Robot multi-objective self-evolutionary collaborative optimization method and system for dynamic task scenarios
By constructing a task semantic graph and a two-layer policy architecture, combined with sparse communication and knowledge distillation mechanisms, the problems of target staticization and resource non-adaptability in dynamic task scenarios of robot systems are solved. This enables efficient autonomous optimization and resource optimization of robot systems in complex environments, improving robustness and response flexibility.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ROBOTICS RESEARCH CENTER OF YUYAO CITY
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-02
Smart Images
Figure CN122131820A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of robot control and artificial intelligence technology, specifically to a multi-objective self-evolutionary cooperative optimization method and system for robots in dynamic task scenarios. Background Technology
[0002] In dynamic task scenarios such as industrial automation and intelligent services, robot systems need to achieve a dynamic balance among multiple objectives, including efficiency, energy consumption, and safety, to ensure the robustness of task execution and environmental adaptability. Multi-objective optimization, as a core component of robot decision-making, directly determines the system's task completion efficiency and response flexibility in complex dynamic environments, and is a key support for promoting the intelligent development of robots.
[0003] Currently, robot target optimization techniques are mainly divided into two categories. The first category is static objective function optimization methods, which solve the Pareto front by pre-setting fixed weights and combining them with traditional optimization algorithms. However, these methods cannot perceive the dynamic evolution of task intent. For example, in a sudden obstacle avoidance scenario in a hospital emergency room, they cannot automatically adjust the target weights, resulting in a serious disconnect between the optimization results and the actual task requirements. The second category is dynamic optimization methods based on reinforcement learning. Although these methods can support target adjustment through environmental interaction, they are difficult to put into practical use due to defects such as low efficiency of multi-robot system collaboration, fixed algorithm computational complexity that cannot be adapted to resource-constrained platforms, and the reliance on manually preset rules for target generation. Multi-robot systems need to rely on centralized control or high-frequency communication to coordinate targets, causing communication overhead to increase exponentially with the number of robots. In addition, the definition of the optimization target of the algorithm lacks the ability to autonomously understand the semantics of the task and cannot achieve closed-loop generation of the target from input to evolution. Summary of the Invention
[0004] To address the technical problems of static objectives, inefficient collaboration, and resource misadaptability in existing technologies, this invention provides a multi-objective self-evolutionary collaborative optimization method and system for robots in dynamic task scenarios. The technical solution is as follows:
[0005] A multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios includes the following steps:
[0006] Step 1: Construct a task semantic graph and parse the environmental context in real time. Use a semantic-objective mapping network to dynamically transform high-level task intent into a multi-objective optimization model and dynamic weights.
[0007] Step 2: Construct a two-layer policy architecture that includes a global population and a local population. The local population runs independently on each robot and interacts with the global population that maintains the Pareto optimal policy through sparse communication and knowledge distillation mechanisms.
[0008] Step 3: Monitor the robot's hardware resource status and task urgency in real time, and dynamically adjust the optimization algorithm through the online computing resource perception optimization engine to output the optimal control strategy at the current moment under resource constraints;
[0009] Step 4: Execute the optimal control strategy, collect environmental response and performance feedback data, correct the semantic-target mapping network, and update the optimization target and strategy for the next time step.
[0010] Furthermore, step 1 constructs a task semantic graph centered on the robot and incorporating dynamic environmental elements, with the following structure:
[0011] ,
[0012] Among them, the node set Includes the robot itself, collaborating teammates, dynamic obstacles, and task objective points, all gathered together. Represents the spatial proximity and logical association between nodes; initialization any node Feature vectors containing position, velocity, category, and state information A graph neural network based on an attention mechanism is used for multi-hop neighborhood feature aggregation. The aggregation formula is as follows:
[0013] ,
[0014] in, For nodes The neighborhood group, The weight matrix is a learnable matrix. It is a non-linear activation function. For nodes computed via self-attention mechanism With nodes The correlation coefficient between them, after After layer aggregation, a graph embedding vector containing rich environmental context information is obtained.
[0015] Furthermore, the graph is embedded into a vector input semantic-target mapping network, and the current time step is decoded using a multilayer perceptron. Multi-objective function configuration vector and target activation mask By defining candidate objective function libraries Dynamically generate multi-objective optimization models. It includes basic indicators such as energy consumption, time, safety, and smoothness. The multi-objective optimization model is as follows:
[0016] ,
[0017] in, As decision variables, Indicates the first Whether the target has been activated. This refers to the dynamic weight or priority coefficient of the objective.
[0018] Furthermore, in step 2, the global population is updated using the third-generation non-dominated sorting genetic algorithm, which uses the non-dominated sorting hierarchy partitioning and hyperplane reference point association mechanism to screen individuals; the local population is updated using a population-based proximal strategy optimization algorithm variant, which uses generalized advantage estimation to calculate the advantage function, and performs stochastic gradient ascent by maximizing the objective function that includes the truncation term, value function loss and entropy regularization term.
[0019] Furthermore, the cross-population knowledge interaction in step 2 specifically involves: employing a sparse communication protocol for collaboration, with each robot periodically broadcasting a policy summary of its local population that is at the Pareto front, receiving the policy summary as a teacher signal, and minimizing the co-evolutionary loss function based on knowledge distillation. The guiding role of the global population is integrated into the local update, and the formula is:
[0020] ,
[0021] in, The multi-objective weighted loss for the current task. The distillation coefficient is... For Kullback-Leibler divergence, For the local policy distribution to be optimized, For the global population and the current state The most suitable guidance strategy distribution is determined; and a strategy diversity protection operator is introduced to calculate the crowding distance and remove redundant solutions in the local population.
[0022] Furthermore, step 3 constructs a resource state vector. ,in , , and These represent CPU load, memory usage, battery level, and task urgency, respectively, based on the current resource status. A computation-precision tradeoff scheduling is performed, and the hyperparameter configuration of the optimization algorithm is output through table lookup or reinforcement learning policy network. :
[0023] ,
[0024] in, For dynamic population size, The maximum number of iterations, The activation depth of the policy network, wherein the scheduling satisfies the control frequency constraint. Maximizing the Pareto hypervolume under the premise of:
[0025] ,
[0026] in, To calculate the cost prediction function, This represents the currently available computing resources.
[0027] Furthermore, step 3 utilizes timing correlation to adjust the previous control cycle. Pareto optimal solution set After geometric transformation, the current time is obtained. The initial population is hot-started, and an approximate target evaluation agent is integrated. The agent model is built using Gaussian process regression or a lightweight neural network. To approximate the true high-cost objective function: During the optimization process, only the uncertainty in the surrogate model prediction is considered. Above the threshold The solution is evaluated using real physics or high-fidelity simulation.
[0028] Further, step 4 specifically involves: the actuator receiving the optimal control strategy and executing actions, collecting environmental response data after execution, including actual energy consumption, task completion rate and obstacle avoidance success rate within the time period, storing the environmental response data in the experience replay pool, and periodically fine-tuning the parameters of the semantic-target mapping network and the decision logic of the online computing resource perception optimization engine.
[0029] A multi-objective self-evolutionary cooperative optimization system for robots in dynamic task scenarios, used to implement any of the methods described above, includes the following modules:
[0030] The dynamic multi-objective semantic modeling module is used to construct a task semantic graph and utilize a semantic-objective mapping network to dynamically generate multi-objective functions and weights based on environmental features.
[0031] The distributed strategy co-evolution module is used to maintain the global population and local populations, which are updated by non-dominated sorting genetic algorithm and proximal strategy optimization algorithm, respectively, and to perform cross-population knowledge distillation based on sparse communication and diversity protection based on crowding distance.
[0032] The online computing resource awareness optimization engine is used to dynamically schedule the computational granularity of optimization algorithms based on the real-time hardware resource status, and combines incremental frontier maintenance and approximate target evaluation agent output of the optimal control strategy.
[0033] Beneficial effects
[0034] This invention breaks free from the constraints of traditional static target pre-setting by introducing a dynamic multi-objective semantic modeling mechanism, endowing robots with the cognitive ability to autonomously generate and evolve targets based on environmental context. At the collaborative level, the proposed distributed architecture, combined with a two-layer population and knowledge distillation mechanism, successfully achieves Pareto front collaboration of large-scale clusters under low bandwidth constraints, significantly improving the robustness and scalability of the system. This invention incorporates computing resources into the optimization closed loop, and by dynamically adjusting the algorithm granularity, allows complex deep optimization models to run in real time on resource-constrained embedded platforms, achieving the optimal balance between decision accuracy and computational efficiency. Attached Figure Description
[0035] Figure 1 A flowchart for a multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios;
[0036] Figure 2 This is a system architecture diagram for a robot's multi-objective self-evolutionary collaborative optimization system for dynamic task scenarios. Detailed Implementation
[0037] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.
[0038] like Figure 1 As shown, the multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios of the present invention includes the following steps:
[0039] Step 1: Dynamic Multi-Objective Semantic Modeling
[0040] Step 1.1: Task Semantic Graph Construction and Feature Aggregation
[0041] To address the element relationships in unstructured environments, this invention first constructs a robot-centric local task semantic graph (TSG), the structure of which is as follows:
[0042] ,
[0043] The set of nodes It includes the robot's own nodes, collaborating teammate nodes, dynamic obstacle nodes, and task target nodes. For any node in the graph... Initialize its feature vector, which includes position, velocity, entity category (represented by one-hot encoding), and task urgency label. A graph neural network (GNN) based on an attention mechanism is employed for multi-hop neighborhood feature aggregation to capture the potential impact of environmental factors on the current task. For the first... Layer feature update, compute nodes Its neighboring nodes Attention coefficient between This coefficient is calculated through a linear transformation layer processed by the LeakyReLU activation function and is used to quantify the correlation between nodes. The feature aggregation formula is as follows:
[0044] ,
[0045] in, Represents a node The neighborhood group, The weight matrix is a learnable matrix. This is a non-linear activation function. After... After layer aggregation, a graph embedding vector containing rich environmental context information is obtained.
[0046] Step 1.2: Dynamic generation of objective function and priority mapping
[0047] The graph embedding vector obtained in step 1.1 is input into the Semantic-Object Mapping Network (ST-Net), which is composed of a multilayer perceptron (MLP). ST-Net can parse the semantic state of the current environment, including "congested", "emergency", and "open", and decode the current time. Multi-objective function configuration vector and target activation mask .
[0048] Preset candidate target function library This encompasses fundamental metrics such as energy consumption, time, safety, and path smoothness. The dynamically generated multi-objective optimization model is represented as:
[0049] ,
[0050] in, To control policy variables, Decision No. Whether a target is activated, such as activating high-priority obstacle avoidance targets only when the obstacle density exceeds a threshold. The dynamic weight of this objective.
[0051] Step 2: Distributed Strategy Co-evolution
[0052] Step 2.1: Construction and Maintenance of a Two-Layer Strategy Population Architecture
[0053] like Figure 2 As shown in the middle module, this embodiment constructs a two-layer strategy population architecture. By coupling optimization mechanisms at different time scales, it effectively balances the global convergence capability of this invention in complex target spaces with the local response speed to sudden environmental changes. This architecture consists of two closely cooperating parts: a global population and a local population.
[0054] For long-term planning and Pareto front approximation at the global level, this method maintains a scale of global population This population contains a set of elite strategy parameters. Each of them This represents the weight vector of a policy neural network. Given that the dynamic targets generated in step 1.2 are typically high-dimensional, this embodiment employs the third-generation non-dominated sorting genetic algorithm (NSGA-III) to iteratively update the global population, overcoming the crowding distance failure problem of the traditional NSGA-II when handling high-dimensional multi-objectives. First, the parent population... Offspring population generated through genetic manipulation Merge and construct combined populations and to Perform a fast nondominated sort to partition the solution into several nondominated levels. To maintain a uniform distribution of solutions in the normalized high-dimensional objective space, this invention introduces a set of predefined reference points. Distributed on the hyperplane, calculate the first... Objective function value of each individual in each nondominated layer The perpendicular distance to each reference line is used to complete the association operation. Based on this, a niche counting method is used to select individuals from the critical layer, prioritizing solutions associated with sparse reference points to fill the blank areas in the target space. Next-generation global population. The generation process follows the following selection mechanism:
[0055] ,
[0056] The offspring generation process employs a simulated binary crossover operator, with two parent weight vectors... , , its first Offspring generated by dimension parameters The calculation is as follows:
[0057] ,
[0058] in It is a random variable determined by the distribution index, thus ensuring the effective exploration of policy parameters in the global space.
[0059] For millisecond-level real-time response at the robot node level, this embodiment deploys a local scale of [scale missing] on each robot. ( Local populations To achieve the dynamically weighted objective function at the current time... To achieve rapid adaptation, a population-based proximal policy optimization algorithm variant is employed in the local population. This method utilizes gradient information to directly drive policy updates, resulting in a faster convergence speed than genetic algorithms. Specifically, each policy in the local population... Parallel sampling of trajectory data in the current environment And calculate the scalar reward based on the currently activated multi-objective weights. :
[0060] ,
[0061] Based on this, the advantage function is calculated using generalized advantage estimation. To evaluate the value of actions, and to prevent performance degradation due to excessively large step sizes in a single update step, a truncated proxy objective function is constructed, using probability ratios. The update range is limited, and its calculation method is as follows:
[0062] ,
[0063] The optimization objective of the local strategy aims to maximize the following functions, including the truncation term and the value function loss. And the entropy regularization term used to encourage exploration :
[0064]
[0065] Based on this, each local strategy independently performs stochastic gradient ascent to update parameters, as shown in the following equation:
[0066] ,
[0067] Learning rate The computing resource optimization engine dynamically adjusts based on resource status, thereby achieving an organic combination of long-term planning to maintain diversity and immediate response to ensure real-time performance.
[0068] Step 2.2: Cross-population sparse communication and knowledge distillation
[0069] To achieve multi-robot collaboration under low bandwidth conditions, a sparse communication protocol is designed: each robot only broadcasts a policy summary from its local population instead of the complete network parameters, that is, by broadcasting the centroid or extremum of the Pareto non-dominated solution set to reduce the communication burden.
[0070] During the algorithm iteration process, a co-evolutionary loss function based on knowledge distillation is introduced. The robot uses the received global population summary or neighbor policy summary as a teacher signal to adjust its local policy network. Perform a bootstrapping update:
[0071] ,
[0072] in, For the multi-objective weighted loss of the current dynamic task, For adaptive distillation coefficient, Let KL divergence be the KL divergence. For the local policy distribution to be optimized, For the global population and the current state The distribution of the best-matching guidance strategies.
[0073] Furthermore, to prevent local populations from prematurely converging due to overfitting to teacher signals during knowledge distillation, thus losing their ability to explore new environments, this embodiment of the invention employs a diversity protection operator based on crowding distance calculation. This operator quantifies the sparsity of individuals in the target space to eliminate redundant solutions, ensuring the uniformity of the solution set distribution. For local populations... Each body in Firstly, regarding the first Objective function Sort all individuals in ascending order. To preserve the boundary range of the Pareto front, initialize the crowding distance of the first and last boundary individuals to infinity. For any individual in the middle... Its crowded distance Defined as the individual in all The sum of the normalized distances between two adjacent objects in each target dimension is calculated using the following formula:
[0074] ,
[0075] in, and They represent the first time. Individuals under dimensional target ranking The objective function values of successor and predecessor individuals, and These represent the maximum and minimum values of the current population in that target dimension, respectively. Based on the calculated values... During the population maintenance phase, a truncation operation is performed, prioritizing the removal of individuals with the shortest crowding distance until the population size recovers to the preset value. .
[0076] Step 3: Online computing resource awareness optimization
[0077] Step 3.1: Calculation-Accuracy Trade-off Scheduling
[0078] The online computational resource-aware optimization engine (RAO) first reads the robot's hardware status in real time and constructs a resource state vector. ,in , , and These represent CPU load, memory usage, battery level, and task urgency, respectively. A computation-precision tradeoff scheduler is used to optimize the current... Hyperparameter configuration of dynamic output optimization algorithm The scheduling logic is implemented by a lightweight decision network, and its mapping relationship is denoted as:
[0079] ,
[0080] in, For dynamically adjusted population size, The maximum number of iterations, This represents the activation depth of the policy network implemented through the early termination mechanism. The scheduling objective is to satisfy the control frequency constraint. Maximize Pareto hypervolume under the premise of:
[0081] ,
[0082] in To calculate the cost prediction function, This represents the currently available computing resources.
[0083] Step 3.2: Incremental Frontier Maintenance and Approximate Evaluation
[0084] To further reduce computational load, this invention utilizes timing dependencies for warm start. The previous control cycle... Pareto optimal solution set After geometric rigid body transformation, it is used as the current moment. The initial population is optimized to avoid searching from scratch. Simultaneously, an approximate target evaluation agent is integrated to establish an agent model based on Gaussian process regression (GPR). To approximate computationally expensive objective functions, such as safety assessments involving complex dynamics: During the optimization process, only when the variance of the prediction uncertainty of the surrogate model is... Higher than the preset threshold Only then should a real physical simulation be used for accurate evaluation.
[0085] Step 4: Closed-loop feedback and evolutionary update
[0086] The actuator receives the optimal strategy selected by the RAO and executes the action. Environmental response data after execution is collected, including actual energy consumption, task completion rate within the time period, and obstacle avoidance success rate data, and stored in the experience replay pool. This data is used to periodically fine-tune the ST-Net network parameters and the RAO scheduler's decision logic, forming a complete closed loop of perceiving the road environment, generating semantic targets, collaboratively optimizing strategies, adapting computing resources, and executing feedback corrections. This enhances the adaptive capabilities of the multi-robot system in dynamic task scenarios.
[0087] like Figure 2 As shown, the present invention provides a multi-objective self-evolutionary cooperative optimization system for robots in dynamic task scenarios. This system comprises a dynamic multi-objective semantic modeling module (SMOM), a distributed policy cooperative evolution module (DPC), and an online computational resource awareness optimization engine (RAO). This embodiment uses a cluster of mobile robots equipped with an NVIDIA Jetson embedded computing platform, LiDAR, and inertial measurement unit as the application object. SMOM constructs a task semantic graph and uses a semantic-objective mapping network to dynamically generate multi-objective functions and weights based on environmental features. DPC maintains the global and local populations, updating them respectively using a non-dominated sorting genetic algorithm and a proximal policy optimization algorithm, and performs cross-population knowledge distillation based on sparse communication and diversity protection based on congestion distance. RAO dynamically schedules the optimization algorithm based on real-time hardware resource status, and combines incremental frontier maintenance and approximate target evaluation agent outputting the optimal control strategy. The above methods are implemented through the system described in this embodiment of the invention.
[0088] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios, characterized in that, Includes the following steps: Step 1: Construct a task semantic graph and parse the environmental context in real time. Use a semantic-objective mapping network to dynamically transform high-level task intent into a multi-objective optimization model and dynamic weights. Step 2: Construct a two-layer policy architecture that includes a global population and a local population. The local population runs independently on each robot and interacts with the global population that maintains the Pareto optimal policy through sparse communication and knowledge distillation mechanisms. Step 3: Monitor the robot's hardware resource status and task urgency in real time, and dynamically adjust the optimization algorithm through the online computing resource perception optimization engine to output the optimal control strategy at the current moment under resource constraints; Step 4: Execute the optimal control strategy, collect environmental response and performance feedback data, correct the semantic-target mapping network, and update the optimization target and strategy for the next time step.
2. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that: Step 1 involves constructing a task semantic graph centered on the robot and incorporating dynamic environmental elements, with the following structure: ， Among them, the node set Includes the robot itself, collaborating teammates, dynamic obstacles, and task objective points, all gathered together. Represents the spatial proximity and logical association between nodes; initialization any node Feature vectors containing position, velocity, category, and state information A graph neural network based on an attention mechanism is used for multi-hop neighborhood feature aggregation. The aggregation formula is as follows: ， in, For nodes The neighborhood group, The weight matrix is a learnable matrix. It is a non-linear activation function. For nodes computed via self-attention mechanism With nodes The correlation coefficient between them, after After layer aggregation, a graph embedding vector containing rich environmental context information is obtained.
3. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 2, characterized in that: The graph is embedded into a vector input semantic-target mapping network, and the current time step is decoded by a multilayer perceptron. Multi-objective function configuration vector and target activation mask By defining candidate objective function libraries Dynamically generate multi-objective optimization models. It includes basic indicators such as energy consumption, time, safety, and smoothness. The multi-objective optimization model is as follows: ， in, As decision variables, Indicates the first Whether the target has been activated. This refers to the dynamic weight or priority coefficient of the objective.
4. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that: In step 2, the global population is updated using the third-generation non-dominated sorting genetic algorithm, which uses the non-dominated sorting hierarchy partitioning and hyperplane reference point association mechanism to screen individuals; the local population is updated using a population-based proximal strategy optimization algorithm variant, which uses generalized advantage estimation to calculate the advantage function, and performs stochastic gradient ascent by maximizing the objective function that includes the truncation term, value function loss and entropy regularization term.
5. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that, The cross-population knowledge interaction in step 2 specifically involves: coordinating using a sparse communication protocol, with each robot periodically broadcasting a policy summary of its local population that is at the Pareto front, receiving the policy summary as a teacher signal, and minimizing the co-evolutionary loss function based on knowledge distillation. The guiding role of the global population is integrated into the local update, and the formula is: ， in, The multi-objective weighted loss for the current task. The distillation coefficient is... For Kullback-Leibler divergence, For the local policy distribution to be optimized, For the global population and the current state The most suitable guidance strategy distribution is determined; and a strategy diversity protection operator is introduced to calculate the crowding distance and remove redundant solutions in the local population.
6. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that: Step 3 involves constructing the resource state vector. ,in , , and These represent CPU load, memory usage, battery level, and task urgency, respectively, based on the current resource status. A computation-precision tradeoff scheduling is performed, and the hyperparameter configuration of the optimization algorithm is output through table lookup or reinforcement learning policy network. : ， in, For dynamic population size, The maximum number of iterations, The activation depth of the policy network, wherein the scheduling satisfies the control frequency constraint. Maximizing the Pareto hypervolume under the premise of: ， in, To calculate the cost prediction function, This represents the currently available computing resources.
7. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that: Step 3 utilizes timing correlation to update the previous control cycle. Pareto optimal solution set After geometric transformation, the current time is obtained. The initial population is hot-started, and an approximate target evaluation agent is integrated. The agent model is built using Gaussian process regression or a lightweight neural network. To approximate the true high-cost objective function: During the optimization process, only the uncertainty in the surrogate model prediction is considered. Above the threshold The solution is evaluated using real physics or high-fidelity simulation.
8. The multi-objective self-evolutionary cooperative optimization method for robots in dynamic task scenarios as described in claim 1, characterized in that, Step 4 specifically involves: the actuator receiving the optimal control strategy and executing actions, collecting environmental response data after execution, including actual energy consumption, task completion rate and obstacle avoidance success rate within the time period, storing the environmental response data in the experience replay pool, and periodically fine-tuning the parameters of the semantic-target mapping network and the decision logic of the online computing resource perception optimization engine.
9. A multi-objective self-evolutionary cooperative optimization system for robots in dynamic task scenarios, characterized in that, To implement the method as described in any one of claims 1 to 8, the following modules are included: The dynamic multi-objective semantic modeling module is used to construct a task semantic graph and utilize a semantic-objective mapping network to dynamically generate multi-objective functions and weights based on environmental features. The distributed strategy co-evolution module is used to maintain the global population and local populations, which are updated by non-dominated sorting genetic algorithm and proximal strategy optimization algorithm, respectively, and to perform cross-population knowledge distillation based on sparse communication and diversity protection based on crowding distance. The online computing resource awareness optimization engine is used to dynamically schedule the computational granularity of optimization algorithms based on the real-time hardware resource status, and combines incremental frontier maintenance and approximate target evaluation agent output of the optimal control strategy.