Task planning method and system for intelligent robot with body

Abstract

The invention belongs to the field of intelligent robot control, and provides an intelligent robot task planning method and system, and the method comprises the steps: obtaining real-time environment data and a target task; encoding the real-time environment data and the target task by using a pre-trained visual language model to obtain unified representation; generating a sub-task sequence based on unified representation by using an actor-commentator algorithm, evaluating the rationality of a planning strategy based on the sub-task sequence by comprehensively considering high efficiency, constraint conformity and performability, if a score is lower than a threshold value, re-generating the sub-task sequence, otherwise, generating and executing a corresponding action, and in the execution process, executing the sub-task sequence. And obtaining a subtask sequence and a target task completion condition, continuously optimizing a planning strategy, and realizing dynamic planning. The method has the capabilities of dynamic task decomposition and online decision optimization, so that the problem of adaptability of a static planning normal form in a dynamic environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of embodied intelligent robot control, specifically relating to an embodied intelligent robot task planning method and system. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Embodied intelligence aims to endow intelligent agents with the ability to autonomously perceive, make decisions, and execute tasks in a physical environment. Task planning is one of its core functional modules, responsible for translating high-level natural language instructions into a sequence of actionable steps. Currently, the task planning module of embodied intelligence generally relies on the Vision-Language-Motion (VLA) model as its technological foundation. The VLA model, by integrating visual perception, language understanding, and robot control into a unified architecture, paves the way for the realization of general-purpose robots capable of following complex instructions. In recent years, VLA models have mainly developed along two technical paths: one is an end-to-end model based on pre-training on large-scale Internet data, which directly realizes the mapping from observation to action through sequence modeling; the other is a hierarchical architecture model, such as the π series released by Physical Intelligence, which improves the interpretability and control accuracy of the system through explicit high-level task planning and low-level action execution. Within this development trend, the π0.5 model represents a significant breakthrough in hierarchical VLA architecture. It demonstrates strong generalization ability for performing long-cycle tasks in unknown family environments by combining explicit high-level semantic planning (e.g., breaking down "stacking plates" into sub-task sequences such as "finding plates" and "picking up plates") with low-level continuous action generation based on flow matching techniques. This hierarchical design enables the model to handle complex tasks requiring multi-step reasoning while maintaining the fluency and accuracy of action execution.

[0004] However, existing VLA models still have noteworthy limitations in their task planning capabilities in complex dynamic environments. From an architectural perspective, although the planning process lacks effective value assessment and replanning capabilities, specifically, after the system generates a sequence of sub-tasks based on the current environmental state, it lacks a systematic evaluation mechanism for the quality of the planned solutions, making it impossible to quantitatively judge the execution effect and long-term value of the current sub-task sequence.

[0005] From the perspective of learning mechanisms, existing VLA model training paradigms also face challenges when dealing with long-term tasks. Imitation learning-based methods heavily rely on high-quality expert demonstration data, and their performance is limited by the coverage of the training set, making it difficult to generalize to all possible environmental changes. In the reinforcement learning paradigm, the core challenge lies in designing a reward function that can effectively balance long-term goals and short-term behavior. Existing methods often face a dilemma: relying on a single sparse reward can clearly define the final goal, but it leads to a serious credit allocation problem, making the learning process slow and inefficient; on the other hand, simply using a hand-designed dense reward makes it extremely difficult to fully characterize the true intent of complex tasks, which can easily cause the model behavior to deviate from the essential goal, get stuck in local optima, or even lead to the phenomenon of "reward hacking". Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes an embodied intelligent robot task planning method and system. Based on the π0.5 architecture, and building upon its superior multimodal perception and low-level skill execution capabilities, this invention innovatively introduces a high-level decision-making core with forward-looking evaluation and reflective replanning capabilities. This endows the VLA model with the ability of dynamic task decomposition and online decision optimization, thereby solving the adaptability problem of static planning paradigms in dynamic environments.

[0007] According to some embodiments, the present invention adopts the following technical solution: A method for task planning of an embodied intelligent robot includes the following steps: Acquire real-time environmental data and target tasks; A pre-trained visual language model is used to encode real-time environmental data and target tasks to obtain a unified representation; The actor-critic algorithm is used to generate sub-task sequences based on a unified representation. The rationality of the planning strategy based on the sub-task sequences is evaluated by comprehensively considering efficiency, constraint compliance, and executability. If the score is lower than the threshold, the sub-task sequence is regenerated; otherwise, the corresponding action is generated and executed. During the execution, the sub-task sequence and the completion status of the target task are obtained, and the planning strategy is continuously optimized to achieve dynamic programming.

[0008] As an alternative implementation, the process of generating a sequence of subtasks based on a unified representation using the actor-critic algorithm includes: implementing the actor policy network function using the Gemma large language model to parse the input target task into a structured, executable sequence of subtasks.

[0009] As a further defined implementation, the immediate reward r of the subtask is directly predicted by the Gemma large language model. After the generated subtask is input into the Gemma large language model, the model outputs the reward value based on its understanding of the task semantics and execution scenario.

[0010] As an alternative implementation method, the process of evaluating the rationality of a planning strategy based on a sub-task sequence, taking into account efficiency, constraint compliance, and executability, includes using a critic module to assess the rationality value of the decomposition strategy based on whether the decomposition scheme can efficiently advance the overall task and whether it meets constraints, thus obtaining a value score: ; ; ; Among them, V progress It is the value of progress, V constraint V represents the constrained value, and S represents the combined value. t Indicates the state at time t, a t This represents the subtask at time t in the actor policy network decomposition.

[0011] As an alternative implementation, the overall policy gradient update formula for the actor-critic algorithm is as follows: ; ; Among them, a t Let s be the action at time t. t Let V be the state at time t, r be the immediate reward predicted by the Gemma large language model from the subtask input, and V be the value of the state at time t. φ It is a state-value function. The entire training process of the framework forms a closed loop with the Gemma large language model as the core: the target task is input into the Gemma large language model to generate a sequence of subtasks; the subtasks are input into the Gemma large language model to predict the immediate reward r; the commentator module outputs state values ​​Vφ(st) and Vφ(st+1) based on the environment state and the subtask execution results; combined with logπθ(a t |s t Instant reward r and value error V φ (s) t +1)-V φ (s) t The parameters of the actor-critic algorithm are updated using the policy gradient formula.

[0012] As an alternative implementation, the process of regenerating the sub-task sequence if the score is lower than the threshold includes: when the critic module's score for the current strategy is lower than its preset threshold, the value scheduler triggers the task replanning process, that is, the original target task is re-input into the Gemma large language model, and the Gemma large language model performs the decomposition operation of complex task into sub-task sequence again. Then, based on the newly generated sub-tasks, the actor-critic strategy generation, value evaluation and parameter update process is restarted until the critic module's score for the strategy meets the threshold requirement.

[0013] As an alternative implementation method, a non-fusion hierarchical density reward mechanism is adopted during the continuous optimization of the planning strategy. This mechanism does not merge sparse rewards and dense rewards into a single signal, but retains the independence of the two types of rewards. This allows the two to play the roles of global goal anchoring and local action guidance respectively in the reinforcement learning process, thus enabling them to cooperate and divide their tasks.

[0014] As a further defined implementation, the non-fusion hierarchical density reward mechanism includes sparse rewards and dense rewards. The sparse reward serves as a rigid constraint on the global task objective, and outputs a signal only when the task is completely terminated. This signal is used to determine whether the task was ultimately successful and to provide a final verification for the robot's task planning with a goal-oriented approach. Dense rewards serve as an optimization guide for subtasks and single-step actions, and are dynamically triggered during task execution to provide process feedback to the robot.

[0015] As a further defined implementation method, the sparse reward is:

[0016] Dense rewards are:

[0017] Where C(t) represents the task completion degree, and G(t) represents the capture quality. , The corresponding weights.

[0018] An embodied intelligent robot task planning system, comprising: The data acquisition module is configured to acquire real-time environmental data and target tasks. The encoding module is configured to encode real-time environmental data and target tasks using a pre-trained visual language model to obtain a unified representation; The dynamic task planning module is configured to generate a sequence of subtasks based on a unified representation using the actor-critic algorithm. It comprehensively considers efficiency, constraint compliance, and executability to evaluate the rationality of the planning strategy based on the subtask sequence. If the score is lower than the threshold, the subtask sequence is regenerated; otherwise, the corresponding action is generated and executed. During the execution, the subtask sequence and the completion status of the target task are obtained, and the planning strategy is continuously optimized to achieve dynamic planning.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention utilizes the Actor-Critic framework to achieve dynamic programming and optimization. On the Actor policy network, it breaks through the limitations of traditional neural networks that directly output the probability distribution of actions. It adopts the Gemma large language model to realize the Actor function, which parses the input high-level task (such as a complex task described by natural language) into a structured and executable sequence of subtasks. It can identify the implicit constraints of the task, resolve the dependencies between subtasks, and generate subtask units with appropriate granularity and clear semantics. This upgrades the action space of reinforcement learning from low-level operations to semantic subtasks, greatly reducing the complexity of policy learning.

[0020] Regarding the design of the reward function, this invention abandons the traditional manual rule design mode and directly predicts the immediate reward r of the subtask by Gemma. After the generated subtask is input into Gemma, the model outputs the reward value based on its understanding of the task semantics and execution scenario. This mechanism makes the reward prediction semantically reasonable, can dynamically adapt to tasks in different domains, and aligns the reward granularity with the semantic granularity of the subtask, making the feedback signal more in line with human perception of task progress.

[0021] When the Critic module's score for the current strategy falls below its preset threshold Vthreshold, this invention will trigger the task replanning process, restarting the Actor-Critic strategy generation, value assessment, and parameter update process until the Critic's score for the strategy meets the threshold requirement. This forms an enhanced closed loop of strategy quality inspection, replanning, and iterative optimization, ensuring the rationality and effectiveness of the task decomposition strategy.

[0022] This invention proposes a non-fusion hierarchical density reward mechanism. Through its core design of functional separation and independent feedback, it precisely addresses the core contradiction of traditional single reward mechanisms in long-term tasks. This mechanism clearly divides sparse and dense rewards into two independent modules: global goal constraint and local process guidance. This avoids the training inefficiency caused by sparse rewards due to feedback lag, and also avoids the risk of dense rewards deviating from the global goal due to excessive focus on local areas. Furthermore, it eliminates the need for weighted fusion to balance the weights of the two, completely eradicating the problem of weakened reward effects or distorted feedback caused by improper weight settings. From triggering timing to computational logic, both types of rewards remain independent and perform their respective functions, laying a crucial feedback foundation for embodied intelligent robots to solve the credit allocation problem and improve the robustness of task planning in long-term tasks.

[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0025] Figure 1 This is a schematic diagram of an adaptive task planning principle in one embodiment; Figure 2 This is a schematic diagram of an embodiment of the Actor-Critic framework. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, 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 invention pertains.

[0028] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0029] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0030] Example 1 A task planning method for embodied intelligent robots is proposed, implemented based on the TP-ACDR model. This embodiment of the TP-ACDR model is a novel task planning model based on the Actor-Critic framework and a hierarchical reward mechanism. This embodiment is based on a π0.5 architecture, and innovatively introduces a high-level decision-making core with forward-looking evaluation and reflective replanning capabilities, building upon its superior multimodal perception and low-level skill execution capabilities. The core idea of ​​TP-ACDR is to endow the VLA model with dynamic task decomposition and online decision optimization capabilities through the Actor-Critic architecture in reinforcement learning, thereby solving the adaptability problem of static planning paradigms in dynamic environments.

[0031] This embodiment proposes the TP-ACDR framework, which constructs a complete closed-loop decision-making circuit of "dynamic task decomposition-value assessment-execution" for hierarchical VLA models. By deeply integrating reinforcement learning with the traditional hierarchical VLA architecture, this framework fundamentally improves the robustness of completing long-cycle tasks in dynamic environments.

[0032] A dynamic task decomposition mechanism based on Actor-Critic was designed. The Actor network acts as the policy network, generating flexible sub-task policies based on real-time environmental states; the Critic network acts as the value function, proactively evaluating the long-term benefits of the current sub-task policy. When the Critic's value falls below an adaptive threshold, the system immediately triggers replanning, effectively solving the planning deadlock problem common in traditional VLA models.

[0033] A density-based reward mechanism was introduced, and a signaling system was designed that combines dense rewards (for subtask completion quality and action efficiency) and sparse rewards (for global task success or failure). This mechanism provides immediate and fine-grained guidance for the model's local policy optimization, while ensuring that the learning process always aims to efficiently achieve the final goal, effectively solving the credit allocation problem in long-cycle tasks.

[0034] The following is a detailed description. The core objective of the adaptive task planning in this embodiment is to achieve efficient decision-making and accurate execution in complex dynamic scenarios, and to solve the problem of poor adaptability of planning methods under environmental uncertainty and task diversity.

[0035] The TP-ACDR framework in this embodiment uses a large language model as the core for semantic understanding and state awareness, integrates reinforcement learning Actor-Critic algorithm and density function modeling, and constructs a closed-loop task planning system of perception parsing, decision generation, and execution, such as... Figure 1 As shown.

[0036] The TP-ACDR model reinforcement learning fusion framework, based on multimodal understanding, with Actor-Critic policy as the core and density reward as the driving force, constructs an end-to-end closed loop from task input to dynamic decomposition and then to execution optimization. First, the acquired real-time environmental state and high-level target task are encoded into a unified representation using a pre-trained visual language model (in this embodiment, including SigLip+Gemma+ViT). Next, Gemma, acting as the "actor," generates a set of sub-tasks based on this representation. Then, the Critic evaluates the rationality of the strategy from the dimensions of efficiency, constraint compliance, and executability. If the score is below a threshold, Gemma is reprogrammed; if the score is met, the Action-expert module generates and executes the action. During execution, feedback is provided on the sub-task sequence and the overall task completion status, driving strategy optimization. Finally, based on the density reward signal and the advantage function calculated by the Critic score, the Actor and Critic parameters are iteratively updated. Simultaneously, the value scheduler coordinates the learning process, continuously optimizing the rationality and execution efficiency of the strategy. This allows the TP-ACDR model to ensure both local efficiency in sub-task execution and global achievability of the overall task objective under complex constraints, significantly improving the robustness and adaptability of dynamic task planning.

[0037] To address the challenges of unintelligent task decomposition and rigid reward function design in traditional Actor-Critic frameworks for complex semantic tasks, this embodiment proposes an innovative Actor-Critic architecture based on the large language model Gemma. This architecture is deeply integrated into the entire process of task decomposition, reward prediction, and policy optimization, enabling intelligent decomposition from high-level tasks to executable subtasks, as well as semantic prediction of immediate rewards at the subtask level. This significantly improves the framework's adaptability and intelligence level for complex semantic tasks.

[0038] In the design of the Actor module, this embodiment breaks through the limitations of traditional neural networks that directly output action probability distributions. It employs the Gemma large language model to implement the Actor function, parsing the input high-level task (such as a complex task described in natural language) into a structured, executable sequence of subtasks. Leveraging Gemma's powerful natural language understanding and generation capabilities, it can identify implicit constraints of the task, resolve subtask dependencies, and generate appropriately granular and semantically clear subtask units. This upgrades the action space of reinforcement learning from low-level operations to semantic subtasks, significantly reducing the complexity of policy learning. Regarding the reward function design, this method abandons the traditional manual rule design approach, allowing Gemma to directly predict the immediate reward *r* of the subtask. After the generated subtask is input into Gemma, the model outputs a reward value based on its understanding of the task's semantics and execution scenario. This mechanism gives the reward prediction semantic rationality, dynamically adapts to different domain tasks, and aligns the reward granularity with the semantic granularity of the subtask, making the feedback signal more consistent with human perception of task progress.

[0039] In the design of the Critic module, the rationality and value of the decomposition strategy are evaluated. Based on whether the decomposition scheme can efficiently advance the overall task and whether it meets the constraints, a value score is output: (1) (2) (3) Where Vprogress is the progress value, Vconstraint is the constraint value, V is the combination value, st represents the state at time t, and at represents the atomic action (subtask) at time t of the actor policy network decomposition.

[0040] For the above scheme, the overall policy gradient update formula is as follows: (4) (5) Among them, a t Let s be the action at time t. t Let be the state at time t, and r be the instantaneous reward predicted by Gemma from the subtask input. It is the state value function. The entire training process of the framework forms a closed loop with Gemma as the core: the high-level task input Gemma (Actor) generates a sequence of subtasks; the subtask input Gemma predicts the immediate reward r; the Critic module outputs the state value based on the environment state and the subtask execution results. (st) and (st+1); combined with logπθ(a) t |st Instant reward r and value error (st+1)- (st) updates the parameters of Gemma (Actor) and Critic through the policy gradient formula, enabling Gemma to continuously learn during the process of understanding the task, generating subtasks, and predicting rewards, thereby improving the rationality of subtask decomposition and the accuracy of reward prediction.

[0041] To further improve the reliability and rationality of the strategy, this study introduces a value scheduler module based on the aforementioned Actor-Critic framework, as shown below.

[0042] (6) When the Critic module's score for the current strategy falls below its preset threshold Vthreshold, the value scheduler will trigger a task replanning process: the original task is re-inputted into the Gemma module, which then performs the decomposition operation of the complex task into a sequence of subtasks. Subsequently, based on the newly generated subtasks, the Actor-Critic strategy generation, value evaluation, and parameter update process is restarted until the Critic's score for the strategy meets the threshold requirement. This forms an enhanced closed loop of "strategy quality inspection - replanning - iterative optimization," ensuring the rationality and effectiveness of the task decomposition strategy.

[0043] To address the core contradictions of traditional single-reward mechanisms in long-term tasks, such as inefficient training with sparse rewards and tendency for dense rewards to deviate from the target, this study designs a non-fusion hierarchical density reward mechanism. Instead of merging sparse and dense rewards into a single signal through weighting or other methods, this mechanism preserves the independence of the two types of rewards, allowing them to play the roles of global target anchoring and local action guidance, respectively, in the reinforcement learning process. This division of labor and collaboration solves the credit allocation problem while avoiding target deviation or feedback distortion that may occur during fusion.

[0044] The non-fusion density reward mechanism is built on the principle of functional separation and clear division of responsibilities. Its core logic can be summarized as non-overlapping objectives: sparse rewards focus solely on the completion of the overall task, providing a binary final feedback to ensure that strategy optimization does not deviate from the core objective; dense rewards focus solely on the execution quality of sub-tasks / single-step actions, providing refined process feedback and resolving the ambiguity in credit allocation for intermediate steps. Since the two have no functional overlap, there is no need to balance weights through fusion, avoiding the weakening of the role of a certain type of reward due to improper weight settings. In the non-fusion mechanism, sparse rewards and dense rewards remain completely independent, providing feedback from the global goal and local process dimensions, respectively. The computational logic, triggering conditions, and application scenarios within the module are completely independent, as detailed below: The sparse reward module acts as a "rigid constraint on the global task objective," outputting a signal only when the task is completely terminated. Its core function is to determine "whether the task was ultimately successful," providing a "goal-oriented final verification" for robot task planning. This avoids neglecting the global objective due to excessive focus on local optimization. The specific formula is as follows: (7) The dense reward module acts as an "optimization guide for subtasks and single-step actions," dynamically triggered during task execution. Its core function is to provide the robot with "instantaneous, fine-grained process feedback," helping it quickly adjust action parameters and subtask strategies. This alleviates the training inefficiency caused by "fuzzy credit allocation in intermediate steps" in long-cycle tasks. The specific formula is as follows: (8) Where C(t) represents the task completion rate and G(t) represents the capture quality.

[0045] In summary, the non-fusion hierarchical density reward mechanism proposed in this embodiment, through its core design of functional separation and independent feedback, accurately addresses the core contradictions of traditional single reward mechanisms in long-cycle tasks. This mechanism clearly separates sparse and dense rewards into two independent modules: global goal constraint and local process guidance. This avoids the training inefficiency caused by sparse rewards due to feedback lag, and also avoids the risk of dense rewards deviating from the global goal due to excessive focus on local aspects. Furthermore, it eliminates the need for weighted fusion to balance the weights of the two, completely eradicating the problem of weakened reward effects or distorted feedback caused by improper weight settings. From triggering timing to computational logic, both types of rewards remain independent and perform their respective functions. Sparse rewards provide binary goal verification at task termination, while dense rewards provide refined optimization basis through dynamic process triggering. Together, they construct a globally consistent and locally efficient reward feedback system, laying a crucial feedback foundation for embodied intelligent robots to solve the credit allocation problem and improve the robustness of task planning in long-cycle tasks.

[0046] Example 2 An embodied intelligent robot task planning system, comprising: The data acquisition module is configured to acquire real-time environmental data and target tasks. The encoding module is configured to encode real-time environmental data and target tasks using a pre-trained visual language model to obtain a unified representation; The dynamic task planning module is configured to generate a sequence of subtasks based on a unified representation using the actor-critic algorithm. It comprehensively considers efficiency, constraint compliance, and executability to evaluate the rationality of the planning strategy based on the subtask sequence. If the score is lower than the threshold, the subtask sequence is regenerated; otherwise, the corresponding action is generated and executed. During the execution, the subtask sequence and the completion status of the target task are obtained, and the planning strategy is continuously optimized to achieve dynamic planning.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0048] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0049] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0050] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0051] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A body-aware robot task planning method, characterized by, The method comprises the following steps: obtaining real-time environment data and a target task; encoding the real-time environment data and the target task by using a pre-trained visual language model to obtain a unified representation; generating a subtask sequence based on the unified representation by using an actor-critic algorithm, comprehensively considering efficiency, constraint compliance and executability, evaluating the rationality of a planning strategy based on the subtask sequence, if the score is lower than a threshold, re-generating the subtask sequence, otherwise, generating a corresponding action and executing, in the execution process, obtaining the subtask sequence and the completion condition of the target task, and continuously optimizing the planning strategy to realize dynamic planning.

2. The embodied intelligent robot task planning method of claim 1, wherein, The process of generating a subtask sequence based on the unified representation by using the actor-critic algorithm comprises: using a Gemma large language model to realize the function of an actor policy network, and parsing the input target task into a structured and executable subtask sequence.

3. The embodied intelligent robot task planning method of claim 2, wherein, The immediate reward r of the subtask is directly predicted by the Gemma large language model, and after the generated subtask is input into the Gemma large language model, the model outputs a reward value based on the understanding of the task semantics and the execution scene.

4. The embodied intelligent robot task planning method of claim 1, wherein, The process of comprehensively considering efficiency, constraint compliance and executability to evaluate the rationality of the planning strategy based on the subtask sequence comprises: using a critic module to evaluate the rationality value of the decomposition strategy based on whether the decomposition scheme can efficiently promote the total task and whether it complies with the constraint, to obtain a value score: ； ； ； where V progress is the progress value, V constraint is the constraint value, V is the combined value, s t denotes the state at time t, a t denotes the subtask at time t decomposed by the actor policy network.

5. The embodied intelligent robot task planning method of claim 1, wherein, The overall strategy gradient update formula of the actor-critic algorithm is as follows: ； ； where a t is the action at time t, s t is the state at time t, r is the immediate reward predicted by the Gemma large language model for the subtask input, V φ is the state value function, and the training process of the entire framework forms a closed loop with the Gemma large language model as the core: the Gemma large language model generates a subtask sequence for the target task input; the Gemma large language model predicts the immediate reward r; the critic module outputs the state values Vφ(st) and Vφ(st+1) based on the environment state and the subtask execution result; and the parameters of the actor-critic algorithm are updated through the policy gradient formula by combining logπθ(a t |s t ), the immediate reward r, and the value error V φ (s t +1) - V φ (s t ).

6. The embodied intelligent robot task planning method of claim 1, wherein, If the score is lower than the threshold, the process of re-generating the subtask sequence comprises: when the score of the current strategy of the critic module is lower than its preset threshold, a value scheduler is used to trigger the task re-planning process, that is, the original target task is input into the Gemma large language model again, the Gemma large language model performs the decomposition operation of the complex task into the subtask sequence again, and then based on the newly generated subtask, the strategy generation, value evaluation and parameter update process of the actor-critic algorithm is restarted until the score of the strategy of the critic module meets the threshold requirement.

7. The embodied intelligent robot task planning method of claim 1, wherein, In the process of continuously optimizing the planning strategy, a non-fusion hierarchical density reward mechanism is adopted, without combining sparse rewards and dense rewards into a single signal, the independence of the two types of rewards is preserved, and the two types of rewards are allowed to perform division of labor and cooperation in the reinforcement learning process, with the global target anchor and the local action guide respectively.

8. The embodied intelligent robot task planning method of claim 7, wherein, The non-fusion hierarchical density reward mechanism comprises sparse rewards and dense rewards, wherein the sparse rewards are rigid constraint devices for global task targets, and only output signals when the task is completely terminated, which are used to clearly indicate whether the task is successful finally, and provide target-oriented final verification for robot task planning; The dense rewards are optimization guides for subtasks and single-step actions, which are triggered dynamically during task execution, and are used to provide process feedback for robots.

9. A body-aware robot task planning method according to claim 7 or 8, characterized in that, The sparse reward is: Dense rewards are: Wherein, C(t) represents the task completion degree, G(t) represents the crawling quality, , are corresponding weights.

10. A body-aware intelligent robot task planning system, characterized by, The method comprises: a data acquisition module configured to obtain real-time environment data and a target task; an encoding module configured to encode the real-time environment data and the target task by using a pre-trained visual language model to obtain a unified representation; The dynamic task planning module is configured to generate a subtask sequence based on the unified representation by using an actor-critic algorithm, comprehensively consider efficiency, constraint compliance and executability, evaluate the rationality of a planning strategy based on the subtask sequence, re-generate the subtask sequence if the score is lower than a threshold, otherwise, generate a corresponding action and execute, acquire the subtask sequence and the target task completion condition in the execution process, and constantly optimize the planning strategy to realize dynamic planning.

Images