequence in the robot coordinate system for each time step in the prediction time domain based on the position and velocity information of the obstacle points and the robot's predicted pose. Distance calculation steps: Infer the distance weight of each point relative to each edge of the robot's outer contour by batch reasoning of all points in the predicted point cloud sequence at once, and calculate the pushing direction of each point according to the distance weight; Trajectory optimization steps: Based on the distance weights and the pushing direction, a multi-objective optimization problem is constructed, and the optimal obstacle avoidance trajectory of the robot in the prediction time domain is obtained by solving the optimization problem; Iterative control steps: drive the predicted point cloud generation step, distance calculation step and trajectory optimization step to perform alternating iterative optimization until the convergence condition is met.
2. The low-latency real-time obstacle avoidance motion control method as described in claim 1, characterized in that, The velocity information generated for each obstacle point in the point cloud preprocessing step includes: Using the fixed beam index of the lidar, beam points with the same index in adjacent frames are paired, and the velocity of each point is calculated by position difference in the global coordinate system. The error vector between the predicted obstacle position in the previous cycle and the actual observed position in the current cycle is fed back into the current velocity estimate according to the gain coefficient.
3. The low-latency real-time obstacle avoidance motion control method as described in claim 1 or 2, characterized in that, The point cloud preprocessing step also includes: Each point in the point cloud is weighted by confidence based on semantic information, with points corresponding to dynamic obstacles given high weight, points corresponding to static obstacles given medium weight, and points corresponding to uncertain regions given low weight. Weighted downsampling is performed based on the confidence level weights, prioritizing the retention of points with higher weights.
4. The low-latency real-time obstacle avoidance motion control method as described in claim 3, characterized in that, The point cloud preprocessing step also includes: The threat level of each obstacle point is calculated based on the distance of the obstacle point relative to the robot, the magnitude of its velocity, and the angle between the velocity direction and the position vector. The sampling quota for different regions of the point cloud is dynamically adjusted based on the threat level, so that points with high threat levels receive a higher sampling probability.
5. The low-latency real-time obstacle avoidance motion control method as described in claim 2, characterized in that, The specific steps for generating the predicted point cloud include: For the i-th obstacle point, based on its current global coordinates and velocity vector Predict its global coordinates at the h-th prediction step. ,in , where h = 0, 1...H, and Δt is the time interval for each prediction step; Obtain the robot's predicted pose at the h-th prediction step, including position T. h and rotation matrix R h ; Through the spacetime coupling coordinate transformation formula The global predicted position of the obstacle is transformed to the robot coordinate system to generate the predicted point cloud sequence; wherein Let h be the coordinates of the obstacle point in the robot coordinate system at step h. This is the transpose of the rotation matrix; The uncertainty of each obstacle's motion is estimated based on the magnitude of the error vector, and the safe distance to the corresponding obstacle is dynamically adjusted based on the uncertainty.
6. The low-latency real-time obstacle avoidance motion control method as described in claim 5, characterized in that, The distance calculation steps specifically include: The point clouds from all time steps in the generated predicted point cloud sequence are concatenated into a total point list, and the distance weights of each point relative to each edge of the robot are calculated and output all at once. , where k=1…l, and l is the total number of robot boundary lines; According to the distance weight Robot orientation angle θ h and the normal direction component of the robot boundary line (a k ,b k The direction of pushback at each point is calculated analytically using the following formula. , ): ， 。 7. The low-latency real-time obstacle avoidance motion control method as described in claim 6, characterized in that, The iterative control step further includes: After each round of distance calculation is completed, for each obstacle point, record the pushing direction f obtained from the analytical calculation in this round. (k) The direction of the push in the previous round, f (k−1) , and the corresponding robot pose; Based on the spatial rate of change of the pushing direction with respect to the obstacle's position, the following relationship is established: Where K is a coefficient matrix determined by the robot's geometric parameters and current pose, v obs Let be the velocity of the obstacle to be solved, and Δt be the iteration time interval; The velocity v of the obstacle is obtained by solving the equation using the least squares method. obs The estimated value; Using this estimate to the velocity vector The corrected velocity vector is then used in the next iteration of the predicted point cloud generation step.
8. The low-latency real-time obstacle avoidance motion control method as described in claim 7, characterized in that, In the trajectory optimization step, the variable safety distance d h As an optimization variable, its value range is (d min d max ), and set a safe distance reward item in the total cost function. Where η is the reward weight; The variable safety distance d h With the corrected obstacle velocity v obs Dynamic adjustment: when v obs When pointing at the robot, reduce the reward weight η so that d h tending towards d min When v obs When moving away from the robot, increase η to make d h tending towards d max .
9. The low-latency real-time obstacle avoidance motion control method as described in claim 1, characterized in that, The iterative control step includes a multi-resolution variable step-size prediction time domain, wherein the time step size of the recent prediction time step is smaller than the time step size of the distant prediction time step.
10. A mobile robot, characterized in that, include: The robot itself; Sensor systems are used to collect environmental data; An actuator is used to drive the robot's movement; as well as A processor for processing the low-latency real-time obstacle avoidance motion control method as described in any one of claims 1 to 9, wherein the processor receives data from the sensor system, generates control commands, and outputs them to the actuator.