An Adaptive Navigation Method and Apparatus for Unstructured Terrain Based on Sim-to-Real Migration

Abstract

This invention discloses an unstructured terrain adaptive navigation method and apparatus based on Sim-to-Real transfer, comprising: architecturally isolating the high-level strategy from the low-level control loop, and introducing a terrain adaptive controller into the low-level control loop; in a simulation environment, using environmental dynamics features encoded from ground truth physical parameters randomized by the domain, performing a first-stage training on the basic control strategy in the terrain adaptive controller, and using interaction history data generated in the first-stage training to perform a second-stage training on the environment predictor in the terrain adaptive controller; and deploying navigation in a real environment, inferring environmental dynamics features in real time through the trained environment predictor, inputting the trained basic control strategy, and outputting the robot's final control commands. This invention achieves zero-sample transfer of the high-level strategy during real-world deployment and is suitable for high-precision and robust adaptive navigation in unstructured terrain.

Description

Technical Field

[0001] This invention belongs to the field of robot autonomous navigation technology, specifically relating to an unstructured terrain adaptive navigation method and device based on Sim-to-Real migration. Background Technology

[0002] Achieving autonomous navigation in complex, unstructured environments has always been a long-standing challenge in the field of field robotics. Existing navigation technologies can be mainly divided into three categories, but all of them have significant technical bottlenecks.

[0003] The first category is end-to-end learning methods, which attempt to directly learn the mapping from raw sensor inputs to vehicle control commands. Among these, imitation learning (IL)-based end-to-end methods perform exceptionally well in structured scenarios such as autonomous driving, but their core drawback lies in their heavy reliance on large-scale, high-quality expert demonstration datasets. For diverse and dynamic robotic tasks, the high cost of acquiring such datasets constitutes a major obstacle to practical deployment. Reinforcement learning (RL)-based end-to-end methods do not require expert data and can learn autonomously through trial and error. However, in mobile robot navigation tasks, especially when facing long-term and sparse reward signals, these methods generally suffer from low sample efficiency and unstable training. Even in some benchmark tests, carefully tuned traditional methods can still outperform end-to-end RL policies.

[0004] The second category consists of traditional trajectory planning methods, such as the Time-Elastic Band (TEB) and Dynamic Window (DWA) methods. These methods are model-based and possess high robustness and interpretability. However, their performance is highly dependent on a set of key parameters (such as maximum velocity and expansion radius) that require fine-tuning based on expert experience, and they lack the ability to adapt to dynamic environments.

[0005] The third category is learning-based hybrid methods, namely automatic parameter tuning, which represents the most promising compromise. Its core idea is to leverage machine learning (especially RL) to dynamically adjust the parameters of the traditional planner, thereby combining the robustness of model planning with the adaptability of learning. Because the planner restricts its exploration to a high-level parameter space and shields itself from unsafe low-level actions, these methods significantly improve sample efficiency and system safety. AppLR and Apple are typical examples of this approach. However, despite the progress made by these hybrid methods, their performance still lags behind high-performance end-to-end imitation learning systems. This invention argues that this performance gap stems not from algorithm design, but from a long-neglected architectural limitation.

[0006] Specifically, the key limitation of this architecture lies in the control frequency. Once trained, end-to-end IL policies can directly output speed commands at extremely high frequencies (e.g., 50Hz), achieving sensitive response. In contrast, the control path of automatic parameter tuning methods is indirect: the learning policy adjusts parameters, the planner must re-optimize or sample to generate a trajectory, and only then can control commands be produced. This additional computational delay limits the frequency of the entire control loop to a relatively low level (e.g., 5-10Hz). This low-frequency, high-latency control loop creates a performance bottleneck for existing parameter tuning methods.

[0007] Furthermore, when navigation tasks are extended to unstructured terrain, all the aforementioned methods face a second serious challenge: the significant gap between simulation and reality. Controllers trained in simulations (whether end-to-end or low-level control in hybrid approaches) assume a low-variance tracking error, but when deployed in the real world, the variance of the tracking error increases significantly due to unmodeled dynamics (such as friction, gravity, and load variations). For hybrid approaches, this variance gap is fatal; the decline in the performance of the low-level controller (i.e., increased variance) directly reduces the optimal value function achievable by the high-level tuning strategy. Therefore, the limitation of existing Sim-to-Real solutions lies in the fact that common controllers (such as PID, MPC, or some adaptive controllers) either rely on high-precision models or only improve tracking performance in isolation, failing to address the core issue of how to ensure the zero-sample transfer capability of the high-level tuning strategy by maintaining the consistency of low-level tracking performance between simulation and reality.

[0008] In summary, existing technologies face two major challenges: First, hybrid methods based on parameter tuning are limited by the architectural bottleneck of low-frequency control loops, resulting in a performance ceiling; second, for unstructured terrain, there is a lack of a low-level controller that can guarantee consistent Sim-to-Real tracking performance, making it difficult for high-level strategies to achieve zero-sample transfer. Summary of the Invention

[0009] In view of the above, the purpose of this invention is to provide an adaptive navigation method and apparatus for unstructured terrain based on Sim-to-Real transfer, aiming to simultaneously solve two key bottlenecks: the low-frequency control loop architecture bottleneck and Sim-to-Real tracking inconsistency. By introducing a novel terrain-adaptive controller, it can infer and compensate for unmodeled dynamic characteristics online, thereby maintaining consistent trajectory tracking performance in diverse real-world environments. At the architectural level, the dynamic differences between Sim-to-Real and real-world scenarios are isolated to the lower-level control loop, allowing the higher-level parameter tuning strategy to only process abstract geometric information, thus achieving zero-sample transfer of the higher-level strategy without retraining or fine-tuning in the real environment. Ultimately, this results in a high-performance, robust adaptive navigation system that can be directly deployed in complex unstructured terrain.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides an unstructured terrain adaptive navigation method based on Sim-to-Real migration, comprising the following steps: The high-level strategy of the navigation system is architecturally isolated from the low-level control loop, and a terrain-adaptive controller is introduced into the low-level control loop to compensate for the dynamic differences between the simulation and the real environment. The first stage of training for the terrain adaptive controller is carried out in a simulation environment: the ground truth physical parameters of the randomized terrain are encoded using an environmental encoder to generate environmental dynamic features, and the basic control strategy is trained based on the environmental dynamic features to output robot control commands. The second stage of training for the terrain adaptive controller is carried out in the simulation environment: the robot state-action interaction history data generated in the first stage of training is collected, and the environment predictor is trained based on the interaction history data to predict the corresponding environmental dynamic characteristics. Deploying the navigation system in a real environment: After the environmental dynamics characteristics are inferred in real time by the trained environment predictor, they are input into the trained basic control strategy and the final control command of the real robot is output to achieve adaptive navigation for unstructured terrain.

[0011] Preferably, the navigation system includes a high-level strategy, a mid-level planner, and a low-level control loop. The high-level strategy makes decisions based on environmental geometry information, the mid-level planner generates an executable reference trajectory based on the decisions, and the low-level control loop uses a terrain-adaptive controller to predict and output control commands for the real robot on the reference trajectory.

[0012] Preferably, the high-level strategy and the mid-level planner maintain functional consistency in the simulation environment and the real environment, and support zero-sample migration when the navigation system is deployed and migrated to the real environment.

[0013] Preferably, in the first stage of training, the encoded features of the normal state, the robot's action in the previous frame, and the environmental dynamic features generated by the environmental encoder are fused and input into the basic control strategy. The basic control strategy is then trained through reinforcement learning so that it can generate motion control commands for the robot when informed of the current environmental dynamic features.

[0014] Preferably, the physical parameters for field randomization include at least the ground friction coefficient, robot load, and center of mass position.

[0015] Preferably, in the second stage of training, the robot state-action interaction history data generated in the first stage of training is used as the input of the environment predictor, and the environmental dynamic features generated by the environment encoder in the first stage of training are used as supervision labels. The environment predictor is trained through supervised learning so that its output predicted environmental dynamic features are aligned with the supervision labels.

[0016] Preferably, when deploying the navigation system in a real environment, the environment encoder is discarded, and the environment predictor trained in the simulation environment is combined with the basic control strategy to form a terrain adaptive controller in the underlying control loop. Without being informed of the current environmental dynamics characteristics, the encoded features of the normal state, the robot's previous frame action, and the environmental dynamics features generated by the trained environment predictor are fused and input into the trained basic control strategy to generate the robot's motion control commands.

[0017] Secondly, embodiments of the present invention also provide an unstructured terrain adaptive navigation device based on Sim-to-Real migration, which is implemented using the above-mentioned unstructured terrain adaptive navigation method based on Sim-to-Real migration, including: an architecture isolation module, a first simulation training module, a second simulation training module, and a real environment deployment module; The architecture isolation module is used to isolate the high-level strategy of the navigation system from the low-level control loop, and introduces a terrain adaptive controller into the low-level control loop to compensate for the dynamic differences between the simulation and the real environment. The first simulation training module is used to perform the first stage training of the terrain adaptive controller in a simulation environment: using an environmental encoder to encode the ground truth physical parameters of the randomized domain, generating environmental dynamic features, and training the basic control strategy based on the environmental dynamic features to output robot control commands; The second simulation training module is used to perform the second stage training of the terrain adaptive controller in a simulation environment: collect the robot state-action interaction history data generated in the first stage training, and train the environment predictor based on the interaction history data to predict the corresponding environmental dynamics characteristics. The real environment deployment module is used to deploy the navigation system in a real environment: after the environmental dynamics characteristics are inferred in real time by the trained environment predictor, they are input into the trained basic control strategy and the final control command of the real robot is output to realize adaptive navigation for unstructured terrain.

[0018] Thirdly, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described unstructured terrain adaptive navigation method based on Sim-to-Real migration when the computer program is executed.

[0019] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a computer, implements the above-described unstructured terrain adaptive navigation method based on Sim-to-Real migration.

[0020] Compared with the prior art, the beneficial effects of the present invention include at least the following: (1) By constructing an architectural isolation scheme between the high-level strategy and the low-level terrain adaptive controller, the present invention restricts the dynamic differences between the simulation and the real environment to be processed in the low-level control loop, so that the high-level strategy only needs to process abstract geometric information, realizes zero-sample transfer of the high-level strategy, and does not need to be retrained or fine-tuned in the real environment, thus fundamentally ensuring the strong versatility and portability of the high-level navigation module.

[0021] (2) This invention designs a terrain adaptive controller that includes a two-stage training process, so that when it is deployed in real time, it does not need to rely on the unknown real dynamic parameters in the environment. Instead, it can infer the hidden environmental dynamic characteristics online by analyzing the interaction history between the robot and the environment, so as to achieve online accurate perception and adaptive compensation.

[0022] (3) Based on the trained terrain adaptive controller, the present invention provides an adaptive navigation mechanism, thereby solving the problem that when the navigation strategy trained in the simulation is deployed to the real world, the performance will drop sharply or the task will fail because it cannot accurately model the complex physical and dynamic characteristics of terrain friction, slope, load and so on. It significantly improves the trajectory tracking accuracy and robustness on unstructured terrains such as grass, gravel road and earth slope. Attached Figure Description

[0023] 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.

[0024] Figure 1 This is a flowchart illustrating the unstructured terrain adaptive navigation method based on Sim-to-Real migration provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the framework of the unstructured terrain adaptive navigation method based on Sim-to-Real migration provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the unstructured terrain adaptive navigation device based on Sim-to-Real migration provided in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0026] The inventive concept of this invention is as follows: Addressing the limitations of existing methods due to low-frequency control loop performance bottlenecks, and the resulting degradation in tracking performance and inability to achieve zero-sample transfer due to differences between simulated and real-world dynamics, this invention provides an unstructured terrain-adaptive navigation method and apparatus based on Sim-to-Real transfer. By architecturally isolating dynamic uncertainties to the underlying control loop, and introducing a terrain-adaptive controller trained through two-stage simulation, this controller can infer environmental dynamic characteristics in real time based on online interaction history and actively compensate for them. This maintains high-precision, adaptive trajectory tracking performance in both simulated and diverse real-world off-road environments, ultimately achieving zero-sample transfer of high-level navigation strategies and efficient and robust autonomous navigation for the entire system.

[0027] like Figure 1 and Figure 2 As shown in the embodiment, an unstructured terrain adaptive navigation method based on Sim-to-Real migration is provided. Its core includes an architectural isolation strategy and a two-stage training and deployment process for a terrain adaptive controller, specifically including the following steps: S1 isolates the high-level strategy of the navigation system from the low-level control loop in the architecture, and introduces a terrain-adaptive controller in the low-level control loop to compensate for the dynamic differences between the simulation and the real environment.

[0028] In this embodiment, the navigation system includes a high-level strategy, a mid-level planner, and a low-level control loop. To achieve adaptive navigation and zero-sample transfer in unstructured terrain, the system architecture isolates the challenges of Sim-to-Real dynamic uncertainties within the low-level control loop.

[0029] High-level strategy (parameter tuning): The high-level parameter tuning strategy is designed to rely solely on abstract geometric information. Its state-space input is a local elevation map (rather than physical parameters), which is encoded into a low-dimensional latent space feature vector through a variational autoencoder (VAE) network. Since the high-level strategy's decisions are based solely on geometric information and are decoupled from physical dynamics, models trained in simulations can be directly transferred to real-world scenarios.

[0030] Mid-level planners: Mid-level model planners (such as CAP or Uneven Planner) also generate trajectories based on geometric and kinematic principles, and their functionality remains consistent in simulation and reality.

[0031] Low-level control loop: In the low-level control loop, a terrain-adaptive controller is introduced. Its goal is to actively compensate for the dynamic differences caused by architectural isolation, maintaining consistent, low-error trajectory tracking performance across various unstructured terrains (such as simulation terrain, grass, gravel roads, etc.). By performing a two-stage training process in a simulation environment, it is possible to predict and output control commands for the real robot on a reference trajectory without relying on real-world dynamic parameters.

[0032] This architectural isolation strategy, along with the support of a terrain-adaptive controller, enables high-level policies and mid-level planners to maintain functional consistency in both simulation and real-world environments, achieving zero-sample migration when the navigation system is deployed to a real-world environment.

[0033] S2, the terrain adaptive controller is trained in the simulation environment in the first stage: the environmental encoder is used to encode the ground truth physical parameters of the randomized terrain to generate environmental dynamic features, and the basic control strategy is trained based on the environmental dynamic features to output robot control commands.

[0034] In the embodiment, the first stage trains the basic control strategy, aiming to train a basic controller that can adapt to different dynamic environments.

[0035] Specifically, such as Figure 2 As shown, the coded features obtained by using a multi-mode encoder for the normal states (including local map, planned trajectory, robot state, control step size, etc.) and the robot's previous frame action are analyzed. And environmental dynamic features generated by the environmental encoder based on domain-randomized physical parameters (including a wide range of domain-randomized physical dynamic parameters obtained by changing the ground friction coefficient, robot load, and center of mass position). (As privileged information), after being fused by a multilayer perceptron (MLP), it is input into the basic control policy. The basic control policy is then trained through reinforcement learning (such as the PPO algorithm) so that it learns which optimal control action should be performed given the current environmental dynamics characteristics (represented by latent vectors). To accurately track the trajectory.

[0036] S3. The second stage of training of the terrain adaptive controller is carried out in the simulation environment: the robot state-action interaction history data generated in the first stage of training is collected, and the environment predictor is trained based on the interaction history data to predict the corresponding environmental dynamic characteristics.

[0037] In this embodiment, the second stage trains the environment predictor by training only one network to replace the environment encoder that requires privileged information.

[0038] Specifically, such as Figure 2 As shown, the robot interaction history data generated in the first stage of training is used as input to the environment predictor, and the environmental dynamics features generated by the environment encoder in the first stage of training are used as input. As a supervisory label, the environment predictor is trained through supervised learning, enabling it to output predicted environmental dynamics features. Align with the supervision labels. This training enables the environment predictor to learn to indirectly infer underlying environmental dynamics from the robot's observable motion performance data (such as the dynamic response when slipping or climbing).

[0039] S4. Deploy the navigation system in a real environment: After the environmental predictor is trained, it infers the dynamic characteristics of the environment in real time, inputs them into the trained basic control strategy, and outputs the final control command of the real robot to achieve adaptive navigation for unstructured terrain.

[0040] In this embodiment, by deploying the simulation-trained navigation system to a real environment, the aim is to achieve zero-sample migration of the navigation system from the simulation environment to real unstructured terrain, in which the high-level strategy and the mid-level planner can run stably without any adjustment.

[0041] Specifically, such as Figure 2 As shown, when deployed on a real robot, the environmental encoder is discarded, and the encoded features obtained from the normal state through a multi-mode encoder and the robot's previous frame action are used instead. And the predicted environmental dynamics features generated by the trained environment predictor (from the second stage) based on interaction history data. After MLP fusion, the trained basic control strategy (from the first stage) is input to generate the robot's next action. Control commands.

[0042] During online adaptive operation, the environment predictor continuously observes recent interaction history data between the robot and real terrain (such as grass or gravel roads), and infers predicted environmental dynamics characteristics representing the current terrain dynamics in real time based on this interaction history data. This predicted environmental dynamics characteristics are then passed to the basic control strategy as input. Based on these predicted environmental dynamics characteristics, the basic control strategy automatically invokes control modes learned in simulation that match the current real terrain, outputting precise compensating motion control commands. In this way, the terrain adaptive controller constructed in this invention can adapt online to unknown real-world dynamics, maintaining high-fidelity trajectory tracking performance (e.g., extremely low and consistent errors on both grass and gravel roads), thereby ensuring the successful zero-sample transfer of the high-level planning strategy.

[0043] In summary, the unstructured terrain adaptive navigation method based on Sim-to-Real migration provided in this invention demonstrates significant advantages in navigation performance, trajectory tracking accuracy, and adaptive behavior capabilities, as verified by experiments. Specific results are as follows: (1) Robust zero-sample Sim-to-Real transfer and navigation performance improvement: This invention solves the core problem of simulation strategies failing due to dynamic differences when deployed to the real world. In real complex off-road scenarios (such as grasslands, slopes, etc.), the baseline planner using default parameters failed in 2 out of 3 cases, while this invention (trained only in simulation) successfully completed all cases. In the only successful case of the baseline, the passage time of this invention (13.8 seconds) was also nearly twice as fast as the baseline (27.1 seconds), demonstrating excellent real-world robustness and efficiency.

[0044] (2) Consistent and high-precision trajectory tracking is achieved: The terrain-adaptive controller of this invention not only performs best in simulations (e.g., achieving 100% success rate in the case study, compared to only 80% for the baseline), but more importantly, it bridges the tracking error gap in Sim-to-Real testing. When testing preset trajectories on real gravel roads and grass, the tracking error of the terrain-adaptive controller of this invention remains at an extremely low level (e.g., the error of the "figure-eight" trajectory is approximately 0.25 meters on both types of surfaces), and its performance remains consistent. In contrast, the error of traditional controllers is as high as 2-5 times, and their performance is extremely unstable across different road surfaces.

[0045] (3) Demonstrating intelligent adaptive navigation behavior in the real world: Due to the successful transfer of high-level strategies, the entire system exhibits intelligent, context-aware navigation capabilities in real-world environments. For example, in real off-road testing, during the first half (0-13 seconds) of obstacle avoidance, the system automatically selects a lower maximum linear velocity and a higher angular velocity for precise maneuvering; after passing through the obstacle area and entering an open slope, it immediately switches to a more aggressive strategy, increasing the linear velocity for rapid passage. This demonstrates the practical ability of this invention to balance safety and efficiency in the real world.

[0046] Based on the same inventive concept, such as Figure 3 As shown, this embodiment of the invention also provides an unstructured terrain adaptive navigation device 300 based on Sim-to-Real migration, including: an architecture isolation module 310, a first simulation training module 320, a second simulation training module 330, and a real environment deployment module 340.

[0047] The architecture isolation module 310 is used to architecturally isolate the high-level strategy of the navigation system from the low-level control loop, and introduces a terrain adaptive controller in the low-level control loop to compensate for the dynamic differences between the simulation and the real environment.

[0048] The first simulation training module 320 is used to perform the first stage training of the terrain adaptive controller in the simulation environment: the environmental encoder is used to encode the ground truth physical parameters of the randomized domain to generate environmental dynamic features, and the basic control strategy is trained based on the environmental dynamic features to output robot control commands.

[0049] The second simulation training module 330 is used to perform the second stage training of the terrain adaptive controller in the simulation environment: it collects the robot state-action interaction history data generated in the first stage training, and trains the environment predictor based on the interaction history data to predict the corresponding environmental dynamic characteristics.

[0050] The real environment deployment module 340 is used to deploy the navigation system in a real environment: after the environmental dynamics characteristics are inferred in real time by the trained environment predictor, they are input into the trained basic control strategy and the final control command of the real robot is output to realize adaptive navigation for unstructured terrain.

[0051] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory and one or more processors, wherein the memory is used to store a computer program, and the processor is used to implement the above-described unstructured terrain adaptive navigation method based on Sim-to-Real migration when executing the computer program.

[0052] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program, which, when executed by a computer, implements the above-described unstructured terrain adaptive navigation method based on Sim-to-Real migration.

[0053] It should be noted that the unstructured terrain adaptive navigation device, electronic device, and computer-readable storage medium based on Sim-to-Real migration provided in the above embodiments all belong to the same inventive concept as the unstructured terrain adaptive navigation method based on Sim-to-Real migration. For details of its specific implementation process, please refer to the embodiments of the unstructured terrain adaptive navigation method based on Sim-to-Real migration, which will not be repeated here.

[0054] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An unstructured terrain adaptive navigation method based on Sim-to-Real migration, characterized in that, Includes the following steps: The high-level strategy of the navigation system is architecturally isolated from the low-level control loop, and a terrain-adaptive controller is introduced into the low-level control loop to compensate for the dynamic differences between the simulation and the real environment. The first stage of training for the terrain adaptive controller is carried out in a simulation environment: the ground truth physical parameters of the randomized terrain are encoded using an environmental encoder to generate environmental dynamic features, and the basic control strategy is trained based on the environmental dynamic features to output robot control commands. The second stage of training for the terrain adaptive controller is carried out in the simulation environment: the robot state-action interaction history data generated in the first stage of training is collected, and the environment predictor is trained based on the interaction history data to predict the corresponding environmental dynamic characteristics. Deploying the navigation system in a real environment: After the environmental dynamics characteristics are inferred in real time by the trained environment predictor, they are input into the trained basic control strategy and the final control command of the real robot is output to achieve adaptive navigation for unstructured terrain.

2. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 1, characterized in that, The navigation system includes a high-level strategy, a mid-level planner, and a low-level control loop. The high-level strategy makes decisions based on environmental geometry information, the mid-level planner generates an executable reference trajectory based on the decisions, and the low-level control loop uses a terrain-adaptive controller to predict and output control commands for the real robot on the reference trajectory.

3. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 2, characterized in that, The high-level strategy and mid-level planner maintain functional consistency in both simulation and real-world environments, and support zero-shot migration when the navigation system is deployed and migrated to a real-world environment.

4. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 1, characterized in that, In the first stage of training, the encoded features of the normal state, the robot's previous frame action, and the environmental dynamic features generated by the environmental encoder are fused and input into the basic control strategy. The basic control strategy is then trained through reinforcement learning so that it can generate motion control commands for the robot when informed of the current environmental dynamic features.

5. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 1, characterized in that, The physical parameters for domain randomization include at least the ground friction coefficient, robot load, and center of mass position.

6. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 1, characterized in that, In the second phase of training, the robot state-action interaction history data generated in the first phase of training is used as the input of the environment predictor, and the environmental dynamic features generated by the environment encoder in the first phase of training are used as supervision labels. The environment predictor is trained through supervised learning so that its output predicted environmental dynamic features are aligned with the supervision labels.

7. The unstructured terrain adaptive navigation method based on Sim-to-Real migration according to claim 1, characterized in that, When deploying the navigation system in a real environment, the environment encoder is discarded. Instead, the environment predictor trained in the simulation environment is combined with the basic control strategy to form the terrain adaptive controller in the underlying control loop. Without being informed of the current environmental dynamics, the encoded features of the normal state, the robot's previous frame action, and the environmental dynamics features generated by the trained environment predictor are fused and input into the trained basic control strategy to generate the robot's motion control commands.

8. An unstructured terrain adaptive navigation device based on Sim-to-Real migration, implemented using the unstructured terrain adaptive navigation method based on Sim-to-Real migration as described in any one of claims 1 to 7, characterized in that, include: Architecture isolation module, first simulation training module, second simulation training module, and real environment deployment module; The architecture isolation module is used to isolate the high-level strategy of the navigation system from the low-level control loop, and introduces a terrain adaptive controller into the low-level control loop to compensate for the dynamic differences between the simulation and the real environment. The first simulation training module is used to perform the first stage training of the terrain adaptive controller in a simulation environment: using an environmental encoder to encode the ground truth physical parameters of the randomized domain, generating environmental dynamic features, and training the basic control strategy based on the environmental dynamic features to output robot control commands; The second simulation training module is used to perform the second stage training of the terrain adaptive controller in a simulation environment: collect the robot state-action interaction history data generated in the first stage training, and train the environment predictor based on the interaction history data to predict the corresponding environmental dynamics characteristics. The real environment deployment module is used to deploy the navigation system in a real environment: after the environmental dynamics characteristics are inferred in real time by the trained environment predictor, they are input into the trained basic control strategy and the final control command of the real robot is output to realize adaptive navigation for unstructured terrain.

9. An electronic device comprising a memory and one or more processors, the memory for storing a computer program, characterized in that, The processor is configured to implement the unstructured terrain adaptive navigation method based on Sim-to-Real migration as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by a computer, it implements the unstructured terrain adaptive navigation method based on Sim-to-Real migration as described in any one of claims 1 to 7.

Images