Trajectory prediction method, controller, robot and storage medium

Abstract

The embodiment of the invention provides a trajectory prediction method, a controller, a robot and a storage medium, relates to the technical field of path planning, and is suitable for the fields of financial science and technology and medical health. The method comprises the following steps: determining an intelligent agent influencing a moving path of a target robot in a target scene; obtaining a basic constant speed model of the intelligent agent, and obtaining at least two basic movement constraint logics of the intelligent agent; calling a large language model to perform population generation on the basic constant speed model according to the target scene and the basic movement constraint logic to obtain a target population; wherein the target population comprises at least two candidate trajectory generation models; calling an evolutionary algorithm to perform individual screening on each candidate trajectory generation model in the target population to obtain a target trajectory generation model; and generating a prediction trajectory of the intelligent agent through the target trajectory generation model. According to the embodiment of the invention, the accuracy of trajectory prediction can be improved.

Description

Technical Field

[0001] This application relates to the field of path planning technology, applicable to the fields of fintech and healthcare, and particularly to a trajectory prediction method, controller, robot, and storage medium. Background Technology

[0002] During robot operation, it is often necessary to predict the movement trajectories of intelligent agents (such as other robots, pedestrians, etc.). For example, in the banking scenario within the fintech field, a bank robot moves through the bank lobby. In this case, the bank robot needs to predict the movement paths of customers in the lobby to avoid collisions. Similarly, in the elderly care scenario within the healthcare field, a care robot can deliver medications, meals, medical equipment, and other items to the elderly. During its movement, the care robot needs to predict the movement paths of other pedestrians or other robots to avoid collisions.

[0003] However, current trajectory prediction methods typically employ hand-designed heuristics, such as social force models and constant velocity models. These methods rely on human experience and trial-and-error processes, making the trajectory prediction results susceptible to subjective influence and limiting their accuracy. Summary of the Invention

[0004] The main objective of this application is to propose a trajectory prediction method, controller, robot, and storage medium, aiming to solve the technical problem that trajectory prediction is highly dependent on subjective experience, resulting in limited accuracy, and to improve the accuracy of trajectory prediction.

[0005] To achieve the above objectives, a first aspect of this application proposes a trajectory prediction method, the method comprising: Identify the agents that influence the movement path of the target robot in the target scenario; Obtain the basic constant velocity model of the agent, and obtain at least two basic movement constraint logics of the agent; The large language model is invoked to generate a population of the basic constant velocity model based on the target scene and the basic movement constraint logic, thereby obtaining a target population; wherein, the target population includes at least two candidate trajectory generation models; An evolutionary algorithm is invoked to perform individual screening on each of the candidate trajectory generation models in the target population to obtain the target trajectory generation model; The target trajectory generation model generates the predicted trajectory of the agent.

[0006] In some embodiments, the step of invoking an evolutionary algorithm to individually screen each of the candidate trajectory generation models in the target population to obtain a target trajectory generation model includes: The trajectory prediction performance of each candidate trajectory generation model is evaluated to obtain a trajectory prediction performance score. Based on the trajectory prediction performance score, model selection is performed among the candidate trajectory generation models to obtain parent model pairs; wherein, the parent model pair includes a first parent trajectory generation model and a second parent trajectory generation model; Based on the differences in trajectory prediction performance scores between the first parent trajectory generation model and the second parent trajectory generation model, as well as the differences in the basic movement constraint logic, the first parent trajectory generation model and the second parent trajectory generation model are fused to obtain a fused trajectory generation model. Obtain an initial historical trajectory generation model, and search for similar scene models in the initial historical trajectory generation model according to the target scene to obtain a similar historical trajectory generation model; The target trajectory generation model is obtained by filtering models based on the fused trajectory generation model and the similar historical trajectory generation model.

[0007] In some embodiments, the step of fusing the first parent trajectory generation model and the second parent trajectory generation model based on the difference in trajectory prediction performance scores between the first parent trajectory generation model and the second parent trajectory generation model, and the difference in the basic movement constraint logic, to obtain a fused trajectory generation model, includes: The basic movement constraint logic of the first parent trajectory generation model is determined as the first movement constraint logic, and the basic movement constraint logic of the second parent trajectory generation model is determined as the second movement constraint logic. If the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model, and the first movement constraint logic does not include the second movement constraint logic, the second movement constraint logic is identified as a defective movement constraint logic. The defective movement constraint logic of the second parent trajectory generation model is deleted to obtain the third parent trajectory generation model; The fused trajectory generation model is obtained by fusing the first parent trajectory generation model and the third parent trajectory generation model.

[0008] In some embodiments, after fusing the model based on the first parent trajectory generation model and the third parent trajectory generation model to obtain the fused trajectory generation model, the method further includes: If the second movement constraint logic does not include the first movement constraint logic, the first movement constraint logic is determined as the target movement constraint logic; Rules are added to the fused trajectory generation model according to the target movement constraint logic, so that the fused trajectory generation model has the target movement constraint logic.

[0009] In some embodiments, before performing model filtering based on the fused trajectory generation model and the similar historical trajectory generation model to obtain the target trajectory generation model, the method further includes: Obtain the historical scenes of the similar historical trajectory generation model, and identify the target scenes that do not belong to the historical scenes as the difference scenes; The success rate of the basic movement constraint logic of the similar historical trajectory generation model is evaluated based on the difference scenario to obtain the scenario trajectory prediction success rate. The basic movement constraint logic of the similar historical trajectory generation model is updated based on the success rate of the scene trajectory prediction to obtain the updated historical trajectory generation model. Replace the similar historical trajectory generation model with the updated historical trajectory generation model.

[0010] In some embodiments, the process of evaluating the trajectory prediction performance of each candidate trajectory generation model to obtain a trajectory prediction performance score includes: Based on the candidate trajectory generation model, trajectory prediction is performed to obtain the predicted trajectory; Based on the error between the predicted trajectory and the preset standard trajectory, a prediction accuracy score is calculated to obtain the trajectory prediction accuracy score. A compliance score is obtained by calculating rule scores based on the predicted trajectory and the basic movement constraint logic of the candidate trajectory generation model; wherein, the compliance score is inversely correlated with the number of basic movement constraint logic violations by the predicted trajectory; The trajectory prediction performance score is obtained by weighted summation of the trajectory prediction accuracy score and the compliance score.

[0011] In some embodiments, the step of selecting parent model pairs from among the candidate trajectory generation models based on the trajectory prediction performance score includes: Based on the trajectory prediction performance score, each candidate trajectory generation model is sorted from largest to smallest to obtain a trajectory generation model sequence. The first N candidate trajectory generation models in the trajectory generation model sequence are determined as the first parent trajectory generation model subsequence, and the other candidate trajectory generation models are determined as the second parent trajectory generation model subsequence; where N is a positive integer; A first model is selected from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the first parent trajectory generation model. A second model is selected from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the second parent trajectory generation model. The first parent trajectory generation model and the second parent trajectory generation model are determined as the parent model pair.

[0012] To achieve the above objectives, a second aspect of this application provides a controller, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0013] To achieve the above objectives, a third aspect of the present application provides a robot, which includes the controller described in the second aspect.

[0014] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0015] The trajectory prediction method, controller, robot, and storage medium proposed in this application identify agents (such as other robots, pedestrian agents, vehicle agents, etc.) that influence the movement path of the target robot in the target scene. They acquire the agent's basic constant velocity model and multiple basic movement constraint logics (such as "vehicles move along lane lines" and "pedestrians experience speed decay when encountering obstacles"). This allows for the integration of multiple constraint logics highly correlated with the agent's movement in the current scene into the basic constant velocity model, enabling targeted prediction of the agent's movement path and reducing the risk of collisions between the target robot and agents in the target scene, rather than using a fixed basic constant velocity model for trajectory prediction. Then, a large language model is invoked to generate a population of basic constant velocity models based on the target scene and basic movement constraint logic, resulting in a target population including multiple candidate trajectory generation models. This leverages the large language model's understanding and reasoning capabilities regarding scene semantics to generate a population of models containing diverse movement constraint logics, providing a diverse model foundation for subsequent evolutionary search. An evolutionary algorithm is then invoked to individually screen each candidate trajectory generation model in the target population, obtaining the target trajectory generation model. This allows for the generation of superior offspring models based on the target population through evolutionary algorithms. It enables the combination of effective movement constraint logic within the models and the selection of candidate trajectory generation models that perform well in trajectory prediction from the target population, thus generating the agent's predicted trajectory. This improves the accuracy of trajectory prediction based on this model, thereby enhancing the overall accuracy of trajectory prediction. Attached Figure Description

[0016] Figure 1 This is a flowchart of the trajectory prediction method provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart for step 104 in the document; Figure 3 yes Figure 2 The flowchart for step 201 in the document; Figure 4 yes Figure 2 The flowchart for step 202 in the document; Figure 5 yes Figure 2 The flowchart for step 203 in the text; Figure 6 This is a flowchart of a trajectory prediction method provided in another embodiment of this application; Figure 7 This is a flowchart of a trajectory prediction method provided in another embodiment of this application; Figure 8 This is a flowchart illustrating an application example provided in an embodiment of this application; Figure 9 This is a schematic diagram of the hardware structure of the controller provided in the embodiments of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of 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 not intended to limit the scope of this application.

[0018] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] First, let's analyze some of the terms used in this application: Artificial Intelligence (AI) is a new technical science that studies and develops theories, methods, technologies, and application systems for simulating, extending, and expanding human intelligence. AI is a branch of computer science that attempts to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can be a simulation of the information processes of human consciousness and thought. AI can also be the theory, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. Basic AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operating / interactive systems, and mechatronics. AI software technologies mainly include computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning. This application can acquire and process relevant data based on AI technology.

[0021] Large Language Models (LLMs) are deep learning models used for text generation and understanding. They can generate natural language text or understand the meaning of spoken language text. LLMs can improve their semantic understanding capabilities by training them on large text datasets.

[0022] During robot operation, it is often necessary to predict the trajectories of intelligent agents (such as other robots, pedestrian agents, etc.) to avoid collisions. Current trajectory prediction methods include hand-designed heuristics, such as social force models and constant velocity models. It's important to note that the social force model (SFM) is a dynamic model used to simulate the movement behavior of pedestrians or groups. The constant velocity model (CVM) is used to predict the trajectory of intelligent agents, its core assumption being that the speed of the agent (such as a smart car, robot, or elderly person in a senior care scenario) remains constant. However, these methods rely on human experience and trial-and-error processes. The trajectory prediction results obtained by these methods are easily influenced by subjective experience, resulting in limited accuracy, poor generalization ability, and difficulty in adapting to dynamically changing and complex scenarios.

[0023] Based on this, embodiments of this application provide a trajectory prediction method, controller, robot, and storage medium, aiming to improve the accuracy of trajectory prediction.

[0024] The trajectory prediction method, controller, robot, and storage medium provided in this application are specifically described through the following embodiments. First, the trajectory prediction method in the embodiments of this application is described.

[0025] It should be noted that in all specific embodiments of this application, when processing data related to user identity or characteristics, such as user information, user behavior data, user historical data, and user location information, user permission or consent is obtained first. Furthermore, the collection, use, and processing of this data comply with relevant laws, regulations, and standards. In addition, when embodiments of this application require access to sensitive personal information of users, separate permission or consent from the user is obtained through pop-ups or redirection to confirmation pages. Only after obtaining the user's separate permission or consent is the necessary user-related data required for the proper functioning of these embodiments acquired.

[0026] Figure 1 This is an optional flowchart of the trajectory prediction method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps 101 to 105.

[0027] Step 101: Identify the agents that influence the movement path of the target robot in the target scene; Step 102: Obtain the basic constant velocity model of the agent and obtain at least two basic movement constraint logics of the agent; Step 103: Invoke the large language model to generate a population of the basic constant velocity model based on the target scene and basic movement constraint logic to obtain the target population; wherein, the target population includes at least two candidate trajectory generation models; Step 104: Use an evolutionary algorithm to screen individual candidate trajectory generation models in the target population to obtain the target trajectory generation model; Step 105: Generate the predicted trajectory of the agent using the target trajectory generation model.

[0028] The beneficial effects of this application's embodiments include, but are not limited to: by identifying agents (such as other robots, pedestrian agents, vehicle agents, etc.) that influence the target robot's movement path in the target scene, and obtaining the agent's basic constant velocity model and multiple basic movement constraint logics (such as "vehicles move along lane lines" and "pedestrians experience speed decay when encountering obstacles"), multiple constraint logics highly correlated with the agent's movement in the current scene can be incorporated into the basic constant velocity model to make targeted predictions of the agent's movement path, reducing the risk of collision between the target robot and agents when moving in the target scene, rather than using a fixed basic constant velocity model for trajectory prediction. Then, a large language model is invoked to generate a population of basic constant velocity models based on the target scene and basic movement constraint logic, resulting in a target population including multiple candidate trajectory generation models. This leverages the large language model's understanding and reasoning capabilities of scene semantics to generate a population of models containing diverse movement constraint logics, providing a diverse model foundation for subsequent evolutionary search. An evolutionary algorithm is then invoked to individually screen each candidate trajectory generation model in the target population to obtain the target trajectory generation model. This allows for the generation of superior offspring models based on the target population through evolutionary algorithms. It enables the combination of effective movement constraint logic within the models and the selection of candidate trajectory generation models that perform well in trajectory prediction from the target population, thus generating the agent's predicted trajectory. This improves the accuracy of trajectory prediction based on this model, thereby enhancing the overall accuracy of trajectory prediction.

[0029] In step 101 of some embodiments, the target robot refers to the robot performing trajectory prediction. The target scene is the scene in which the target robot is located. Specifically, the target scene can be an indoor scene or an outdoor scene. It should be noted that when the target robot moves in the target scene, pedestrians, vehicles, other robots, etc. in the target scene may obstruct the movement of the target robot. Therefore, they can be treated as intelligent agents (such as pedestrian intelligent agents, vehicle intelligent agents, robot intelligent agents, etc.) and trajectory prediction can be performed on the intelligent agents.

[0030] In some embodiments, such as in a banking scenario within the fintech field, a bank robot moves through a bank lobby. The bank robot needs to predict the movement paths of customers in the lobby to avoid collisions. In this case, the target robot is specifically a bank robot, the target scenario is the bank lobby where the bank robot is located, and the intelligent agent is the customer in the bank lobby.

[0031] In another embodiment, for example, in the elderly care scenario within the healthcare field, target robots (such as meal delivery robots, medicine delivery robots, nursing robots, companion robots, etc.) can move within target settings such as senior living apartments and nursing homes to deliver meals, medications, medical equipment, and other items to the elderly. During their movement, the target robots need to predict the movement paths of pedestrians (such as the elderly) or other robots to avoid collisions. For instance, a meal delivery robot can start from the kitchen and deliver meals to a designated room; it needs to use trajectory prediction to anticipate the movement paths of elderly people walking in the corridor, other robots, or temporarily stored nursing equipment to avoid collisions. Similarly, when a nursing robot carries medical equipment to an elderly person's room, it needs to use trajectory prediction to avoid unexpected situations (such as an elderly person suddenly leaving the room) to ensure the safety of both the medical equipment and the personnel.

[0032] In some embodiments, the target scenario can be a complex scenario with dense crowds or numerous obstacles. Specifically, in the elderly care scenario within the healthcare field, the target scenario could be a restaurant in an elderly care facility, where elderly people move around in large numbers during mealtimes. The target scenario could also be an activity hall in an elderly care facility, where people are densely packed and their movements are chaotic during group activities. The target scenario could also be a corridor corner in an elderly care facility, where visibility is obstructed and elderly people are prone to suddenly crossing the corridor. It should be noted that in the aforementioned complex scenarios, the movement of the elderly is unpredictable (e.g., they may suddenly stop to talk or bend down to pick up items), and there may also be temporary obstacles such as wheelchairs or nursing carts, which places extremely high demands on the adaptability of trajectory prediction. This application's embodiments address such complex scenarios by incorporating multiple constraint logics highly correlated with the movement of the agent in the target scenario, based on a basic constant velocity model, to specifically predict the agent's movement path, thereby improving the adaptability of trajectory prediction.

[0033] In step 102 of some embodiments, the basic constant velocity model is a Constant Velocity Model (CVM). It should be noted that the constant velocity model is the simplest heuristic method in trajectory prediction. Its core assumption is that the agent's (such as a smart car, robot, or elderly person in a senior care scenario) speed remains constant; that is, the agent's future position can be calculated by summing the product of its current position, speed, and time. The basic constant velocity model provides fundamental trajectory generation logic, giving Large Language Models (LLMs) a clear starting point for expansion and optimization.

[0034] In some embodiments, the basic motion constraint logic is a natural language description of the scene rules of the target robot's environment. For example, the basic motion constraint logic may include, but is not limited to, any one or more of the following: "When the agent type is a wheelchair, the maximum speed does not exceed 0.8 meters per second (m / s)," "If a static obstacle (such as a hospital bed) is detected in the semantic map, the predicted trajectory must maintain a distance of at least 0.5 meters from the static obstacle," "In outdoor rainy weather scenarios, the vehicle's speed is reduced by 20%," and so on. Specifically, the basic motion constraint logic is data in text form. Text format facilitates understanding and parsing by the LLM model, enabling it to transform these basic motion constraint logics into conditional judgments or parameter adjustment logic in heuristic code (i.e., candidate trajectory generation models).

[0035] In step 103 of some embodiments, the target population is a collection of multiple candidate trajectory generation models. A candidate trajectory generation model is a new model generated by integrating basic movement constraint logic into a basic constant velocity model using a large language model. Specifically, the candidate trajectory generation model can be heuristic code. In some embodiments, any two candidate trajectory generation models may have different basic movement constraint logic.

[0036] For example, when an LLM model generates a target population, its input can include three parts: First, a detailed description of the basic constant velocity model, such as a textual explanation of "predicted position = current position + velocity * time," or the corresponding simplified code. Second, textual rules for the basic movement constraint logic, such as the correspondence between agent type and velocity limits, and basic principles of obstacle avoidance. Third, details of the trajectory prediction task, such as the input needing to include historical positions, agent types, and semantic map data, the output needing to be the position coordinates for the next 10 time steps, and the optimization objective being to minimize prediction error while satisfying safety constraints, etc. Based on these inputs, LLM, by understanding the logic of the basic constant velocity model and the multimodal basic movement constraint logic, generates N different heuristic codes (i.e., candidate trajectory generation models) as outputs. Each code incorporates different basic movement constraint logic on top of the basic constant velocity model. For example, some candidate trajectory generation models focus on the influence of agent type on velocity, while others focus on the adjustment of trajectory offset by obstacle distance. This allows for the formation of diverse target populations, providing a rich starting point for subsequent evolutionary searches.

[0037] In step 104 of some embodiments, the target trajectory generation model is the final model selected from multiple candidate trajectory generation models through an evolutionary algorithm. It should be noted that an evolutionary algorithm (EA) is an optimization algorithm that mimics the natural biological evolution mechanism. Evolutionary algorithms utilize the adaptive differences among individuals in a population to gradually approach the optimal solution. The process of individual selection can be referred to the detailed description of steps 201 to 205 below, and will not be repeated here.

[0038] In step 105 of some embodiments, the target robot may execute a target trajectory generation model to obtain a predicted trajectory for the agent output by the target trajectory generation model.

[0039] Please see Figure 2 In some embodiments, step 104 may include, but is not limited to, steps 201 to 205: Step 201: Evaluate the trajectory prediction performance of each candidate trajectory generation model to obtain a trajectory prediction performance score. Step 202: Based on the trajectory prediction performance score, the model is screened among the candidate trajectory generation models to obtain parent model pairs; wherein, the parent model pairs include the first parent trajectory generation model and the second parent trajectory generation model. Step 203: Based on the difference in trajectory prediction performance scores between the first parent trajectory generation model and the second parent trajectory generation model, as well as the difference in basic movement constraint logic, the first parent trajectory generation model and the second parent trajectory generation model are fused to obtain a fused trajectory generation model. Step 204: Obtain the initial historical trajectory generation model, and search for similar scene models in the initial historical trajectory generation model according to the target scene to obtain a similar historical trajectory generation model; Step 205: Based on the fusion trajectory generation model and the similar historical trajectory generation model, the model is selected to obtain the target trajectory generation model.

[0040] The advantage of this embodiment lies in that by evaluating the trajectory prediction performance of each candidate trajectory generation model and obtaining a trajectory prediction performance score, the predictive ability of each model can be objectively measured based on quantitative indicators, providing a basis for subsequent selection. Parent model pairs are selected from each candidate trajectory generation model based on the trajectory prediction performance score, ensuring that subsequent fusion operations are based on diverse individuals with superior performance, thus improving evolutionary efficiency. Then, model fusion is performed based on the performance differences and differences in basic movement constraint logic between parent models. This allows for the generation of more adaptive and diverse fusion models by combining the advantageous constraint logic of different models. Furthermore, by acquiring the initial historical trajectory generation models and searching for similar historical models of the target scene, effective experience from historical scenes can be introduced, enhancing the model's adaptability to the target scene. By selecting the target model from the fused trajectory generation model and similar historical trajectory generation models, the current evolutionary results and historical experience can be combined to select the trajectory generation model with higher prediction accuracy, thereby improving the accuracy of trajectory prediction.

[0041] In step 201 of some embodiments, the trajectory prediction performance score is used to characterize the ability of the candidate trajectory generation model to generate trajectories. For example, the trajectory prediction performance score can be the sum of the trajectory prediction accuracy score and the compliance score. The calculation process of the trajectory prediction performance score can be referred to the detailed description of steps 301 to 304 below, and will not be repeated here.

[0042] In step 202 of some embodiments, the first parent trajectory generation model and the second parent trajectory generation model are models selected from multiple candidate trajectory generation models, respectively, for model fusion. For example, two models can be randomly selected from candidate trajectory generation models whose trajectory prediction performance scores are greater than a preset threshold, serving as the first parent trajectory generation model and the second parent trajectory generation model. Alternatively, different selection probabilities can be generated based on different trajectory prediction performance scores for model selection, wherein the trajectory prediction performance score of the candidate trajectory generation model is positively correlated with the selection probability. In some embodiments, for example, candidate trajectory generation models with trajectory prediction performance scores greater than a preset threshold can be considered elite individuals, and their selection probability can be set to 30%. Candidate trajectory generation models that are not elite individuals but can run successfully can be considered ordinary individuals, and their selection probability can be set to 70%. Then, based on the 30% and 70% probabilities, one model is selected from the elite individuals and the ordinary individuals to obtain the first parent trajectory generation model. The selection process for the second parent trajectory generation model is consistent with the selection process for the first parent trajectory generation model, and will not be repeated here. In another embodiment, model selection can also be performed in other ways, not limited to these.

[0043] In step 203 of some embodiments, the fused trajectory generation model is a model obtained by fusing the first parent trajectory generation model and the second parent trajectory generation model. Specifically, the performance of the first parent trajectory generation model and the second parent trajectory generation model can be determined based on the difference between their trajectory prediction performance scores. Furthermore, the differences in the basic motion constraint logic of these two models can be compared to ensure that the fused trajectory generation model retains the basic motion constraint logic of the parent model with better performance and eliminates the basic motion constraint logic of the parent model with poorer performance, allowing effective strategies from the parent models to be combined. In another embodiment, effective constraint logic with higher frequency among multiple basic motion constraint logics can be determined statistically to increase the weight of this strategy in the descendant model (i.e., the fused trajectory generation model). For example, suppose that constraint logic k1 of parent model A (i.e., the first parent trajectory generation model) has a higher frequency and can adapt to the low-speed characteristics of wheelchairs. Constraint logic k2 of parent model B (i.e., the second parent trajectory generation model) has a higher frequency and can adapt to corridor corner scenarios. Then, by judging whether the applicable scenarios of these high-frequency logics (such as the agent type being a wheelchair, the environment including corridor corner scenarios, etc.) are consistent with the current scenario (i.e. the target scenario), if they are consistent, the constraint logic k1 in parent model A that adapts to wheelchairs and the constraint logic k2 in parent model B that handles corridor corners can be merged. Furthermore, redundant or conflicting parts in the low-frequency logic of both can be eliminated to generate a descendant model that retains the high-quality logic of the parent model and conforms to statistical laws.

[0044] In step 204 of some embodiments, the similar historical trajectory generation model refers to the initial historical trajectory generation model that is similar to the target scene. The similarity between the initial historical trajectory generation model and the target scene can be analyzed using a Large Language Model (LLM), and models with a similarity greater than a preset threshold are considered similar historical trajectory generation models. Other methods can also be used to find similar scene models, and are not limited to these.

[0045] In some embodiments, after obtaining multiple similar historical trajectory generation models, a probability distribution (such as a Softmax distribution) can be generated based on the historical performance (such as historical trajectory prediction performance score) of the similar historical trajectory generation models. The final similar historical trajectory generation model can then be obtained by sampling from the multiple similar historical trajectory generation models based on the probability distribution. This allows for the priority selection of historical models with higher performance scores and similar application scenarios, thereby improving the diversity and accuracy of trajectory prediction.

[0046] In step 205 of some embodiments, the target trajectory generation model is a model selected from the fused trajectory generation model and the similar historical trajectory generation model. In some embodiments, the trajectory prediction performance of the fused trajectory generation model and the similar historical trajectory generation model can be evaluated separately to obtain trajectory prediction performance scores for the fused trajectory generation model and the similar historical trajectory generation model, respectively. Then, the model with the highest trajectory prediction performance score is selected as the target trajectory generation model.

[0047] Please see Figure 3 In some embodiments, step 201 may include, but is not limited to, steps 301 to 304: Step 301: Perform trajectory prediction based on the candidate trajectory generation model to obtain the predicted trajectory; Step 302: Calculate the prediction accuracy score based on the error between the predicted trajectory and the preset standard trajectory to obtain the trajectory prediction accuracy score. Step 303: Calculate the rule score based on the basic movement constraint logic of the predicted trajectory and candidate trajectory generation model to obtain the compliance score; wherein, the compliance score is inversely correlated with the number of basic movement constraint logic violations by the predicted trajectory; Step 304: The trajectory prediction performance score is obtained by weighted summation of the trajectory prediction accuracy score and the compliance score.

[0048] The advantage of this embodiment lies in that it predicts the trajectory based on the candidate trajectory generation model and calculates the error between the predicted trajectory and the preset standard trajectory to obtain a trajectory prediction accuracy score. This quantifies the model's accuracy in trajectory prediction, reflecting its deviation from the actual trajectory. Furthermore, it calculates a compliance score based on violations of the basic movement constraint logic (such as collision avoidance and lane rule compliance) of the predicted trajectory and the candidate trajectory generation model, with the compliance score inversely correlated with the number of constraint violations. This assesses the model's adherence to preset movement rules, ensuring that the generated trajectory meets the safety and rationality requirements of the scenario. Then, by weighted summing the trajectory prediction accuracy score and the compliance score, a comprehensive trajectory prediction performance score is obtained. This calculated performance score not only reflects the model's prediction accuracy but also incorporates rule compliance, making the model evaluation more comprehensive and thus improving the overall comprehensiveness of trajectory prediction.

[0049] In step 301 of some embodiments, the predicted trajectory is the trajectory output by the candidate trajectory generation model after trajectory prediction is performed. The predicted trajectory is used to characterize the possible movement paths of the agent.

[0050] In step 302 of some embodiments, the preset standard trajectory is the actual movement path traversed by the agent. In some embodiments, the minimum average displacement error (minADE) between the predicted trajectory and the preset standard trajectory can be calculated. For example, the predicted trajectory includes the predicted position of the agent at each time step, and the standard trajectory includes the actual position of the agent at each time step. In this case, the average difference between the predicted position and the actual position at each time step can be calculated to obtain the average displacement error. Then, the minimum value of the average displacement errors among the candidate trajectory generation models in the target population can be selected as the minimum average displacement error (minADE). The minimum average displacement error (minADE) can be determined as the trajectory prediction accuracy score.

[0051] In another embodiment, other types of errors can also be calculated, such as calculating the minimum final distance error (minFDE) between the predicted trajectory and the preset standard trajectory, but not limited thereto.

[0052] In step 303 of some embodiments, a compliance score is used to characterize the degree to which the predicted trajectory violates the underlying motion constraint logic. The compliance score is inversely correlated with the number of underlying motion constraint logics violated by the predicted trajectory. For example, suppose the candidate trajectory generation model is candidate model M1, and the underlying motion constraint logic of candidate model M1 includes speed limit logic ("the speed of the wheelchair agent does not exceed 0.8 meters per second (m / s), obstacle avoidance logic ("the distance between the robot and the hospital bed is greater than or equal to 0.5 meters"), and map rules ("the intelligent vehicle does not deviate from the lane line"). If the predicted trajectory generated by candidate model M1 violates only one of these logics, then the compliance score can be set to 70 points. If the predicted trajectory generated by candidate model M1 violates all three of these logics, then the compliance score can be set to 0 points.

[0053] In step 304 of some embodiments, the weight of the trajectory prediction accuracy score can be set to 60%, the weight of the compliance score can be set to 40%, and the weighted sum can be obtained to obtain the trajectory prediction performance score.

[0054] Please see Figure 4 In some embodiments, step 202 may include, but is not limited to, steps 401 to 404: Step 401: Sort each candidate trajectory generation model from largest to smallest according to the trajectory prediction performance score to obtain the trajectory generation model sequence. Step 402: Determine the first N candidate trajectory generation models in the trajectory generation model sequence as the first parent trajectory generation model subsequence, and determine the other candidate trajectory generation models as the second parent trajectory generation model subsequence; where N is a positive integer; Step 403: Select a first model from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the first parent trajectory generation model; select a second model from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the second parent trajectory generation model. Step 404: The first parent trajectory generation model and the second parent trajectory generation model are determined as the parent model pair.

[0055] The advantage of this embodiment lies in that each candidate trajectory generation model is sorted from highest to lowest based on its trajectory prediction performance score, resulting in a trajectory generation model sequence. The top N models in this sequence are designated as the first parent trajectory generation model subsequence, and the remaining models are designated as the second parent trajectory generation model subsequence. This distinguishes between high-performance model groups and ordinary model groups, providing a foundation for subsequent parent model selection. Then, from the first and second parent trajectory generation model subsequences, first and second model selections are performed respectively to obtain parent model pairs. This approach not only considers models from the high-performance subsequence to retain superior prediction characteristics during the selection process but also has the opportunity to introduce diversity features from the ordinary model subsequence, avoiding selection bias. In other words, there is a certain probability of selecting high-scoring excellent models, and a certain probability of selecting other models, thereby increasing the diversity of parent model pairs and consequently improving the diversity of trajectory prediction results.

[0056] In step 401 of some embodiments, the trajectory generation model sequence includes candidate trajectory generation models sorted from largest to smallest according to trajectory prediction performance scores.

[0057] In step 402 of some embodiments, N is a positive integer, specifically, N can be 10. For example, the sequence of the first 10 candidate trajectory generation models can be determined as the first parent trajectory generation model subsequence, and the sequence of other candidate trajectory generation models after the 10th model can be determined as the second parent trajectory generation model subsequence. It is evident that the trajectory prediction performance score of each model in the first parent trajectory generation model subsequence is greater than or equal to the trajectory prediction performance score of any model in the second parent trajectory generation model subsequence. The first parent trajectory generation model subsequence is a sequence composed of models with high trajectory prediction performance scores. The second parent trajectory generation model subsequence is a sequence composed of models with average or low trajectory prediction performance scores.

[0058] In step 403 of some embodiments, a model can be selected from either the first parent trajectory generation model subsequence or the second parent trajectory generation model subsequence as the first parent trajectory generation model. Specifically, the models in the first parent trajectory generation model subsequence can be considered elite individuals, with the selection probability of elite individuals set to 30%, while the models in the second parent trajectory generation model subsequence can be considered ordinary individuals, with the selection probability of ordinary individuals set to 70%. Then, based on the probabilities of 30% and 70%, a model is selected from each elite individual and ordinary individual to obtain the first parent trajectory generation model. The selection process for the second parent trajectory generation model is consistent with the selection process for the first parent trajectory generation model, and will not be repeated here.

[0059] In step 404 of some embodiments, the parent model pair includes a first parent trajectory generation model and a second parent trajectory generation model.

[0060] Please see Figure 5 In some embodiments, step 203 may include, but is not limited to, steps 501 to 504: Step 501: Determine the basic movement constraint logic of the first parent trajectory generation model as the first movement constraint logic, and determine the basic movement constraint logic of the second parent trajectory generation model as the second movement constraint logic; Step 502: If the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model, and the first movement constraint logic does not contain the second movement constraint logic, then the second movement constraint logic is identified as the defective movement constraint logic. Step 503: Delete the defective movement constraint logic of the second parent trajectory generation model to obtain the third parent trajectory generation model. Step 504: Perform model fusion based on the first parent trajectory generation model and the third parent trajectory generation model to obtain the fused trajectory generation model.

[0061] The advantage of this embodiment lies in that, when the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model, and the basic movement constraint logic (i.e., the first movement constraint logic) of the first parent trajectory generation model does not contain the basic movement constraint logic (the second movement constraint logic) of the second parent trajectory generation model, the second movement constraint logic is identified as a defective movement constraint logic and removed from the second parent trajectory generation model to obtain the third parent trajectory generation model. This allows for the identification and removal of poor strategies in low-performance models based on performance differences and logical inclusion relationships, preventing poor movement constraint logic from affecting the fusion process. Then, model fusion is performed based on the first and third parent trajectory generation models to obtain a fused trajectory generation model, generating a trajectory prediction model more adapted to the target scenario. This improves the accuracy of trajectory prediction in the target scenario, further enhancing the accuracy and specificity of trajectory prediction.

[0062] In step 501 of some embodiments, the first movement constraint logic is the basic movement constraint logic possessed by the first parent trajectory generation model. The second movement constraint logic is the basic movement constraint logic possessed by the second parent trajectory generation model.

[0063] In step 502 of some embodiments, if the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model, it indicates that the trajectory prediction performance of the first parent trajectory generation model is superior. Therefore, logic possessed by the second parent trajectory generation model but not by the first parent trajectory generation model is used as constraint logic to reduce model performance. Defective movement constraint logic is logic that leads to poor trajectory prediction performance of the second parent trajectory generation model. For example, suppose the first movement constraint logic sets a speed limit for the wheelchair agent, and the second movement constraint logic does not set a speed limit for the wheelchair agent, and when predicting the trajectory of the wheelchair agent, the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model. Then, it can be determined that the second movement constraint logic is the logic that leads to poor performance of the second parent trajectory generation model.

[0064] In step 503 of some embodiments, the third parent trajectory generation model is the model obtained after deleting the defect movement constraint logic from the second parent trajectory generation model. Specifically, the code portion corresponding to the defect movement constraint logic in the second parent trajectory generation model can be deleted or adjusted.

[0065] In step 504 of some embodiments, a fused trajectory generation model is obtained by fusing the first parent trajectory generation model and the third parent trajectory generation model to eliminate defective movement constraint logic that has been verified to affect the model's prediction performance.

[0066] Please see Figure 6 In some embodiments, after step 504, the trajectory prediction method may also include, but is not limited to, steps 601 to 602: Step 601: If the second movement constraint logic does not contain the first movement constraint logic, the first movement constraint logic is determined as the target movement constraint logic. Step 602: Add rules to the fusion trajectory generation model according to the target movement constraint logic so that the fusion trajectory generation model has the target movement constraint logic.

[0067] The advantage of this embodiment lies in determining whether the second movement constraint logic includes the first movement constraint logic, and identifying the first movement constraint logic as the target movement constraint logic when the second movement constraint logic does not include the first movement constraint logic. This identifies the unique and effective constraint logic in the high-performance parent model (i.e., the first parent trajectory generation model). Then, rules are added to the fused trajectory generation model according to the target movement constraint logic, ensuring that the fused model contains verified and effective movement constraint logic and avoiding the loss of key movement constraint logic during model fusion. This preserves the high-quality strategies of the parent model, thereby enhancing the scene adaptability of the fused model and improving the accuracy of trajectory prediction in the target scene.

[0068] In step 601 of some embodiments, the target movement constraint logic is a constraint logic that the first parent trajectory generation model possesses but the second parent trajectory generation model does not. Since the trajectory prediction performance score of the first parent trajectory generation model is higher than that of the second parent trajectory generation model, meaning the trajectory prediction performance of the first parent trajectory generation model is superior, the logic possessed by the first parent trajectory generation model but not by the second parent trajectory generation model is used as the constraint logic that leads to performance improvement. For example, suppose the first movement constraint logic is an obstacle avoidance logic that "dynamically adjusts the trajectory offset based on the obstacle distance," while the second movement constraint logic does not include this obstacle avoidance logic. Then, the obstacle avoidance logic can be used as the target movement constraint logic, i.e., effective logic that can improve the model's prediction performance.

[0069] In step 602 of some embodiments, code can be generated based on the target movement constraint logic, and the code corresponding to the target movement constraint logic can be added to the fusion trajectory generation model.

[0070] Please see Figure 7 In some embodiments, prior to step 205, the trajectory prediction method may include, but is not limited to, steps 701 to 704: Step 701: Obtain the historical scenes of the similar historical trajectory generation model, and identify the target scenes that do not belong to the historical scenes as the difference scenes; Step 702: Evaluate the success rate of the basic movement constraint logic of the similar historical trajectory generation model based on the differences in scenarios to obtain the success rate of scenario trajectory prediction. Step 703: Update the basic movement constraint logic of the similar historical trajectory generation model based on the success rate of scene trajectory prediction to obtain the updated historical trajectory generation model. Step 704: Replace the similar historical trajectory generation model with the updated historical trajectory generation model.

[0071] The advantage of this embodiment lies in acquiring historical scenarios from similar historical trajectory generation models and identifying the current scenario (i.e., the target scenario) that does not belong to these historical scenarios as differing scenarios. This allows for the identification of specific scenario differences between the current scenario and the historical scenarios targeted by the similar historical trajectory generation models. The success rate of the basic movement constraint logic of the similar historical trajectory generation model is evaluated based on the differing scenarios, yielding the scenario trajectory prediction success rate. This quantifies the adaptability of the historical model in new scenarios. Then, the basic movement constraint logic is updated based on the success rate, resulting in an updated historical trajectory generation model. This allows for dynamic adjustment of the constraint logic parameters or weights to optimize the model's performance in differing scenarios, making the historical model more adaptable to the current trajectory generation task scenario and improving the accuracy of trajectory prediction in the target scenario.

[0072] In step 701 of some embodiments, the historical scene is the scene targeted by the similar historical trajectory generation model. For example, suppose the similar historical trajectory generation model is used to predict an agent moving in a corridor of a nursing home, then the historical scene is the corridor of the nursing home. The difference scene is the target scene that does not belong to the historical scene. For example, suppose the target scene includes the corridor and the dining room of the nursing home, then the difference scene is the dining room of the nursing home.

[0073] In step 702 of some embodiments, the scene trajectory prediction success rate refers to the success rate of the similar historical trajectory generation model in predicting the trajectories of one or more agents in different scenes. It should be noted that if the error between the predicted trajectory output by the similar historical trajectory generation model and the actual trajectory of the agent is less than a preset error threshold, the prediction is considered successful.

[0074] In step 703 of some embodiments, the weights of the basic motion constraint logic in the similar historical trajectory generation model can be adjusted based on the success rate of scene trajectory prediction. In some embodiments, the success rate of scene trajectory prediction is positively correlated with the weights of the basic motion constraint logic.

[0075] In step 704 of some embodiments, the similar historical trajectory generation model is replaced with an updated historical trajectory generation model so as to use the updated historical trajectory generation model for subsequent model individual selection.

[0076] Please see Figure 8 In one application example, the trajectory prediction method may include the following steps: Initialize population generation; specifically, the population can be generated by combining a large language model (LLM) with a constant velocity model and basic movement constraint logic, which is to generate the target population mentioned above.

[0077] Select parent individuals; specifically, select model individuals that can run successfully from the target population with a 70% probability, and select elite model individuals, that is, model individuals with high trajectory prediction performance scores, with a 30% probability.

[0078] Cross-fusion; specifically, the LLM model can be used to fuse the parent individuals' models (also known as short-term reflection) to generate offspring models, that is, to generate the fusion trajectory generation model mentioned above.

[0079] Mutation optimization; specifically, the model can be obtained by sampling from the historical elite model (i.e., the similar historical trajectory generation model mentioned above) through the LLM model, and the model can be optimized and adjusted to obtain the updated historical trajectory generation model mentioned above.

[0080] Evaluate the performance of each trajectory generation model; specifically, the trajectory prediction performance of the model (i.e., the heuristic code) can be evaluated using metrics such as minimum average distance error (minADE).

[0081] The distribution of indicators such as minADE is fed back to the LLM model using the Statistical Feedback Loop (SFL) technique.

[0082] Based on the evaluated performance, the target population is updated and iterated to select new parent models from the next generation.

[0083] When the termination conditions are met, the optimized executable recursive code, i.e., the target trajectory generation model, is output as the final result. Specifically, there are three main types of termination conditions: First, reaching the preset maximum number of evolutionary generations (e.g., 100 generations); second, population performance convergence, i.e., the improvement in the optimal individual comprehensive score (minADE + compliance) over multiple consecutive generations (e.g., 10 generations) is less than a threshold (e.g., 1%), indicating that further iterations are unlikely to significantly optimize the model; third, the performance of the generated heuristic code on the test set reaches the preset target (e.g., minADE less than or equal to 0.3 meters, and obstacle avoidance success rate greater than or equal to 99%), meeting the safety and accuracy requirements of the elderly care scenario. The iteration terminates when any one of these conditions is met.

[0084] This application also provides a controller, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the trajectory prediction method described above. The controller can include any smart terminal such as a tablet computer or an in-vehicle computer.

[0085] Please see Figure 9 , Figure 9 The hardware structure of a controller according to another embodiment is illustrated. The controller includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the operating system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 using the trajectory prediction method of the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0086] This application also provides a robot, which includes the controller described above.

[0087] The specific implementation of this robot is basically the same as the specific implementation of the trajectory prediction method described above, and will not be repeated here.

[0088] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the trajectory prediction method described above.

[0089] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0090] It should be noted that the software tools or components not belonging to our company that appear in the embodiments of this application are merely examples and do not represent actual use.

[0091] The embodiments described in this application are intended to more clearly illustrate the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. Those skilled in the art will know that with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0092] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0093] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0095] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0096] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0097] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, or indirect coupling or communication connection between the apparatus or units, and may be electrical, mechanical, or other forms.

[0098] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0099] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0100] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A trajectory prediction method, characterized in that, The method includes: Identify the agents that influence the movement path of the target robot in the target scenario; Obtain the basic constant velocity model of the agent, and obtain at least two basic movement constraint logics of the agent; The large language model is invoked to generate a population of the basic constant velocity model based on the target scene and the basic movement constraint logic, thereby obtaining a target population; wherein, the target population includes at least two candidate trajectory generation models; An evolutionary algorithm is invoked to perform individual screening on each of the candidate trajectory generation models in the target population to obtain the target trajectory generation model; The target trajectory generation model generates the predicted trajectory of the agent.

2. The method according to claim 1, characterized in that, The step of invoking an evolutionary algorithm to perform individual screening on each candidate trajectory generation model in the target population to obtain the target trajectory generation model includes: The trajectory prediction performance of each candidate trajectory generation model is evaluated to obtain a trajectory prediction performance score. Based on the trajectory prediction performance score, model selection is performed among the candidate trajectory generation models to obtain parent model pairs; wherein, the parent model pair includes a first parent trajectory generation model and a second parent trajectory generation model; Based on the differences in trajectory prediction performance scores between the first parent trajectory generation model and the second parent trajectory generation model, as well as the differences in the basic movement constraint logic, the first parent trajectory generation model and the second parent trajectory generation model are fused to obtain a fused trajectory generation model. Obtain an initial historical trajectory generation model, and search for similar scene models in the initial historical trajectory generation model according to the target scene to obtain a similar historical trajectory generation model; The target trajectory generation model is obtained by filtering models based on the fused trajectory generation model and the similar historical trajectory generation model.

3. The method according to claim 2, characterized in that, The step of fusing the first parent trajectory generation model and the second parent trajectory generation model based on the difference in trajectory prediction performance scores between the first parent trajectory generation model and the second parent trajectory generation model, and the difference in the basic movement constraint logic, to obtain a fused trajectory generation model, includes: The basic movement constraint logic of the first parent trajectory generation model is determined as the first movement constraint logic, and the basic movement constraint logic of the second parent trajectory generation model is determined as the second movement constraint logic. If the trajectory prediction performance score of the first parent trajectory generation model is greater than that of the second parent trajectory generation model, and the first movement constraint logic does not include the second movement constraint logic, the second movement constraint logic is identified as a defective movement constraint logic. The defective movement constraint logic of the second parent trajectory generation model is deleted to obtain the third parent trajectory generation model; The fused trajectory generation model is obtained by fusing the first parent trajectory generation model and the third parent trajectory generation model.

4. The method according to claim 3, characterized in that, After fusing the first parent trajectory generation model and the third parent trajectory generation model to obtain the fused trajectory generation model, the method further includes: If the second movement constraint logic does not include the first movement constraint logic, the first movement constraint logic is determined as the target movement constraint logic; Rules are added to the fused trajectory generation model according to the target movement constraint logic, so that the fused trajectory generation model has the target movement constraint logic.

5. The method according to claim 2, characterized in that, Before performing model filtering based on the fused trajectory generation model and the similar historical trajectory generation model to obtain the target trajectory generation model, the method further includes: Obtain the historical scenes of the similar historical trajectory generation model, and identify the target scenes that do not belong to the historical scenes as the difference scenes; The success rate of the basic movement constraint logic of the similar historical trajectory generation model is evaluated based on the difference scenario to obtain the scenario trajectory prediction success rate. The basic movement constraint logic of the similar historical trajectory generation model is updated based on the success rate of the scene trajectory prediction to obtain the updated historical trajectory generation model. Replace the similar historical trajectory generation model with the updated historical trajectory generation model.

6. The method according to claim 2, characterized in that, The step of evaluating the trajectory prediction performance of each candidate trajectory generation model to obtain a trajectory prediction performance score includes: Based on the candidate trajectory generation model, trajectory prediction is performed to obtain the predicted trajectory; Based on the error between the predicted trajectory and the preset standard trajectory, a prediction accuracy score is calculated to obtain the trajectory prediction accuracy score. A compliance score is obtained by calculating rule scores based on the predicted trajectory and the basic movement constraint logic of the candidate trajectory generation model; wherein, the compliance score is inversely correlated with the number of basic movement constraint logic violations by the predicted trajectory; The trajectory prediction performance score is obtained by weighted summation of the trajectory prediction accuracy score and the compliance score.

7. The method according to claim 2, characterized in that, The step of selecting parent model pairs from among the candidate trajectory generation models based on the trajectory prediction performance score includes: Based on the trajectory prediction performance score, each candidate trajectory generation model is sorted from largest to smallest to obtain a trajectory generation model sequence. The first N candidate trajectory generation models in the trajectory generation model sequence are determined as the first parent trajectory generation model subsequence, and the other candidate trajectory generation models are determined as the second parent trajectory generation model subsequence; where N is a positive integer; A first model is selected from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the first parent trajectory generation model. A second model is selected from the first parent trajectory generation model subsequence and the second parent trajectory generation model subsequence to obtain the second parent trajectory generation model. The first parent trajectory generation model and the second parent trajectory generation model are determined as the parent model pair.

8. A controller, characterized in that, The controller includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the trajectory prediction method according to any one of claims 1 to 7.

9. A robot, characterized in that, The robot includes the controller as described in claim 8.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the trajectory prediction method according to any one of claims 1 to 7.

Images