Strategy migration method and system applied to cross-domain unmanned system

Abstract

The invention provides a strategy migration method and system applied to a cross-domain unmanned system, and relates to the technical field of strategy migration, and the method comprises the steps: obtaining a task instruction, platform information and environment information of the cross-domain unmanned system; computing platform-independent ownership potential representations and serializing the ownership potential representations to obtain an ownership token sequence; processing the task instruction and the token sequence with the body through a pre-constructed model to obtain a high-level strategy token sequence; the high-level strategy token sequence is converted into a high-level strategy; based on a cross-domain strategy distillation loss processing high-level strategy, obtaining a root domain high-level strategy and a target domain high-level strategy after alignment; receiving a high-level strategy and a real-time platform state detected by a state space through a pre-constructed lightweight control head, generating a bottom-layer control instruction and outputting the bottom-layer control instruction; and detecting and generating an execution result of the underlying control instruction, and adaptively updating the model and the lightweight control head. According to the invention, a unified deployment framework is constructed, and plug-and-play cross-platform deployment is realized through modes of cloud scheduling and end-side loading.

Description

Technical Field

[0001] This invention relates to the field of strategy migration technology, and specifically to a strategy migration method and system applied to cross-domain unmanned systems. Background Technology

[0002] In the development of cross-domain unmanned systems and embodied intelligence, how to achieve efficient migration and unified deployment of control strategies among heterogeneous platforms such as unmanned aerial vehicles (UAVs), unmanned ships (USVs), unmanned vehicles (UGVs), and underwater robots (AUVs), while ensuring the safety and reliability of the system, has become a key challenge in current intelligent system research and engineering applications.

[0003] The industry hopes that various unmanned systems can overcome the limitations of "working in isolation" and share a unified cognitive model, task planning capabilities, and decision-making strategies, thereby forming a collaborative intelligent cluster. However, due to significant differences in appearance, perception perspective, dynamic characteristics, morphological configuration, sensor modes, and task constraints among different platforms, directly migrating strategies from one platform to another often leads to severe performance degradation, or even strategy failure or execution failure. Traditional methods generally achieve strategy transfer by mitigating appearance or scene differences, such as reducing the visual discrepancy between simulation and display through texture and lighting perturbations. In addition, some studies have improved the adaptability of strategies to the dynamic differences of different platforms through dynamic perturbations or system recognition.

[0004] However, because the above methods address visual domain differences or dynamic domain differences separately, their effectiveness in migrating various types of unmanned systems with significant differences in platform structure is extremely limited. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a strategy migration method and system for cross-domain unmanned systems, solving the problem of limited strategy migration effectiveness for various types of unmanned systems with significant differences in platform structure.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a strategy migration method applied to a cross-domain unmanned system, the method comprising: Acquire mission instructions, platform information, and environmental information from cross-domain unmanned systems; Based on the task instructions, platform information and environmental information, calculate and serialize the platform-independent embodied potential representation to obtain the embodied token sequence; The task instructions and the embodied token sequence are processed by a pre-built strategy generation model to obtain a high-level strategy token sequence; and the high-level strategy token sequence is converted into a high-level strategy. The high-level policy is processed using cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy. The system receives the real-time platform status from the high-level policies and state space detection through a pre-built lightweight control head, generates low-level control commands, and outputs them. The execution result of the underlying control command is detected, and the strategy generation model and the lightweight control head are adaptively updated based on the execution result using efficient parameter fine-tuning technology.

[0007] Preferably, the computation of platform-independent embodied potential representations based on the task instructions, platform information, and environmental information includes: The task instructions, platform information, and environmental information are abstracted into a domain-labeled Markov decision process, defining the state space, action space, state transition probability, reward function, and discount factor. The domain embedding vector and modality indication vector are calculated based on the Markov decision process. The raw observation data obtained from the state space observation, the domain embedding vector, and the modality indicator vector are input into a pre-constructed cross-modal encoder to obtain an embodied latent representation.

[0008] Preferably, the step of converting the high-level policy token sequence into a high-level policy includes: Pre-build a cross-platform, reusable skills library; Based on the skill library, the high-level policy token sequence is converted into a high-level policy through skill decoding mapping; wherein, the high-level policy includes skills and skill parameters, and each skill is defined by a set of corresponding sub-policies and termination functions.

[0009] Preferably, the optimization objective of the parameter high-efficiency fine-tuning technology is: In the formula, Indicates task loss; Indicates cross-domain consistency. Regular expression terms representing the modules to be adapted; , Indicates the weighting coefficient; This indicates the backbone parameters of the strategy-generated model; The parameters of the adaptation module are used for efficient fine-tuning. Indicates a lightweight control head; Indicates the sequence of embodied tokens; cIndicates natural language task instructions; Represents the root domain The distribution of high-level strategies is output. Represents the target domain The distribution of high-level strategies is output.

[0010] Preferably, the step of adaptively updating the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology includes: The system detects preset indicators and calculates a risk score using a comprehensive risk coefficient function. The preset indicators include monitoring and control command distribution deviation, constraint violation rate, and performance degradation rate. In response to the risk score exceeding a preset threshold, the output weight of the strategy generation model is reduced or a stable strategy is switched.

[0011] Preferably, the expression for the comprehensive risk coefficient function is: In the formula, This is the strategy offset. To control error, For safety constraints and regulations, , , These are weighting coefficients. Continuously adjusted to ensure the overall risk coefficient is maintained. Less than the risk threshold .

[0012] Preferably, before generating and outputting the underlying control instructions, the method further includes: Anomaly detection is performed on the underlying control commands, including physical feasibility checks, collision prediction, and energy margin assessment. In response to the result of the anomaly detection being an anomaly, the underlying control command is switched to a conservative strategy.

[0013] Secondly, the present invention also provides a strategy migration system applied to cross-domain unmanned systems, the system comprising: The acquisition module retrieves task instructions, platform information, and environmental information from the cross-domain unmanned system. The first processing module calculates and serializes a platform-independent embodied potential representation based on the task instructions, platform information and environmental information to obtain an embodied token sequence. The second processing module processes the task instructions and the embodied token sequence through a pre-built strategy generation model to obtain a high-level strategy token sequence; and converts the high-level strategy token sequence into a high-level strategy. The third processing module processes the high-level policy based on cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy. The control generation module receives the real-time platform status from the high-level strategy and state space detection through a pre-built lightweight control header, generates low-level control commands, and outputs them. The update module detects the execution result of the underlying control command and adaptively updates the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology.

[0014] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program for policy migration applied to a cross-domain unmanned system, wherein the computer program causes a computer to execute the policy migration method for cross-domain unmanned systems as described above.

[0015] Fourthly, the present invention also provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing policy migration applied to cross-domain unmanned systems as described above.

[0016] (III) Beneficial Effects This invention provides a strategy transfer method and system for cross-domain unmanned systems. Compared with existing technologies, it has the following advantages: This invention enables various unmanned systems to operate collaboratively under a shared high-level command system by establishing a consistent embodied representation space, cross-modal policy interface, and unified large model decision framework, ensuring the consistency and continuity of behavioral logic across platforms and tasks.

[0017] This invention explicitly models platform differences and constructs a unified embodied representation space, enabling heterogeneous platforms such as different types of UAVs, unmanned ships, unmanned vehicles, and underwater robots to achieve "platform-independent task planning + platform-related execution control" within a shared large model strategy framework. This ensures both strategy consistency and the accuracy of platform physical layer control, fundamentally solving problems in existing technologies such as broken cross-platform capabilities, low migration efficiency, and the inability of resource-constrained devices to support large models. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a strategy migration method for cross-domain unmanned systems provided in an embodiment of this application; Figure 2 This is a flowchart illustrating a strategy migration method for cross-domain unmanned systems, provided as another embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] This application provides a strategy migration method and system for cross-domain unmanned systems, which solves the long-standing problems of decision-making fragmentation and difficulty in strategy universality in the execution of tasks by heterogeneous platforms such as different types of UAVs, unmanned ships, unmanned vehicles, and underwater robots. It realizes the comprehensive needs of cross-domain unmanned systems for a unified strategy hub, efficient migration capabilities, and safe and reliable execution under multi-platform, multi-scenario, and multi-task conditions.

[0022] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: In recent years, some studies have attempted to construct morphology-independent intermediate skill representations to achieve policy reuse across different robotic platforms. However, most of these methods are limited to situations with similar limb structures or relatively simple task scenarios, and still have significant shortcomings in supporting multimodal observation inputs and unified representations across platforms.

[0023] With the development of large-scale model technology, multimodal large-scale models and vision-language-action (VLA) models are gradually being introduced into embodied intelligence scenarios. Some studies have achieved end-to-end high-level policy generation from natural language commands to robot actions by fusing visual features with large language models; others have proposed a "brain-cerebellum" architecture, using large models for high-level task planning and lightweight controllers to perform low-level motion control, thereby giving robots a certain degree of cross-task generalization ability. However, existing methods generally rely on cloud computing resources, making them difficult to deploy directly on small unmanned platforms with limited computing, storage, and energy consumption; at the same time, they lack a unified policy interface for multiple platforms and have not designed a safety guidance mechanism in conjunction with the large model, which can easily lead to erroneous decisions, unauthorized control, or security risks during cross-platform migration.

[0024] In summary, practical cross-domain unmanned system applications often face multiple challenges, including differences in multi-source perception domains, heterogeneous platform forms, highly diverse tasks, and strict security constraints. Based on these issues, the policy migration method and system for cross-domain unmanned systems in this application provide a novel overall architecture integrating a "unified large-scale model policy hub—efficient cross-domain adaptation—intrinsic security control" to support the collaborative intelligence and reliable operation of heterogeneous platforms such as unmanned aerial vehicles (UAVs), unmanned vehicles (UGVs), unmanned ships (USVs), and underwater robots (AUVs).

[0025] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0026] like Figure 1 As shown in the figure, this application provides a policy migration method applied to a cross-domain unmanned system, and the specific implementation steps include: Step S110: Obtain the task instructions, platform information, and environmental information of the cross-domain unmanned system.

[0027] Step S120: Calculate and serialize the platform-independent embodied potential representation based on the task instructions, platform information, and environmental information to obtain an embodied token sequence.

[0028] Step S130: Process the task instructions and the embodied token sequence through a pre-built strategy generation model to obtain a high-level strategy token sequence; and convert the high-level strategy token sequence into a high-level strategy.

[0029] Step S140: Process the high-level policy based on cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy.

[0030] Step S150: Receive the real-time platform status from the high-level strategy and state space detection through the pre-built lightweight control head, generate low-level control commands, and output them.

[0031] Step S160: Detect the execution result of the underlying control command, and adaptively update the strategy generation model and the lightweight control head based on the execution result using parameter efficient fine-tuning technology.

[0032] See Figure 2 The strategy migration method proposed in this application constructs a cross-domain unified architecture of "cloud-based large model strategy hub + edge-side lightweight execution stack". Through the connection of four subsystems, namely "cross-domain multimodal state coding and embodied token module - large model-driven high-level strategy and skill library module - lightweight control head (platform-specific control adaptation and parameter high-efficiency migration module) - unified deployment and migration stability management module", the end-to-end integrated design of high-level task planning, cross-domain skill migration and low-level control output is realized.

[0033] The policy migration method for cross-domain unmanned systems provided in this application constructs a cross-domain policy migration mechanism that combines generalization ability, interpretability, and engineering feasibility. This enables various types of unmanned systems to quickly adapt to new mission requirements in complex and ever-changing real-world scenarios, supports cross-scenario deployment and multi-domain collaborative operations, and possesses secure autonomous decision-making capabilities. This lays a unified technical foundation for the intelligent collaboration of large-scale heterogeneous unmanned systems in the future.

[0034] In one embodiment, step S110 involves acquiring the task instructions, platform information, and environmental information of the cross-domain unmanned system.

[0035] This embodiment aims to acquire data to provide an input basis for subsequent policy migration.

[0036] In one embodiment, step S120 involves calculating and serializing a platform-independent embodied potential representation based on task instructions, platform information, and environmental information to obtain an embodied token sequence.

[0037] The specific implementation of this embodiment includes the following steps: Step S121: The task instructions, platform information and environmental information are abstracted into a domain-labeled Markov decision process, defining the state space, action space, state transition probability, reward function and discount factor.

[0038] Step S122: Calculate the domain embedding vector and modality indication vector based on the Markov decision process.

[0039] Step S123 involves inputting the original observation data obtained from the state space observation, the domain embedding vector, and the modality indication vector into a pre-constructed cross-modal encoder to obtain an embodied latent representation.

[0040] This invention first abstracts each type of platform and environment into a domain-labeled Markov decision process. ,in This represents a composite domain of "platform-environment-task", and its state. It can include multimodal data such as visual images, LiDAR point clouds, radar echoes, IMU tracks, GPS tracks, rudder angles / motor speeds, etc. This indicates the action space that can be executed within this domain. Represents the state transition probability. Represents the reward function, This is the discount factor.

[0041] To achieve unified input across different modalities, a cross-modal encoder is constructed. The original observations Domain embedding vector and mode indicator vector The input uniform mapping is shown in the following formula: in, This provides platform-independent embodied latent representations. The encoder internally consists of a visual encoder (CNN / ViT structure), a temporal sensor encoder (RNN / Transformer), and a modality fusion layer. It is jointly trained through self-supervised objectives such as contrastive learning, reconstruction prediction, and temporal consistency constraints, enabling the encoder to learn representations that are "task-sensitive but platform-insensitive," providing a unified input for large-scale model policy generation. It is important to clarify that the domain embedding vector and modality indicator vector are not pre-set fixed vectors, but rather intermediate representations calculated through learning based on domain-labeled Markov decision process modeling. The domain embedding vector represents the differences in features across different platform / environment domains, while the modality indicator vector represents different perceptual modalities. Both can be jointly learned during training through the embedding layer or cross-modal coding network and used as input to calculate the embodied latent representation, rather than being pre-set as static prior parameters.

[0042] This application's embodiments solve the problem of inconsistent observation space caused by differences in sensor modes, perspectives, sampling dimensions, and environmental structures of different unmanned platforms by constructing a cross-modal encoder, enabling large models to uniformly understand the environment without relying on the original observation structure of a specific platform.

[0043] In one embodiment, step S130 involves processing the task instructions and the embodied token sequence using a pre-built policy generation model to obtain a high-level policy token sequence; and then converting the high-level policy token sequence into a high-level policy. The specific implementation of this embodiment includes the following steps: Step S131: Build a cross-platform reusable skill library.

[0044] Step S132: Based on the skill library, the high-level policy token sequence is converted into a high-level policy through skill decoding mapping; wherein, the high-level policy includes skills and skill parameters, and each skill is defined by a set of corresponding sub-policies and termination functions.

[0045] In this embodiment, it is important to clarify that a sub-strategy is a low-level executable control strategy unit used to implement a skill, distinct from the abstract behavioral description of an intermediate strategy. Specifically, a "sub-strategy" refers to a low-level executable control strategy unit used to implement a certain skill, which concretizes high-level strategies or skill parameters into executable control behaviors. For example, and not limitingly, consider the skill "target point navigation": this skill can correspond to multiple sub-strategies, such as a straight-line tracking sub-strategy, an obstacle avoidance adjustment sub-strategy, and a deceleration approach sub-strategy; each sub-strategy is responsible for executing the corresponding control behavior under specific state conditions, and a termination function determines whether the current sub-strategy is completed or switches to another sub-strategy. In contrast, an "intermediate strategy" typically refers to an abstract behavioral layer between the task-level strategy and the underlying control, used to describe what type of action to take (such as cruising, following, or searching), while a sub-strategy focuses more on describing how the action is completed at the specific execution level. The two differ in their abstract level and functional granularity.

[0046] In this embodiment, the policy generation model is driven by a large model; to enable high-level policies to be migrated to different platforms, this invention pre-constructs a cross-platform reusable skill library. Each skill This corresponds to a type of platform-independent but parameterizable intermediate strategy, such as directional movement, area inspection, target circling, and trajectory following. Each skill consists of a set of sub-strategies. With termination function Define and ensure that skills have a consistent semantic structure across different platforms.

[0047] Through skill decoding mapping , strategy token sequence Converting to skills and their parameters (i.e., high-level policies), the expression is: in These are skill parameters (such as speed and direction, target point coordinates, track length, altitude maintenance, etc.).

[0048] In one embodiment, step S140 involves processing the high-level policy based on cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy.

[0049] In this embodiment, it should be clarified that the high-level policy token sequence is an intermediate representation obtained during the unified policy generation process, and its semantic structure remains consistent across different domains; and the distinction between the "root domain high-level policy" and the "target domain high-level policy" is introduced during the subsequent alignment process based on cross-domain policy distillation loss, and is used to describe the output results of the same high-level policy under different domain conditions.

[0050] To enhance cross-domain consistency, in the root domain With the target domain The same task was performed on both sides to obtain the trajectory dataset, and the distribution of the high-level policy output by the large model in both domains was statistically analyzed. , By using cross-domain strategy distillation loss, we can ensure that different platforms output as consistent a high-level strategy structure as possible under the same task instructions. In the formula; Indicates the sequence of embodied tokens; c Indicates natural language task instructions; Represents the root domain The distribution of high-level strategies is output. Represents the target domain The distribution of high-level strategies is output.

[0051] This loss ensures structural consistency in strategy generation across platforms, and is the core step in this invention to achieve write-once strategy, shared execution across multiple platforms.

[0052] In one embodiment, step S150 involves receiving the real-time platform state from the high-level policy and state space detection via a pre-built lightweight control header, generating and outputting low-level control commands. This embodiment further includes the following steps: Step S151: Perform anomaly detection on the underlying control command. The anomaly detection includes physical feasibility check, collision prediction, and energy margin judgment.

[0053] Step S152: In response to the result of the anomaly detection being an anomaly, the underlying control command is switched to a conservative strategy.

[0054] In this embodiment, the skills output from the skill library are... With skill parameters (High-level strategies) are mapped to the underlying control commands of each platform, enabling the execution and implementation of cross-platform strategies.

[0055] For each platform The embodiments of the present invention construct a lightweight control head. Receive skill parameters and current platform status Generate low-level control commands ,have: The control head can adopt a combination of "small neural network + traditional controller (PID / MPC)". The neural network is responsible for nonlinear mapping of skill parameters, while the traditional controller ensures system stability and physical interpretability.

[0056] In one embodiment, step S160 involves detecting the execution result of the underlying control instructions and adaptively updating the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology. This embodiment further includes the following steps: Step S161: Detect preset indicators and calculate risk scores using a comprehensive risk coefficient function. The preset indicators include monitoring control command distribution offset, constraint violation rate, and performance degradation rate. Step S162: In response to the risk score exceeding a preset threshold, reduce the output weight of the strategy generation model or switch to a stable strategy.

[0057] In this embodiment, it should be clarified that "conservative strategy" and "stable strategy" are not the same specific strategy, but they have similarities. Both are used to describe the strategy execution method of imposing safety constraints or adjustments on the output of an existing strategy under risk assessment trigger conditions. Specifically, the conservative strategy focuses on limiting the magnitude of the strategy output when the risk is high, while the stable strategy focuses on smoothing the strategy output when the system fluctuates greatly.

[0058] Regarding migration and deployment, to enable the large model capabilities to be deployed on resource-constrained platforms, this module introduces a parameter efficient fine-tuning technique (PEFT). Its core idea is to adjust only a small number of additional parameters (such as LoRA weights, the Adapter layer, or the quantized lightweight module) while keeping the backbone model frozen. Specifically: Large model side: Freeze the backbone and only update LoRA / Adapter to fine-tune for new tasks; Control head side: Rapid platform adaptation through a very small-scale adaptation layer (such as MLP, low-rank update); Lightweight deployment: The control head can be distilled and quantized to meet the real-time requirements of edge inference.

[0059] The optimization goal of the parameter fine-tuning technique is: in, Indicates task loss; Indicates cross-domain consistency. Regular expression terms representing the modules to be adapted; , Indicates the weighting coefficient; Indicates a lightweight control head; τ represents an embodied token sequence; c Indicates natural language task instructions; Represents the root domain The distribution of high-level strategies is output. Represents the target domain The distribution of high-level strategies is output. This indicates the backbone parameters of the strategy-generated model; This indicates the parameters of the adaptation module for efficient fine-tuning.

[0060] The unified deployment framework encapsulates the large model strategy hub, cross-modal encoder, skill library, and control heads for each platform into standardized modules, which are centrally stored, updated, and scheduled in the cloud. The corresponding control heads and adaptation modules are loaded on the edge as needed, enabling a "plug-and-play" multi-platform deployment method.

[0061] During runtime, the system automatically selects appropriate skill subsets and large model strategy versions based on platform type, task semantics, environmental status, and security level, achieving adaptive combination of task-level and platform-level approaches.

[0062] To ensure the security of the migration process, this invention incorporates a two-level security mechanism: (1) Safety barrier layer: Before the control output, the physical feasibility of the command is checked, collision prediction is performed, energy margin is judged, etc. Once an abnormality is detected, a conservative strategy (hovering, deceleration, return to home, emergency stop, etc.) is immediately executed.

[0063] (2) Migration Risk Assessment and Adaptive Termination Mechanism: Risk scores are calculated by monitoring indicators such as control instruction distribution shift, constraint violation rate, and performance degradation rate. If the risk exceeds a threshold, the output weights of the large model are reduced or a stable strategy is switched to prevent negative migration from causing uncontrollable behavior. Specifically, a comprehensive risk coefficient function is constructed, with the following expression: The expression for the comprehensive risk coefficient function is: In the formula, This is the strategy offset. To control error, For safety constraints and regulations, , , These are weighting coefficients. Continuously adjusted to ensure the overall risk coefficient is maintained. Less than the risk threshold .

[0064] The migration risk assessment and adaptive termination mechanism provided in this implementation ensures the stability, security, and controllability of the large model during cross-domain execution, avoiding potential uncertainties in policy migration. To ensure the safe and reliable operation of the system during migration, a security barrier layer and a migration risk assessment module are embedded in the policy migration and execution closed loop to monitor and constrain the skills and control variables output by the large model in real time, triggering a rollback strategy when necessary. By continuously monitoring indicators such as policy deviation, control error, and safety constraint violation rate, the system instability and potential security risks caused by negative migration can be effectively suppressed. Ultimately, this enables the cross-domain unmanned system to maintain a controllable, explainable, and recoverable safe behavior pattern even in unfamiliar platforms, complex scenarios, and high-risk environments.

[0065] This invention also provides a policy migration system for cross-domain unmanned systems, comprising: The acquisition module retrieves task instructions, platform information, and environmental information from the cross-domain unmanned system. The first processing module calculates and serializes a platform-independent embodied potential representation based on the task instructions, platform information, and environmental information to obtain an embodied token sequence.

[0066] The second processing module processes the task instructions and the embodied token sequence through a pre-built strategy generation model to obtain a high-level strategy token sequence; and converts the high-level strategy token sequence into a high-level strategy.

[0067] The third processing module processes the high-level policy based on cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy.

[0068] The control generation module receives the high-level strategy and the real-time platform state detected by the state space through a pre-built lightweight control header, generates low-level control commands, and outputs them.

[0069] The update module detects the execution result of the generated underlying control instructions and adaptively updates the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology.

[0070] It is understood that the policy migration system for cross-domain unmanned systems provided in this embodiment of the invention corresponds to the policy migration method for cross-domain unmanned systems described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the policy migration method for cross-domain unmanned systems, and will not be repeated here.

[0071] This invention also provides a computer-readable storage medium storing a computer program for a policy migration method applied to a cross-domain unmanned system, wherein the computer program causes a computer to execute the policy migration method for a cross-domain unmanned system as described above.

[0072] This application also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing policy migration methods applied to cross-domain unmanned systems as described above.

[0073] In summary, compared with existing technologies, it has the following beneficial effects: 1. This application's embodiments construct a "cloud-based large-scale model strategy hub + standardized embodied interface," enabling multiple types of unmanned platforms to share a unified large-scale model strategy framework and skill library structure. All high-level strategy logic is uniformly stored and updated in the cloud, and the edge only needs to load a lightweight platform control header and a few adaptation parameters to execute tasks. This avoids problems such as "multiple versions of large models running in parallel, repetitive training of task logic, and inconsistent platform configurations" in traditional multi-platform systems, significantly simplifying the model update process and reducing system maintenance complexity. The unified strategy hub allows model iteration, optimization, and hot updates to be performed centrally in the cloud, achieving one update and simultaneous effect across multiple platforms, significantly reducing model management, deployment, and maintenance costs.

[0074] 2. The embodiments of this application can handle the multi-source differences existing in real cross-domain environments. By leveraging a unified high-level policy structure and a cross-platform reusable skill library, heterogeneous platforms such as UAVs, unmanned ships, and unmanned vehicles can share similar policy combinations when performing different tasks. When adding a new platform or task, there is no need to recollect a large amount of platform-specific data or completely retrain the policy network. Only minor, lightweight adjustments to the high-level policy mapping or skill parameters are needed to quickly complete the migration. Compared with the traditional reinforcement learning or imitation learning schemes where "each platform and each task requires separate training," this invention significantly reduces training and data acquisition costs, and greatly improves the scalability and economy of the policy in the actual engineering implementation process.

[0075] 3. This application's embodiments employ a strategy of efficient parameter fine-tuning (PEFT), lightweight control heads, and modular deployment to compress and securely migrate the capabilities of large models originally running in the cloud to resource-constrained platforms. For example, small drones, unmanned vessels, and mobile robots are typically limited by computing power, power supply, and real-time control requirements. However, this invention only requires loading a small number of LoRA / Adapter parameters and a very small number of control heads to execute high-level strategies for generating large models on the edge with low inference overhead. This capability deployment mechanism achieves a closed-loop engineering process of "cloud training, edge execution," making the deployment of large models on edge devices feasible and facilitating stable operation in scenarios such as the wild, at sea, and underground spaces where long-term reliance on cloud communication is not possible.

[0076] 4. This application's embodiments simultaneously model cross-platform observation appearance, viewpoint changes, dynamic differences, sensor differences, and platform morphology differences within a unified embodied latent space. Inconsistencies in the input space are eliminated through cross-modal encoders and embodied token mechanisms. Combined with cross-domain policy distillation loss, this ensures consistency in high-level policy structure across different platforms when executing the same task instructions, preventing policy drift or abnormal control behavior due to platform differences. This joint modeling approach enables the policy to maintain stable performance under different environmental conditions, different sensor combinations, and even platform replacements, thus possessing stronger cross-domain generalization ability and robustness, making it suitable for complex, dynamic, and frequently disturbed real-world environments.

[0077] 5. This application's embodiments overcome the problem of unstable strategies in the early stages of migration by constructing a systematic safety barrier and adaptive termination mechanism to control potential hallucination decisions or decisions that violate physical rules from large models. The invention introduces a dual safety mechanism: a safety barrier layer and a migration risk assessment mechanism. The safety barrier layer performs physical feasibility checks, collision predictions, control limits, and energy margin judgments on the skills and control quantities output by the large model. Once a risk is detected, a conservative strategy, such as hovering, emergency stop, or return to base, is immediately triggered. The migration risk control mechanism updates the risk coefficient in real time based on indicators such as strategy drift, control error, and safety constraint violation rate, adjusting model weights or switching to a stable execution mode when necessary. This mechanism effectively reduces potential safety hazards in cross-domain migration, improves the overall controllability, reliability, and scalability of the system, and enables it to operate reliably in scenarios with high safety requirements, such as disaster relief, inspection, and maritime operations.

[0078] 6. The overall architecture of "unified large model + embodied token + skill library + control head + security barrier" formed by this invention is not only applicable to typical cross-domain collaborations such as UAVs and unmanned ships, but can also be extended to a wide range of embodied intelligent systems such as unmanned vehicles, industrial robots, special operation robots, and warehousing and logistics robots. Its unified interface specification, unified strategy structure, skill reuse mechanism, and lightweight platform adaptation method can serve as a general technical foundation for future multi-platform, multi-task, and multi-agent collaborative systems. This technical system is expected to drive the evolution of embodied intelligence from single-platform intelligence to cross-domain collaboration and multi-agent collaboration, laying the foundation for large-scale deployment and networked collaboration of intelligent agents.

[0079] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A strategy transfer method applied to cross-domain unmanned systems, characterized in that, The method includes: Acquire mission instructions, platform information, and environmental information from cross-domain unmanned systems; Based on the task instructions, platform information and environmental information, calculate and serialize the platform-independent embodied potential representation to obtain the embodied token sequence; The task instructions and the embodied token sequence are processed by a pre-built strategy generation model to obtain a high-level strategy token sequence; and the high-level strategy token sequence is converted into a high-level strategy. The high-level policy is processed using cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy. The system receives the real-time platform status from the high-level policies and state space detection through a pre-built lightweight control head, generates low-level control commands, and outputs them. The execution result of the underlying control command is detected, and the strategy generation model and the lightweight control head are adaptively updated based on the execution result using efficient parameter fine-tuning technology.

2. The strategy migration method for cross-domain unmanned systems according to claim 1, characterized in that, The computation of platform-independent embodied potential representations based on the task instructions, platform information, and environmental information includes: The task instructions, platform information, and environmental information are abstracted into a domain-labeled Markov decision process, defining the state space, action space, state transition probability, reward function, and discount factor. The domain embedding vector and modality indication vector are calculated based on the Markov decision process. The raw observation data obtained from the state space observation, the domain embedding vector, and the modality indicator vector are input into a pre-constructed cross-modal encoder to obtain an embodied latent representation.

3. The policy migration method for cross-domain unmanned systems according to claim 1, characterized in that, The step of converting the high-level policy token sequence into a high-level policy includes: Pre-build a cross-platform, reusable skills library; Based on the skill library, the high-level policy token sequence is converted into a high-level policy through skill decoding mapping; wherein, the high-level policy includes skills and skill parameters, and each skill is defined by a set of corresponding sub-policies and termination functions.

4. The strategy migration method for cross-domain unmanned systems according to claim 1, characterized in that, The optimization objective of the parameter fine-tuning technology is: In the formula, Indicates task loss; Indicates cross-domain consistency. Regular expression terms representing the modules to be adapted; , Indicates the weighting coefficient; This indicates the backbone parameters of the strategy-generated model; The parameters of the adaptation module are used for efficient fine-tuning. Indicates a lightweight control head; Indicates the sequence of embodied tokens; c Indicates natural language task instructions; Represents the root domain The distribution of high-level strategies is output. Represents the target domain The distribution of high-level strategies is output.

5. The strategy migration method for cross-domain unmanned systems according to claim 1, characterized in that, The step of adaptively updating the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology includes: The system detects preset indicators and calculates a risk score using a comprehensive risk coefficient function. The preset indicators include monitoring and control command distribution deviation, constraint violation rate, and performance degradation rate. In response to the risk score exceeding a preset threshold, the output weight of the strategy generation model is reduced or a stable strategy is switched.

6. The strategy migration method for cross-domain unmanned systems according to claim 5, characterized in that, The expression for the comprehensive risk coefficient function is as follows: In the formula, This is the strategy offset; To control errors; For safety constraints and regulations; , , The weighting coefficients are continuously adjusted to ensure the overall risk coefficient. Less than the risk threshold .

7. The strategy migration method for cross-domain unmanned systems according to claim 1, characterized in that, Before generating and outputting the underlying control commands, the method further includes: Anomaly detection is performed on the underlying control commands, including physical feasibility checks, collision prediction, and energy margin assessment. In response to the result of the anomaly detection being an anomaly, the underlying control command is switched to a conservative strategy.

8. A strategy migration system applied to cross-domain unmanned systems, characterized in that, The system includes: The acquisition module retrieves task instructions, platform information, and environmental information from the cross-domain unmanned system. The first processing module calculates and serializes a platform-independent embodied potential representation based on the task instructions, platform information and environmental information to obtain an embodied token sequence. The second processing module processes the task instructions and the embodied token sequence through a pre-built strategy generation model to obtain a high-level strategy token sequence; and converts the high-level strategy token sequence into a high-level strategy. The third processing module processes the high-level policy based on cross-domain policy distillation loss to obtain the aligned root domain high-level policy and target domain high-level policy. The control generation module receives the real-time platform status from the high-level strategy and state space detection through a pre-built lightweight control header, generates low-level control commands, and outputs them. The update module detects the execution result of the underlying control command and adaptively updates the strategy generation model and the lightweight control head based on the execution result using efficient parameter fine-tuning technology.

9. A computer-readable storage medium, characterized in that, It stores a computer program for policy migration applied to cross-domain unmanned systems, wherein the computer program causes a computer to execute the policy migration method for cross-domain unmanned systems as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing policy migration methods for cross-domain unmanned systems as described in any one of claims 1 to 7.

Images