Robot control method, terminal equipment and medium
By automatically generating and correcting robot motion models and control strategies through pre-trained models, the consistency and adaptability issues in the construction of behavior tree systems are solved, enabling efficient and reliable robot task execution.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NAT UNIV OF DEFENSE TECH
- Filing Date
- 2026-03-13
- Publication Date
- 2026-04-28
AI Technical Summary
In existing technologies, behavior tree systems rely on expert knowledge for construction, the action model and execution strategy are inconsistent, the system has poor adaptability, and existing methods have high computational complexity or lack physical environment interaction feedback, which makes the generated plan unexecutable.
Candidate action models and control strategies are automatically generated using pre-trained large language models and visual language models. These are then verified and corrected using behavior tree planning algorithms. By combining planning and execution contexts, the consistency and completeness of the action models and control strategies are ensured.
It reduces the reliance on manually defined motion models by domain experts, improves the system's adaptability and execution reliability, solves the problems of illusion and execution failure in robot control, and is suitable for industrial, logistics and service scenarios.
Smart Images
Figure CN121928565A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a robot control method, terminal equipment, and medium. Background Technology
[0002] In the field of robotics, reliable high-level planning and robust low-level control strategies are required. In recent years, Behavior Trees (BTs) have gained popularity as an intelligent robot control architecture due to their modularity, interpretability, high responsiveness, and safety. BT planning methods provide theoretical guarantees for generating reliable BTs to automatically complete tasks.
[0003] However, traditional BT planning methods typically assume the existence of a well-designed, "grounded" BT system, which includes a well-defined high-level action model and effective low-level control strategies. In practice, building such a BT system often requires domain-specific expert knowledge and significant human investment to manually define atomic actions, their preconditions, and effects.
[0004] The main challenge currently faced is known as the "BT Grounding Problem," which refers to how to automatically construct a BT system that satisfies both completeness (the ability to plan solutions) and consistency (the ability for the planned actions to be actually executed by the underlying controller) for a given task set. Existing methods, such as evolutionary computation or reinforcement learning, are typically computationally complex and struggle to guarantee the logical reliability of task execution. While methods based on Large Language Models (LLMs) demonstrate reasoning capabilities, they often lack interactive feedback with the physical environment, resulting in the generated plans being unexecutable on actual robots (i.e., illusions or physical infeasibility). Summary of the Invention
[0005] This invention aims to provide a robot control method, terminal device, and medium to solve the problems of behavior tree system construction relying on expert knowledge, inconsistency between action models and execution strategies, and poor system adaptability in the prior art.
[0006] In a first aspect, the present invention provides a robot control method, the method comprising the following steps: Obtain a task set consisting of at least one task to be performed by the robot; each task to be performed by the robot includes an initial physical state and a target physical state; the task to be performed by the robot is at least an industrial assembly task or a logistics sorting task. A candidate action model set is generated based on the task set using the first pre-trained model, and the behavior tree planning algorithm is used to verify whether the candidate action model set is sufficient to plan a solution to complete all the tasks to be performed by the robot in the task set, and a planning context is generated. The candidate action model set includes at least one candidate action model, and each candidate action model represents a repeatable robot atomic operation, including the preconditions required to perform the robot atomic operation, the physical state expected to be added to the environment after execution, and the physical state expected to be removed from the environment after execution. Using a second pre-trained model, based on the visual observation data of the environment and the planning context, control policies corresponding to candidate action models are sampled from the control policy space; the control policy is the control parameters that drive the movement of robot joints or end effectors; the visual observation data includes objects and the spatial relationships between objects; The control strategy is executed in a simulated or physical environment. An execution context is generated based on the feedback of the environment state after execution, and the consistency between the actual physical state changes generated by the control strategy and the expected physical state changes defined by the candidate action model is verified. When the validation results are inconsistent, the candidate action model is modified by using the second pre-trained model and combining the planning context and execution context until a behavior tree system that meets the requirements of completeness and consistency is generated. The behavior tree system is parsed, and the behavior tree nodes are converted into driving instructions for the robot. The robot is then driven to perform operations based on the driving instructions to complete the task set.
[0007] Optionally, the first pre-trained model is a large language model trained on a large scale of text and code, used for symbolic reasoning and logical planning; The second pre-trained model is a visual language model jointly trained with visual language multimodal data, used to understand physical scenes and generate executable instructions or parameters related to physical scenes.
[0008] Optionally, when validation results are inconsistent, the candidate action model is modified using a second pre-trained model, combined with the planning and execution contexts, including: If a control strategy that meets the consistency requirement cannot be sampled for a candidate action model within a preset number of times, the candidate action model is determined to be infeasible in the current context. The planning context and execution context are input as prompts into the second pre-trained model. The planning context includes at least one or more of the following information: task information that failed due to lack of effective actions during behavior tree planning, action sequences that have been attempted to expand but failed to achieve the goal, or physical state conditions that are still missing to achieve the goal. The execution context includes at least one or more of the following information: images or point cloud data of the physical state of the environment after execution obtained by robot sensors or simulation environment engine, error codes or warning messages triggered during the execution of control strategies, or descriptions of the differences between actual physical state changes and expected physical state changes. The modified candidate action model is inferred using the second pre-trained model. The modified action model includes modifications to the preconditions, added physical states, or deleted physical states of the original candidate action model, so that the modified candidate action model is physically feasible and logically solvable.
[0009] Optionally, verify whether the actual physical state changes generated by the control strategy are consistent with the expected physical state changes defined by the candidate action model, including: Obtain the set of actual environmental physical states after the execution of the control strategy; Determine whether the actual set of physical states contains the union of the set of preconditions defined by the candidate action model and the set of added physical states, and exclude the set of deleted physical states defined by the candidate action model. If the judgment result is yes, then the verification result is determined to be consistent, and the candidate action model and control strategy are bound as an action node in the behavior tree system; If the judgment result is negative, then the verification result is determined to be inconsistent.
[0010] Optionally, a candidate action model set is generated based on the task set using the first pre-trained model, and the behavior tree planning algorithm is used to verify whether the candidate action model set is sufficient to plan a solution to complete all the tasks to be performed by the robot in the task set, generating a planning context, including: Initialize the set of unexplored action models; For each robot task to be performed in the task set, a behavior tree planning algorithm is run using the currently validated and consistent set of action models. If there is a robot task that the behavior tree planning algorithm cannot generate a solution for, then the robot task is marked as a planning failure task, and the missing conditions or breakpoints that caused the planning failure are recorded as part of the planning context. The failed planning task, the current set of unexplored action models, and the planning context are input into the first pre-trained model. The first pre-trained model infers and generates a new supplementary action model, and then adds the new supplementary action model to the set of unexplored action models.
[0011] Optionally, using a second pre-trained model, based on visual observation data of the environment and planning context, control policies corresponding to candidate action models are sampled from the control policy space, including: For each candidate action model to be validated, initialize its corresponding set of explored policies to empty; Determine whether the number of elements in the explored strategy set is less than the preset maximum number of samples; If so, the current candidate action model, the set of explored policies, and the current execution context are input into the second pre-trained model. The second pre-trained model then infers a new combination of parameters or an algorithm interface from the control policy space based on the previous failure reasons recorded in the execution context, which serves as the control policy for the current sampling.
[0012] Optionally, the behavior tree system includes a set of condition nodes and a set of action nodes; Each action node in the action node set is represented as a tuple, which consists of a validated action model and a control policy consistent with the action model. The action model includes a set of preconditions, a set of added physical states, and a set of deleted physical states.
[0013] Optionally, the control strategy space includes at least one or more of the following types of robot control strategies: An end-to-end neural network strategy is used to directly map visual input to robot joint movements; The hierarchical control strategy includes upper-level mission planning for generating intermediate waypoints and lower-level motion control for performing trajectory tracking. Rule-based parameterized action primitives, including predefined motion scripts and application programming interfaces for object-specific operations.
[0014] In a second aspect, the present invention provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described robot control method.
[0015] Thirdly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described robot control method.
[0016] The present invention has at least the following beneficial effects: By automatically generating candidate action models and sampling control strategies using pre-trained large models, this invention eliminates the reliance on manually defined action models by domain experts, significantly reducing the cost of deploying robot systems. It ensures that high-level logic covers all tasks through planning context, and that the physical effects of the abstract model match the underlying control strategy through execution context and physical simulation / physical verification. This effectively solves the common "illusion" and execution failure problems of large models in robot control. The method is independent of specific tasks or robot platforms and can be widely applied in industrial, logistics, and service scenarios. Attached Figure Description
[0017] The accompanying drawings are provided to further understand the technical solutions of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the technical solutions of the present invention, and do not constitute a limitation on the technical solutions of the present invention.
[0018] Figure 1 This is a flowchart of a robot control method in one embodiment of this application; Figure 2 This is a schematic diagram comparing the evaluation results of two high-level model proposals using two LLMs in another embodiment of this application. Two main metrics are used to evaluate the completeness of the high-level model: (1) Average Planning Success Rate (ASR): the average planning success rate of all individual tasks in a given task set. (2) Complete Planning Success Rate (CSR): the success rate of all tasks in the task set being successfully planned simultaneously. Figure 3 This is a schematic diagram comparing the evaluation results of low-level policy sampling of five typical action models using VLM in another embodiment of this application. Figure 4 This diagram illustrates the success rate of correcting five specific examples of motion model errors in another embodiment of this application. Figure 5 This is a schematic diagram of the structure of a terminal device in one embodiment of this application. Detailed Implementation
[0019] The technical solution of the present invention will now be described in detail and completely with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0020] In the description of this invention, it should be noted that the terms "upper", "lower", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.
[0021] Example 1 This invention provides a robot control method that aims to automatically construct a complete and consistent behavior tree (BT) system for a given task set. For example... Figure 1 As shown, the method specifically includes the following steps: Step 11: Obtain a task set consisting of at least one task to be performed by the robot.
[0022] In this embodiment of the invention, each task to be performed by the robot includes an initial physical state and a target physical state. The task to be performed by the robot is at least an industrial assembly task or a logistics sorting task.
[0023] In one feasible implementation, the task set can be represented as Each robot has a task to perform. ,in, Indicates the first The initial physical state (e.g., object position and robot posture) of a robot to perform a task. Indicates the first The target physical state of the robot to perform a task (e.g., an object being placed at a target location). Task types may include pick-and-place, switch operation, open and place, rearrange, etc.
[0024] For example, in logistics sorting tasks, the task can be represented as: Initial state: Goods are in area A, robot is empty-handed. Target state: Goods are inside box B. Initial state: Box B is closed. Target state: Box B is open.
[0025] It should be noted that, in actual implementation, the task set Provided by the user or upstream system, and the task set It should cover all the skills the robot needs to master. For example, a complete task set includes combinations of tasks such as picking and placing, switching, opening and putting in, etc.
[0026] Step 12: Use the first pre-trained model to generate a set of candidate action models based on the task set, and use the behavior tree planning algorithm to verify whether the set of candidate action models is sufficient to plan a solution to complete all the tasks to be performed by the robot in the task set, and generate a planning context.
[0027] In this embodiment of the invention, the candidate action model set includes at least one candidate action model. Each candidate action model represents a repeatable robot atomic operation, including the preconditions required to perform the robot atomic operation, the physical state expected to be added to the environment after execution, and the physical state expected to be deleted from the environment after execution.
[0028] Preferably, the first pre-trained model is a large language model trained on a large scale of text and code, used for symbolic reasoning and logical planning. For example, the first pre-trained model is GPT-4 or DeepSeek.
[0029] In one feasible implementation, step 12 specifically includes steps 12.1 to 12.4.
[0030] Step 12.1: Initialize the set of unexplored action models.
[0031] Specifically, define the set of valid condition nodes. (Provided by domain common sense or users, such as the positions and states of all possible objects). Initialize the set of unexplored action models. At the beginning, It can be empty, or it can contain some data determined by the LLM based on the task set. The initial candidate model is generated directly.
[0032] In one feasible implementation, a prompt is sent to the LLM (such as GPT-4): "Given the following robot task descriptions: [Task 1 description], [Task 2 description]... Please list the basic atomic operations that may be needed to complete these tasks, and define the precondition (pre), add effect (add, added physical state), and delete effect (del, deleted physical state) for each operation. The format is: operation name(pre=[condition list], add=[condition list], del=[condition list])." The LLM's response will be parsed into an initial set of candidate action models.
[0033] Step 12.2: For each robot task to be performed in the task set, run the behavior tree planning algorithm using the currently validated and consistent set of action models.
[0034] Specifically, let the set of currently validated and consistent action models be... The corresponding action model set is For each robot's task to be performed. Using classic BT programming algorithms (such as the BT Expansion algorithm), to The action model in the model is used as the planning domain to attempt to define the task to be performed by the robot. Generate a solution BT.
[0035] Step 12.3: If there is a robot task that the behavior tree planning algorithm cannot generate a solution for, then mark the robot task as a planning failure task and record the missing conditions or breakpoints that caused the planning failure as part of the planning context.
[0036] If the robot has a task to perform If the planning is successful, it is recorded as successful; if it fails, the planning algorithm usually reveals the reason for the failure. For example, the algorithm may try to extend a certain condition but cannot find an action that can produce that condition. This "lack of physical state conditions to achieve the goal" and "the robot's pending tasks that caused the planning failure" are also factors. The robot itself, along with the sequence of actions that have been attempted to expand but failed to achieve the goal, together constitute the task to be performed by the robot. Planning context .
[0037] Step 12.4: Input the failed planning task, the current set of unexplored action models, and the planning context into the first pre-trained model. The first pre-trained model infers and generates a new supplementary action model, and adds the new supplementary action model to the set of unexplored action models.
[0038] Specifically, for each robot task that failed to be planned... ,Will The current set of unexplored models As input, LLM is invoked again. In one possible implementation, the LLM prompt is: "To complete the task 'Put Goods A into the box', the current action model library cannot generate a plan. We need to satisfy the condition Holding(Goods) in the planning, but no action can generate it. The existing action models are: [list]" [Referring to the model in the text]. Based on the task requirements and existing models, please deduce and propose possible new action models to address this missing link.
[0039] The new motion model output by LLM has been added. and from Move to the explored model set .
[0040] It should be noted that, in practice, steps 12.2 to 12.4 are repeated until the BT planning algorithm can solve the task set. A solution has been found for each task that the robots in the set need to perform. At this point, the set... The action models contained therein satisfy completeness.
[0041] It's worth noting that this step transforms the exponentially large action model space search problem into an LLM-driven, heuristic search based on planning failure feedback, significantly improving efficiency. The generated planning context... It provides semantic information for subsequent steps on "why a certain action model is needed".
[0042] Step 13: Using the second pre-trained model, based on the visual observation data of the environment and the planning context, sample the control policy corresponding to the candidate action model from the control policy space.
[0043] This step aims to provide a framework for each candidate action model proposed in step 12. To find a control strategy at the physical level that can achieve this. .
[0044] In this embodiment of the invention, the control strategy is the control parameters that drive the movement of robot joints or end effectors. Visual observation data includes objects and the spatial relationships between them. The control strategy space includes at least one or more of the following types of robot control strategies: An end-to-end neural network strategy is used to directly map visual input to robot joint movements; The hierarchical control strategy includes upper-level mission planning for generating intermediate waypoints and lower-level motion control for performing trajectory tracking. Rule-based parameterized action primitives, including predefined motion scripts and application programming interfaces for object-specific operations.
[0045] Preferably, the second pre-trained model is a visual language model jointly trained with visual language multimodal data, used to understand the physical scene and generate executable instructions or parameters related to the physical scene. For example, the second pre-trained model is GPT-4o.
[0046] In one feasible implementation, the process of sampling control strategies corresponding to candidate action models from the control strategy space using a second pre-trained model based on visual observation data of the environment and planning context includes steps 13.1 to 13.3.
[0047] Step 13.1: For each candidate action model to be verified, initialize its corresponding set of explored strategies to be empty.
[0048] Specifically, initialize an empty set of explored strategies. Used for the current action model .
[0049] Step 13.2: Determine whether the number of elements in the explored strategy set is less than the preset maximum number of samplings.
[0050] Step 13.3 If yes, the current candidate action model, the explored policy set, and the current execution context are input into the second pre-trained model. The second pre-trained model infers a new parameter combination or algorithm interface from the control policy space based on the previous failure reasons recorded in the execution context, which is used as the currently sampled control policy.
[0051] Specifically, it is associated with the current action model. Relevant planning context (For example, It was introduced when trying to solve the "placing goods into a box" task. Then, it uses a visual camera to acquire the current RGB-D image, object segmentation mask, pose estimation, and other information, and generates prompts for a second pre-trained model based on this information. For example: [RGB image of the current scene] "As shown in the figure, there is a piece of goods A on the table in the environment. The robot needs to perform the atomic operation 'grab the goods' (prerequisite: hands are empty and close to the goods; target effect: holding the goods). Currently tried strategies include: [list]" [Strategies and reasons for failure in the process]. Based on the current visual observation, please generate or select a control strategy (such as an API interface, or output a piece of position-based impedance control code) to complete this grabbing action.
[0052] The output of the second pre-trained model is interpreted as a specific control policy. For example, control strategies It could be: a) A piece of code that calls a predefined motion primitive API; b) A set of parameters used for motion generation (such as the 3D coordinates of the gripping point and the posture of the robotic arm's end effector); c) Configuration parameters for a neural network strategy.
[0053] Step 14: Execute the control strategy in a simulated or physical environment, generate an execution context based on the environmental state feedback after execution, and verify whether the actual physical state changes generated by the control strategy are consistent with the expected physical state changes defined by the candidate action model.
[0054] It should be noted that step 14 is used to verify the sampled strategy. Is it really possible to implement motion models? The effect of the statement.
[0055] Specifically, the process of executing control strategies in a simulated or physical environment and generating an execution context based on the feedback from the environmental state after execution includes: setting an initial state in the simulation environment or physical robot. To make it accurately meet the motion model All prerequisites For example, for Ensure the robot arm is empty and that the end effector has moved to the vicinity of the goods. Run the control strategy within the defined scenario. Record the entire execution process to obtain the execution context. Execution context It should include at least one or more of the following information: images or point cloud data of the physical state of the environment after execution, acquired through robot sensors or a simulated environment engine; error codes or warning messages triggered during the execution of the control strategy; or a description of the difference between the actual physical state change and the expected physical state change. For example: "After executing the grasping strategy, the expected state includes IsHolding(Goods), but in the actual state, this condition is false. It was actually observed that the robotic arm collided with the goods as it approached them, and the goods were knocked off." (Execution context) It will include information such as error logs and comparisons of state snapshots before and after execution.
[0056] In one feasible implementation, the process of verifying whether the actual physical state change generated by the control strategy is consistent with the expected physical state change defined by the candidate action model includes steps 14.1 to 14.4.
[0057] Step 14.1: Obtain the set of actual environmental physical states after the execution of the control strategy.
[0058] Operation control strategy Then, the actual physical state of the environment is obtained through environmental interfaces or sensors. .
[0059] Step 14.2: Determine whether the actual set of physical states contains the union of the set of preconditions defined by the candidate action model and the set of added physical states, and exclude the set of deleted physical states defined by the candidate action model.
[0060] Specifically, based on the motion model Calculate the expected physical state .
[0061] Step 14.3: If the judgment result is yes, then the verification result is determined to be consistent, and the candidate action model and control strategy are bound as an action node in the behavior tree system.
[0062] Specifically, check Whether it holds true, that is: whether the actual physical state contains at least all the expected added effects and does not contain all the expected deleted effects.
[0063] Step 14.4: If the judgment result is negative, then the verification result is determined to be inconsistent.
[0064] if If true, then the verification results are considered consistent, and... As a verified "consistent action pair", it is added to the final set of action nodes. .mark The grounding has been successfully completed.
[0065] like If the result is not valid, the verification result is determined to be inconsistent, and step 15 is executed.
[0066] Step 15: When the verification results are inconsistent, the candidate action model is modified by using the second pre-trained model and combining the planning context and execution context until a behavior tree system that meets the requirements of completeness and consistency is generated.
[0067] In this embodiment of the invention, the behavior tree system includes a set of condition nodes and a set of action nodes; each action node in the action node set is represented as a tuple, which consists of a validated action model and a control strategy consistent with the action model; the action model includes a set of preconditions, a set of added physical states, and a set of deleted physical states.
[0068] Specifically, this includes steps 15.1 to 15.3.
[0069] Step 15.1: When a control strategy that meets the consistency requirements cannot be sampled for a candidate action model within a preset number of times, the candidate action model is determined to be infeasible in the current context.
[0070] Specifically, when targeting a certain action model After multiple samplings (e.g., 3 times) have all failed to validate the strategy, the decision is made to directly determine the outcome in the current context. Finding a strategy may be very difficult or impossible.
[0071] Step 15.2: Input the planning context and execution context as prompts into the second pre-trained model.
[0072] Unlike conventional methods, in determining this action model If it is not feasible in the current context, the present invention will not simply abandon the action model. Instead, it initiates a detailed process to integrate the following information: Problematic motion model ,and Relevant planning context (For illustration) Why it is needed), the set of execution contexts for all attempted strategies. (Used to explain why it failed) and the set of unexplored action models (Used to provide directions for modification), and input the integrated information into the second pre-trained model, and generate prompts to request action models. Make modifications to make it easier to implement at the physical level.
[0073] In one possible implementation, the prompt word is: "Action Model" This is necessary when attempting to complete the 'placing goods into a box' task. However, we tried three grasping strategies, all of which failed due to collisions upon approach. Please analyze the reasons for the failures and suggest improvements to this action model. For example, are the preconditions too lenient? Is it necessary to add... or What are the prerequisites? Or does the effect definition need to be adjusted? Indicate whether the robot's hands are empty. Indicate whether you are near the goods. This indicates the action of grabbing goods. Indicates whether the goods are stable. This indicates whether the grasping posture is appropriate.
[0074] Step 15.3: Use the second pre-trained model to infer the corrected candidate action model.
[0075] In this embodiment of the invention, the modified action model includes modifications to the preconditions, added physical states, or deleted physical states of the original candidate action model, so that the modified candidate action model is physically feasible and logically solvable.
[0076] Specifically, VLM outputs one or more modified motion models. For example, modify the premise to The revised model It will be sent back to step 12 to re-participate in the planning verification. If the revised action model... If it remains intact, proceed to steps 13-14 to attempt grounding. This process achieves an intelligent feedback loop from execution failure to high-level model correction.
[0077] Step 16: Parse the behavior tree system, convert the behavior tree nodes into robot drive instructions, and drive the robot to perform operations according to the drive instructions to complete the task set.
[0078] Specifically, when the loop iteration (steps 12 to 15) ends, it means that a set of action models has been found. And for each of these action models They all found a consistent control strategy .These The action node set A constitutes the behavior tree system. All conditions appearing in these action models constitute the condition node set C. Thus, a complete and consistent BT system is formed. Construction complete.
[0079] For a specific task instance submitted by a user (its initial state) and target Included in the task set Within the defined scope), using the standard BT planner, based on Quickly generate a specific executable BitTorrent (BT). Parse this BT into a series of underlying control strategies. The call sequence is sent to the robot controller for execution until the task is completed.
[0080] Example 2 In this embodiment of the invention, the effectiveness of the robot control method provided by the present invention was verified in a set of robot operation experiments based on the OmniGibson and isaacsim simulation platforms. The task set includes 7 tasks in 3 categories, involving the Franka single-arm robot, the Franka dual-arm robot, and the Fetch mobile robotic arm.
[0081] like Figure 2 As shown, the robot control method provided by this invention exhibits significant performance improvements across the entire task set after integrating the planning context. Specifically, the average planning success rate (ASR) of GPT-4o jumps from 66.2% without using the context to 89.4%, while the ASR of GPT-3.5-turbo also improves from 67.1% to 74.8%. Regarding planning integrity (CSR), GPT-4o's integrity significantly improves from 50% to 88.3% (even jumping to over 90%), and GPT-3.5-turbo also improves from 42.9% to 64.3%. This performance improvement is particularly pronounced in complex dual-arm and mobile manipulation tasks. This demonstrates the effectiveness of step 12 in the robot control method provided by this invention. Figure 2 In this context, FC represents the average number of feedback cycles.
[0082] like Figure 3 As shown, using VLM (GPT-4o) in conjunction with execution context for policy sampling, the average success rate (ESR) reaches 62% for five typical actions: Pick, Place, Open, Close, and Toggle. This is superior to sampling without context (44%) and methods that consistently use a single policy type (such as end-to-end or rule-based). This demonstrates the advantages of context-aware sampling in steps 13 and 14 of the robot control method provided by this invention.
[0083] like Figure 4 As shown, for inconsistent motion models, i.e., discrepancies or errors between high-level prediction results and low-level execution results, the correction method provided by this invention can successfully correct these inconsistencies, increasing the correction rate from 44% to 76%. This demonstrates the effectiveness of step 15 in the robot control method provided by this invention.
[0084] As can be seen, the robot control method provided by this invention automatically generates candidate action models and sampled control strategies by utilizing a pre-trained large model. This invention eliminates the reliance on manual definition of action models by experts in specific domains, significantly reducing the cost of deploying robot systems. By planning the context, it ensures that the high-level logic can cover all tasks, and by using the execution context and physical simulation / physical verification, it ensures that the physical effects of the abstract model and the underlying control strategy match. This effectively solves the common problems of "illusion" and execution failure in robot control caused by large models.
[0085] like Figure 5 As shown, embodiments of the present invention provide a terminal device, such as... Figure 5 As shown, the terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 5 The diagram shows only one processor, a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 executes the computer program D102 to implement the steps in any of the above method embodiments.
[0086] Specifically, when the processor D100 executes the computer program D102, it acquires a task set consisting of at least one robot task to be performed; it generates a candidate action model set based on the task set using a first pre-trained model, and verifies whether the candidate action model set is sufficient to plan a solution to complete all robot tasks in the task set using a behavior tree planning algorithm, generating a planning context; it samples control strategies corresponding to the candidate action models from the control strategy space based on the visual observation data of the environment and the planning context using a second pre-trained model; it executes the control strategies in a simulated or physical environment, generates an execution context based on the environmental state feedback after execution, and verifies whether the actual physical state changes generated by the control strategies are consistent with the expected physical state changes defined by the candidate action models; when the verification results are inconsistent, it uses the second pre-trained model, combined with the planning context and the execution context, to correct the candidate action models until a behavior tree system that meets the requirements of completeness and consistency is generated; it parses the behavior tree system, converts the behavior tree nodes into robot drive instructions, and drives the robot to perform operations according to the drive instructions to complete the task set. By automatically generating candidate action models and sampling control strategies using pre-trained large models, this invention eliminates the reliance on manually defined action models by domain experts, significantly reducing the cost of deploying robot systems. It ensures that high-level logic covers all tasks through planning context, and that the physical effects of the abstract model and the underlying control strategy match through execution context and physical simulation / physical verification. This effectively solves the common "illusion" and execution failure problems of large models in robot control. The method is independent of specific tasks or robot platforms and can be widely applied in industrial, logistics, and service scenarios.
[0087] The processor D100 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.
[0088] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may be an external storage device of the terminal device D10, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the terminal device D10. Furthermore, the memory D101 may include both internal and external storage units of the terminal device D10. The memory D101 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the computer program. The memory D101 can also be used to temporarily store data that has been output or will be output.
[0089] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps described in the various method embodiments above.
[0090] This application provides a computer program product that, when run on a terminal device, enables the terminal device to implement the steps described in the various method embodiments above.
[0091] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.
[0092] One or more embodiments in this application are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments in this application should be included within the protection scope of this application.
Claims
1. A robot control method, characterized in that, include: Obtain a task set consisting of at least one robot task to be performed; each robot task to be performed includes an initial physical state and a target physical state; the robot task to be performed is an industrial assembly task or a logistics sorting task. A candidate action model set is generated based on the task set using a first pre-trained model, and the behavior tree planning algorithm is used to verify whether the candidate action model set is sufficient to plan a solution to complete all the robot tasks to be performed in the task set, thereby generating a planning context. The candidate action model set includes at least one candidate action model, and each candidate action model represents a repeatable robot atomic operation, including the preconditions required to perform the robot atomic operation, the physical state expected to be added to the environment after execution, and the physical state expected to be deleted from the environment after execution. Using a second pre-trained model, based on the visual observation data of the environment and the planning context, control policies corresponding to the candidate action models are sampled from the control policy space; The control strategy refers to the control parameters that drive the movement of robot joints or end effectors; the visual observation data includes objects and the spatial relationships between them. The control strategy is executed in a simulated or physical environment. An execution context is generated based on the environmental state feedback after execution, and it is verified whether the actual physical state change generated by the control strategy is consistent with the expected physical state change defined by the candidate action model. When the verification results are inconsistent, the candidate action model is modified by using the second pre-trained model, combined with the planning context and the execution context, until a behavior tree system that meets the requirements of completeness and consistency is generated. The behavior tree system is parsed, and the behavior tree nodes are converted into driving instructions for the robot. The robot is then driven to perform operations according to the driving instructions to complete the task set.
2. The robot control method according to claim 1, characterized in that, The first pre-trained model is a large language model trained on a large scale of text and code, used for symbolic reasoning and logical planning; The second pre-trained model is a visual language model jointly trained with visual language multimodal data, used to understand physical scenes and generate executable instructions or parameters related to the physical scenes.
3. The robot control method according to claim 2, characterized in that, When the verification results are inconsistent, the candidate action model is corrected using the second pre-trained model, combined with the planning context and the execution context, including: If a control strategy that meets the consistency requirement cannot be sampled for a candidate action model within a preset number of times, the candidate action model is determined to be infeasible in the current context. The planning context and the execution context are input as prompts into the second pre-trained model; wherein, the planning context includes at least one or more of the following information: task information that failed due to lack of effective actions during behavior tree planning, action sequences that have been attempted to expand but failed to achieve the goal, or physical state conditions that are still lacking to achieve the goal; the execution context includes at least one or more of the following information: images or point cloud data of the physical state of the environment after execution obtained by robot sensors or simulation environment engine, error codes or warning information triggered during the execution of the control strategy, or a description of the difference between the actual physical state change and the expected physical state change; The modified candidate action model is inferred using the second pre-trained model; the modified action model includes modifications to the preconditions, added physical states or deleted physical states of the original candidate action model, so that the modified candidate action model is physically feasible and logically solvable.
4. The robot control method according to claim 1, characterized in that, The verification of whether the actual physical state changes generated by the control strategy are consistent with the expected physical state changes defined by the candidate action model includes: Obtain the set of actual environmental physical states after the control strategy is executed; Determine whether the actual environmental physical state set contains the union of the precondition set defined by the candidate action model and the added physical state set, and exclude the deleted physical state set defined by the candidate action model. If the judgment result is yes, then the verification result is determined to be consistent, and the candidate action model is bound to the control strategy as an action node in the behavior tree system; If the judgment result is negative, then the verification result is determined to be inconsistent.
5. The robot control method according to claim 1, characterized in that, The process involves using a first pre-trained model to generate a set of candidate action models based on the task set, and verifying, through a behavior tree planning algorithm, whether the set of candidate action models is sufficient to plan a solution to complete all the tasks to be performed by the robots in the task set, generating a planning context, including: Initialize the set of unexplored action models; For each robot task to be performed in the task set, a behavior tree planning algorithm is run using the currently validated and consistent set of action models. If there is a robot task to be executed that the behavior tree planning algorithm cannot generate a solution for, then the robot task to be executed is marked as a planning failure task, and the missing conditions or breakpoints that caused the planning failure are recorded as part of the planning context. The failed planning task, the current set of unexplored action models, and the planning context are input into the first pre-trained model. The first pre-trained model infers and generates a new supplementary action model, and the new supplementary action model is added to the set of unexplored action models.
6. The robot control method according to claim 1, characterized in that, The step of using a second pre-trained model to sample control policies corresponding to the candidate action models from the control policy space based on the visual observation data of the environment and the planning context includes: For each candidate action model to be validated, initialize its corresponding set of explored policies to empty; Determine whether the number of elements in the explored strategy set is less than the preset maximum number of samplings; If so, the current candidate action model, the explored policy set, and the current execution context are input into the second pre-trained model. The second pre-trained model then infers a new parameter combination or algorithm interface from the control policy space based on the previous failure reasons recorded in the execution context, which serves as the currently sampled control policy.
7. The robot control method according to claim 1, characterized in that, The behavior tree system includes a set of condition nodes and a set of action nodes; Each action node in the set of action nodes is represented as a tuple, which consists of a validated action model and a control strategy consistent with the action model. The action model includes a set of preconditions, a set of added physical states, and a set of deleted physical states.
8. The robot control method according to claim 1, characterized in that, The control strategy space includes at least one or more of the following types of robot control strategies: An end-to-end neural network strategy is used to directly map visual input to robot joint movements; The hierarchical control strategy includes upper-level mission planning for generating intermediate waypoints and lower-level motion control for performing trajectory tracking. Rule-based parameterized action primitives, including predefined motion scripts and application programming interfaces for object-specific operations.
9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the robot control method as described in any one of claims 1 to 7.
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 robot control method as described in any one of claims 1 to 7.