Unstructured terrain adaptive navigation method and device based on Simto-Real migration

Abstract

The invention discloses an unstructured terrain adaptive navigation method and device based on Simto-Real migration, and the method comprises the steps: carrying out the architecture isolation of a high-level strategy and a bottom-layer control loop, and introducing a terrain adaptive controller into the bottom-layer control loop; the method comprises the following steps: in a simulation environment, performing first-stage training on a basic control strategy in a terrain adaptive controller by utilizing environment dynamic characteristics generated by encoding field randomized ground true value physical parameters; performing second-stage training on an environment predictor in the terrain adaptive controller by utilizing interaction historical data generated in the first-stage training; navigation deployment is carried out in a real environment, environment dynamic characteristics are deduced in real time through the trained environment predictor, a trained basic control strategy is input, and a final robot control instruction is output. According to the method, zero sample migration of a high-level strategy during real environment deployment is realized, and the method is suitable for high-precision and robust adaptive navigation of unstructured terrains.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robot autonomous navigation, and particularly relates to a Sim-to-Real migration-based unstructured terrain adaptive navigation method and device. BACKGROUND

[0002] It is always a long-term challenge in the field of field robots to achieve autonomous navigation in complex and unstructured environments. Existing navigation technologies are mainly divided into three categories, but all have significant technical bottlenecks.

[0003] The first category is the End-to-End learning method, which attempts to directly learn the mapping from raw sensor inputs to vehicle control commands. Among them, the End-to-End method based on imitation learning (IL) performs well in structured scenarios such as autonomous driving, but its core defect lies in its heavy reliance on large-scale and high-quality expert demonstration datasets. For diversified and dynamic robot tasks, the cost of collecting such datasets is high, which constitutes a major obstacle to actual deployment. The End-to-End method based on reinforcement learning (RL) does not require expert data and can learn autonomously through a trial-and-error mechanism. However, in mobile robot navigation tasks, especially when facing long-term and sparse reward signals, this method generally has the problems of low sample efficiency and unstable training. Even in some benchmark tests, the traditional method with careful parameter tuning can still outperform the End-to-End RL strategy in performance.

[0004] The second category is the traditional trajectory planning method, such as Time Elastic Band (TEB) and Dynamic Window Approach (DWA) and the like. This method is based on a model and has high robustness and interpretability. However, its performance is highly dependent on a set of key parameters (such as maximum speed, inflation radius, etc.) that require expert experience for fine-tuning, and lacks the ability to adapt to dynamic environments.

[0005] The third category is the hybrid method based on learning, i.e. automatic parameter tuning, which is the most promising compromise at present. The core idea is to use machine learning (especially RL) to dynamically adjust the parameters of the traditional planner, thereby combining the robustness of model planning and the adaptability of learning. Since the planner limits the exploration range to the high-level parameter space and shields unsafe low-level actions, this method significantly improves the sample efficiency and system safety. Methods such as APPLR and APPLE are typical representatives in this direction. However, despite the progress made by this hybrid method, its performance cannot match that of high-performance End-to-End imitation learning systems. The present application believes that this performance gap is not due to algorithm design, but due to a long-neglected architectural limitation.

[0006] Specifically, the key of this architecture limitation lies in the control frequency. The end-to-end IL policy, once trained, can directly output velocity commands at a very high frequency (e.g., 50 Hz) to achieve a sensitive response. In contrast, the control path of the auto-tuning method is indirect: the learning policy adjusts parameters, the planner has to re-optimize or sample to generate trajectories, and finally the control commands are generated. This additional computational delay limits the frequency of the entire control loop to a relatively low level (e.g., 5-10 Hz). This low-frequency, high-delay control loop causes the performance bottleneck of existing parameter tuning methods.

[0007] In addition, when the navigation task is extended to unstructured terrain, all the above methods face a second serious challenge: the huge gap between simulation and reality (Sim-to-Real). The controller trained in simulation (whether end-to-end or bottom-level control in hybrid methods) assumes a low-variance tracking error, but when deployed to the real world, the variance of the tracking error increases significantly due to unmodeled dynamics (such as friction, gravity, load changes). For hybrid methods, this variance gap is fatal, and the decline in the performance quality of the bottom-level controller (i.e., the increase in variance) directly reduces the optimal value function that the high-level tuning policy can achieve. Therefore, the limitations of existing Sim-to-Real solutions are that common controllers (such as PID, MPC, or some adaptive controllers) either rely on high-precision models or only improve tracking performance in isolation, and none of them can solve the core problem of how to guarantee the zero-sample transfer ability of the high-level tuning policy by keeping the bottom-level tracking performance consistent between simulation and reality.

[0008] In summary, the existing technology faces two core problems: first, the hybrid method based on parameter tuning is limited by the architecture bottleneck of the low-frequency control loop, and the performance has an upper limit; second, for unstructured terrain, there is a lack of a bottom-level controller that can guarantee the consistency of Sim-to-Real tracking performance, making it difficult for the high-level strategy to achieve zero-sample transfer. SUMMARY

[0009] In view of the above, the purpose of the present application is to provide a Sim-to-Real migration based unstructured terrain adaptive navigation method and device, aiming to solve the two key bottlenecks of low-frequency control loop architecture bottleneck and Sim-to-Real tracking inconsistency. By introducing a new type of terrain adaptive controller, it can infer and compensate the unmodeled dynamic characteristics online, so as to maintain consistent trajectory tracking performance in diversified real environment. At the architecture level, the Sim-to-Real dynamic difference is isolated to the bottom control loop, so that the high-level parameter tuning strategy only needs to deal with abstract geometric information, thereby realizing zero-sample migration of high-level strategy without retraining or fine-tuning in real environment. Finally, a set of high-performance, high-robustness and directly deployable adaptive navigation system on complex unstructured terrain is formed.

[0010] To achieve the above-mentioned purpose of the application, the technical solutions provided by the present application are as follows: In the first aspect, the present application provides a Sim-to-Real migration based unstructured terrain adaptive navigation method, comprising the following steps: The high-level strategy of the navigation system is isolated from the bottom control loop, and a terrain adaptive controller is introduced in the bottom control loop to compensate for the dynamic difference between simulation and real environment; First stage training of the terrain adaptive controller in the simulation environment: using the environment encoder to encode the field randomization ground truth physical parameters to generate the environment dynamic characteristics, and training the basic control strategy based on the environment dynamic characteristics to output the robot control instruction; Second stage training of the terrain adaptive controller in the simulation environment: collecting the robot state-action interaction history data generated in the first stage training, and training the environment predictor based on the interaction history data to predict the corresponding environment dynamic characteristics; Navigation system deployment in real environment: after the trained environment predictor infers the environment dynamic characteristics in real time, the final control instruction of the real robot is input into the trained basic control strategy to realize adaptive navigation for unstructured terrain.

[0011] Preferably, the navigation system includes high-level strategy, middle-level planner and bottom control loop, wherein the high-level strategy makes decisions according to the environment geometric information, the middle-level planner generates executable reference trajectory according to the decisions, and the bottom control loop predicts and outputs the control instruction of the real robot on the reference trajectory by using the terrain adaptive controller.

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

[0013] Preferably, in the first stage training, the encoded features of the regular state, the last frame action of the robot, and the environment dynamics features generated by the environment encoder are fused and input into the base control policy, and the base control policy is trained through reinforcement learning to generate the action control instructions of the robot under the condition of being informed of the current environment dynamics features.

[0014] Preferably, the domain-randomized physical parameters at least include the ground friction coefficient, the robot load, and the center of mass position.

[0015] Preferably, in the second stage training, the robot state-action interaction history data generated in the first stage training is taken as the input of the environment predictor, and the environment dynamics features generated by the environment encoder in the first stage training are taken as the supervision labels, and the environment predictor is trained through supervised learning to align the predicted environment dynamics features output by the environment predictor with the supervision labels.

[0016] Preferably, when the navigation system is deployed in the real environment, the environment encoder is discarded, the environment predictor trained in the simulation environment is combined with the base control policy to form a terrain adaptive controller in the bottom control loop, and under the condition of not being informed of the current environment dynamics features, the action control instructions of the robot are generated by inputting the encoded features of the regular state, the last frame action of the robot, and the environment dynamics features generated by the trained environment predictor into the trained base control policy after being fused.

[0017] In a second aspect, the embodiments of the present application also provide a Sim-to-Real migration based unstructured terrain adaptive navigation device, which is implemented by using the above-mentioned Sim-to-Real migration based unstructured terrain adaptive navigation method, and includes 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 configured to isolate the high-level policy of the navigation system from the bottom control loop, and introduce a terrain adaptive controller in the bottom control loop to compensate for the dynamics difference between the simulation and the real environment. The first simulation training module is configured to perform first stage training of the terrain adaptive controller in the simulation environment: encode the domain-randomized ground truth physical parameters by using an environment encoder to generate environment dynamics features, and train a base control policy based on the environment dynamics features to output robot control instructions. The second simulation training module is configured to perform second stage training of the terrain adaptive controller in the simulation environment: collect the robot state-action interaction history data generated in the first stage training, and train an environment predictor based on the interaction history data to predict corresponding environment dynamics features. The real environment deployment module is used for navigation system deployment in a real environment: after real-time inference of environmental dynamic characteristics by the trained environment predictor, the trained basic control strategy is input and the final control instruction of the real robot is output, so that adaptive navigation facing unstructured terrain is realized.

[0018] In a third aspect, an embodiment of the present application further provides an electronic device, comprising a memory and one or more processors, the memory is used to store a computer program, and the processor is used to realize the Sim-to-Real migration based adaptive navigation method for unstructured terrain described above when the computer program is executed.

[0019] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, and the storage medium stores a computer program, and the Sim-to-Real migration based adaptive navigation method for unstructured terrain described above is realized when the computer program is executed by a computer.

[0020] Compared with the prior art, the present application has beneficial effects at least including: (1) The present application limits the difference in dynamics between simulation and reality to the bottom control loop by constructing an architectural isolation scheme of high-level strategy and bottom terrain adaptive controller, so that the high-level strategy only needs to process abstract geometric information, realizes zero-sample migration of the high-level strategy, and does not need to be retrained or fine-tuned in the real environment, which fundamentally guarantees the strong universality and migratability of the high-level navigation module.

[0021] (2) The present application designs a terrain adaptive controller containing a two-stage training process, so that it does not need to rely on unknown real dynamics parameters in the environment when deployed in reality, but can infer the hidden environmental dynamics characteristics online by analyzing the interaction history of the robot and the environment, to realize online accurate perception and adaptive compensation.

[0022] (3) The present application provides an adaptive navigation mechanism based on the trained terrain adaptive controller, thereby solving the problem that the navigation strategy trained in simulation cannot accurately model the complex physical dynamics characteristics such as terrain friction, slope and load when deployed to the real world, resulting in a sharp decline in performance or task failure, and significantly improving the trajectory tracking accuracy and robustness on unstructured terrain such as grassland, gravel road and soil slope. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the accompanying drawings required by the embodiments or prior art description will be briefly introduced. Obviously, the accompanying drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0024] Figure 1 is a flowchart of a Sim-to-Real migration based unstructured terrain adaptive navigation method provided by an embodiment of the present application; Figure 2 is a framework diagram of a Sim-to-Real migration based unstructured terrain adaptive navigation method provided by an embodiment of the present application; Figure 3 is a structural diagram of a Sim-to-Real migration based unstructured terrain adaptive navigation device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application 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 only used to explain the present application and do not limit the protection scope of the present application.

[0026] The inventive concept of the present application is that: in order to solve the problems that the existing method is limited by the low-frequency control loop performance bottleneck and the tracking performance is reduced due to the difference between simulation and reality dynamics, and zero-sample migration cannot be realized, an embodiment of the present application provides a Sim-to-Real migration based unstructured terrain adaptive navigation method and device, which limits the dynamic uncertainty processing to the bottom control loop through architectural isolation, and introduces a terrain adaptive controller trained through two-stage simulation, which can infer the environmental dynamics in real time according to the online interaction history and actively compensate, so as to maintain high-precision and adaptive trajectory tracking performance in simulation and diversified real off-road environments, and finally realize zero-sample migration of high-level navigation strategy and efficient and high-robustness autonomous navigation of the whole system.

[0027] As shown in Figure 1 and Figure 2 , an embodiment provides a Sim-to-Real migration based unstructured terrain adaptive navigation method, which includes an architectural isolation strategy, a two-stage training and deployment process of a terrain adaptive controller, and specifically includes the following steps: S1, the high-level strategy of the navigation system is isolated from the bottom control loop, and a terrain adaptive controller is introduced in the bottom control loop to compensate for the dynamics difference between simulation and reality environment.

[0028] In an embodiment, the navigation system includes a high-level strategy, a middle-level planner, and a bottom-level control loop. To achieve adaptive navigation for unstructured terrain and zero-shot transfer, the dynamics uncertainty challenge of Sim-to-Real is isolated in the bottom-level control loop in the system architecture.

[0029] High-level strategy (parameter tuning): The high-level parameter tuning strategy is designed to rely only on abstract geometric information, with a state space input of a local elevation map (rather than physical parameters), which is encoded into a low-dimensional hidden space feature vector by a variational autoencoder (VAE) network. Since the high-level strategy's decision is based only on geometric information, decoupled from physical dynamics, its model trained in simulation can be directly transferred to real-world scenarios.

[0030] Middle-level planner: The middle-level model planner (such as CAP or Uneven Planner) also generates trajectories based on geometric and kinematic principles, and its function remains consistent in simulation and reality.

[0031] Bottom-level control loop: In the bottom-level control loop, a terrain-adaptive controller is introduced, aiming to actively compensate for the architectural isolation of dynamics differences to maintain consistent, low-error trajectory tracking performance on different unstructured terrains (such as simulation, grass, gravel, etc.). Through a two-stage training process in the simulation environment, the final prediction output reference trajectory can ultimately predict the control instructions for the real robot without relying on real-world dynamics parameters.

[0032] Through this architectural isolation strategy and the support of the terrain-adaptive controller, the high-level strategy and the middle-level planner maintain functional consistency in simulation and real environments, enabling zero-shot transfer when the navigation system is deployed to a real environment.

[0033] S2, first-stage training of the terrain-adaptive controller in the simulation environment: The environment encoder is used to encode the ground truth physical parameters of the domain randomization, generating environment dynamics features, and the basic control strategy is trained based on the environment dynamics features to output robot control instructions.

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

[0035] Specifically, as shown in Figure 2 the encoding features obtained by passing the regular state (including local map, planned trajectory, robot state, control step, etc.) through a multi-modal encoder, the robot's last frame of action environment dynamics features generated by the environment encoder according to the field randomized physical parameters (including a wide range of field randomized physical dynamics parameters obtained by changing the ground friction coefficient, robot load and center of mass position, etc.) , as privileged information, are fused through a multi-layer perceptron (MLP), input into the base control policy, and trained through reinforcement learning (such as the PPO algorithm) to make the base control policy learn what optimal control action should be performed under the condition of being informed of the current environment dynamics features (represented by the latent vector) to accurately track the trajectory.

[0036] S3, in the simulation environment, the terrain-adaptive controller is trained in the second stage: 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 environment dynamics features.

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

[0038] Specifically, as shown in Figure 2 , the robot interaction history data generated in the first stage of training is input into the environment predictor, and the environment dynamics features generated by the environment encoder in the first stage of training are used as supervised labels to train the environment predictor through supervised learning, so that the predicted environment dynamics features output by the environment predictor are aligned with the supervised labels. Through this training, the environment predictor can learn to indirectly infer the hidden environment dynamics features from the observable motion performance data of the robot (such as the dynamic response when slipping or climbing).

[0039] S4, navigation system deployment in the real environment: after the trained environment predictor infers the environment dynamics features in real time, the environment dynamics features are input into the trained base control policy and output the final control instructions of the real robot, realizing adaptive navigation for unstructured terrain.

[0040] In the embodiment, the simulation-trained navigation system is deployed in the real environment to realize zero-shot transfer of the navigation system from the simulation environment to the real unstructured terrain, where the high-level strategy and the middle-level planner can be directly and stably operated without any adjustment.

[0041] Specifically, as shown in Figure 2 , when deployed on the real robot, the environment encoder is discarded, and the encoded features obtained by the multi-modal encoder from the regular state, the last action of the robot and the predicted environment dynamics features generated by the trained environment predictor (from the second stage) from the interaction history data After being fused by the MLP, the trained base control policy (from the first stage) is inputted to generate the next action control instruction of the robot .

[0042] When the robot is running online, the environment predictor continuously observes the recent interaction history data of the robot and the real terrain (such as grassland and gravel road), and real-time infers the predicted environment dynamics features representing the current terrain dynamics characteristics according to the interaction history data, and transmits the predicted environment dynamics features to the base control policy as input. The base control policy automatically calls the control mode learned in simulation and matched with the current real terrain according to the predicted environment dynamics features, and outputs accurate compensation action control instructions. In this way, the terrain adaptive controller constructed by the application can adapt to unknown real world dynamics online, maintain high fidelity trajectory tracking performance (such as low and consistent errors on grassland and gravel road), and thus ensure the success of zero-shot transfer of high-level planning strategy.

[0043] In summary, the Sim-to-Real transfer based unstructured terrain adaptive navigation method provided by the embodiment of the application has been verified by experiments to have significant advantages in navigation performance, trajectory tracking accuracy and adaptive behavior ability, and the specific results are as follows: (1) Robust zero-shot Sim-to-Real transfer and navigation performance improvement are achieved: the application solves the core problem that the simulation strategy fails when deployed to the real world due to the difference in dynamics. In real complex off-road scenes (such as grassland and soil slope), the baseline planner with default parameters failed in 2 out of 3 cases, while the application (trained only in simulation) successfully completed all cases. In the only case where the baseline succeeded, the passage time of the application (13.8 seconds) was nearly twice as fast as that of the baseline (27.1 seconds), showing excellent real robustness and efficiency.

[0044] (2) Consistent and high-precision trajectory tracking is achieved: the terrain adaptive controller of the application not only performs best in simulation (for example, it can achieve 100% success rate in the case, while the baseline is only 80%), but more importantly, it solves the tracking error gap of Sim-to-Real. When testing the preset trajectory on real gravel road and grassland, the tracking error of the terrain adaptive controller of the application always remains at a very low level (such as the "8-shaped" trajectory error is about 0.25 meters on both road surfaces), and the performance remains consistent. In contrast, the error of the traditional controller is 2-5 times higher, and the performance is extremely unstable between different road surfaces.

[0045] (3) The adaptive navigation behavior in the real world shows intelligence: due to the successful migration of the high-level strategy, the whole system shows intelligent and context-aware navigation capability in the real environment. For example, in the real off-road test, the system automatically selects a lower maximum linear speed and a higher angular speed in the first half (0-13 seconds) of the obstacle avoidance stage to perform precise steering; after passing through the obstacle area and entering the open slope, it immediately switches to a more aggressive strategy to increase the linear speed to achieve rapid passage. This proves the practical ability of the application to balance safety and efficiency in the real world.

[0046] Based on the same inventive concept, as shown in Figure 3 The embodiment of the application also provides a Sim-to-Real migration based unstructured terrain adaptive navigation device 300, which comprises 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 isolate the high-level strategy of the navigation system from the bottom control loop, and introduce a terrain adaptive controller in the bottom control loop to compensate for the dynamics difference between simulation and reality.

[0048] The first simulation training module 320 is used to perform first-stage training of the terrain adaptive controller in the simulation environment: encode the field randomized ground truth physical parameters by using an environment encoder to generate environment dynamics features, and train the basic control strategy based on the environment dynamics features to output robot control instructions.

[0049] The second simulation training module 330 is used to perform second-stage training of the terrain adaptive controller in the 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 environment dynamics features.

[0050] The real environment deployment module 340 is used to deploy the navigation system in the real environment: after the trained environment predictor infers the environment dynamics features in real time, the environment dynamics features are input to the trained basic control strategy to output the final control instructions of the real robot, and adaptive navigation for unstructured terrain is realized.

[0051] Based on the same inventive concept, the embodiment of the application also provides an electronic device comprising a memory and one or more processors, the memory being used to store a computer program, and the processor being used to implement the above-mentioned Sim-to-Real migration based unstructured terrain adaptive navigation method when executing the computer program.

[0052] Based on the same inventive concept, the embodiment of the present application also provides a computer readable storage medium, and the computer program is stored in the storage medium, and when the computer program is executed by a computer, the Sim-to-Real migration based unstructured terrain adaptive navigation method is realized.

[0053] It should be noted that the Sim-to-Real migration based unstructured terrain adaptive navigation device, the electronic device, and the computer readable storage medium provided by the above embodiment all belong to the same inventive concept as the Sim-to-Real migration based unstructured terrain adaptive navigation method, and the specific implementation process is detailed in the Sim-to-Real migration based unstructured terrain adaptive navigation method embodiment, which will not be repeated here.

[0054] The specific embodiments described above have described the technical solutions and beneficial effects of the present application in detail. It should be understood that the above description is only the most preferred embodiment of the present application and is not intended to limit the present application. Any modification, supplement, and equivalent replacement within the principle range of the present application should be included in the protection scope of the present application.

Claims

1. An unstructured terrain adaptive navigation method based on Sim-to-Real transfer, characterized in that, The method comprises the following steps: The high-level strategy of the navigation system is isolated from the bottom control loop, and a terrain adaptive controller is introduced into the bottom control loop to compensate for the difference in dynamics between simulation and real environment; In the simulation environment, the terrain adaptive controller is trained in the first stage: the environment encoder encodes the field randomized ground truth physical parameters to generate environment dynamics characteristics, and the basic control strategy is trained based on the environment dynamics characteristics to output robot control instructions; In the simulation environment, the terrain adaptive controller is trained in the second stage: the robot state-action interaction history data generated in the first stage training is collected, and the environment predictor is trained based on the interaction history data to predict the corresponding environment dynamics characteristics; In the real environment, the navigation system is deployed: after the trained environment predictor infers the environment dynamics characteristics in real time, it is input into the trained basic control strategy to output the final control instructions of the real robot, realizing adaptive navigation for unstructured terrain.

2. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 1, wherein, The navigation system includes a high-level strategy, a middle-level planner, and a bottom control loop, wherein the high-level strategy makes decisions based on environmental geometric information, the middle-level planner generates executable reference trajectories based on the decisions, and the bottom control loop uses a terrain adaptive controller to predict and output control instructions for the real robot on the reference trajectory.

3. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 2, wherein, The high-level strategy and the middle-level planner maintain functional consistency in the simulation environment and the real environment, supporting zero-shot transfer when the navigation system deployment is migrated to the real environment.

4. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 1, wherein, In the first stage training, the encoding features of the regular state, the last frame of action of the robot, and the environment dynamics characteristics generated by the environment encoder are fused and input into the basic control strategy, and the basic control strategy is trained through reinforcement learning to generate the action control instructions of the robot under the condition of being informed of the current environment dynamics characteristics.

5. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 1, wherein, The field randomized physical parameters include at least the ground friction coefficient, the robot load, and the center of mass position.

6. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 1, wherein, In the second stage training, the robot state-action interaction history data generated in the first stage training is used as the input of the environment predictor, and the environment dynamics characteristics generated by the environment encoder in the first stage training are used as the supervision label, and the environment predictor is trained through supervised learning to align the predicted environment dynamics characteristics with the supervision label.

7. The Sim-to-Real transfer based unstructured terrain adaptive navigation method according to claim 1, wherein, When deploying the navigation system in the real environment, the environment encoder is discarded, and the trained environment predictor and the basic control strategy in the simulation environment are combined as the terrain adaptive controller in the bottom control loop, which generates the action control instructions of the robot by fusing the encoding features of the regular state, the last frame of action of the robot, and the environment dynamics characteristics generated by the trained environment predictor and inputting them into the trained basic control strategy without being informed of the current environment dynamics characteristics.

8. An unstructured terrain adaptive navigation device based on Sim-to-Real transfer, implemented by the unstructured terrain adaptive navigation method based on Sim-to-Real transfer of any one of claims 1-7. It comprises: 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 configured to isolate high-level strategies of the navigation system from low-level control loops, and introduce a terrain adaptive controller in the low-level control loops to compensate for the difference in dynamics between simulation and real environment. The first simulation training module is configured to perform first-stage training of the terrain adaptive controller in a simulation environment: encode domain-randomized ground truth physical parameters by using an environment encoder to generate environment dynamics features, and train a base control strategy based on the environment dynamics features to output robot control instructions. The second simulation training module is configured to perform second-stage training of the terrain adaptive controller in the simulation environment: collect robot state-action interaction history data generated in the first-stage training, and train an environment predictor based on the interaction history data to predict corresponding environment dynamics features. The real environment deployment module is configured to perform navigation system deployment in a real environment: after the trained environment predictor infers environment dynamics features in real time, the environment dynamics features are input into the trained base control strategy to output final control instructions for a real robot, thereby realizing adaptive navigation for unstructured terrain.

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

10. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a computer, implements the Sim-to-Real migration based unstructured terrain adaptive navigation method of any one of claims 1-7.

Images