Robot navigation method and system based on semantic perception and multi-level cost map

Abstract

This invention discloses a robot navigation method and system based on semantic perception and a multi-level cost map. This invention overcomes the limitations of traditional path planning techniques that rely solely on geometric obstacle information while ignoring road surface physical properties. It utilizes visual semantic segmentation technology to deeply perceive the material properties of the road surface, constructs a multi-level cost map that integrates physical attribute weights, and quantifies the road surface "traffic risk" as a resistance value for the search algorithm. This achieves an intelligent trade-off between path length and traffic safety, as well as adaptive speed control, significantly improving the operational safety and robustness of mobile robots in complex, unstructured environments. It realizes end-to-end intelligent decision-making from "environmental perception" to "motion control." Not only can it automatically learn human-like driving strategies such as "slowing down" or "detouring" when encountering low-adhesion or high-vibration areas, but it also significantly improves the robot's adaptability and navigation safety in unknown, unstructured environments.

Description

Technical Field

[0001] This invention belongs to the fields of mobile robot autonomous navigation technology, computer vision applications and intelligent control, and particularly relates to a robot navigation method and system based on semantic perception and multi-level cost maps. Background Technology

[0002] As the application scenarios of mobile robots (especially outdoor inspection vehicles) become increasingly complex, the demand for navigation in unstructured environments is growing rapidly. Adaptive path planning can improve the efficiency and success rate of robot operations, reduce the frequency of manual intervention, reduce hardware wear and safety accidents caused by terrain factors, and provide basic support for the stable operation of robots around the clock.

[0003] Existing mobile robot path planning methods typically involve first constructing an environmental geometry map using sensors and marking obstacles, then planning a collision-free shortest path based solely on spatial location information. These algorithms all rely on geometric connectivity as a prerequisite for path planning; however, on complex surfaces, issues such as missing surface material identification and neglecting physical properties often arise. Furthermore, slippage, severe vibrations, and wheel spin caused by traversing water, oil, or gravel surfaces can significantly impact the localization system. Finally, the diversity of environments means that a single distance-optimal rule is not applicable in many dangerous scenarios. All these problems lead to insufficient safety and robustness in such algorithms. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a robot navigation method and system based on semantic perception and multi-level cost maps, so as to solve the problems disclosed in the background art that traditional path planning ignores the physical properties of the road surface and that traditional search algorithms are difficult to achieve end-to-end intelligent decision-making in complex, dynamic and unstructured environments.

[0005] Technical solution: The robot navigation method and system based on semantic perception and multi-level cost map of the present invention includes the following steps:

[0006] Step 1: Collect environmental images of the mobile robot's direction of travel using the vehicle-mounted vision sensor and perform preprocessing.

[0007] Step 2: Use a deep learning-based semantic segmentation network to process the preprocessed image, identify the road surface material category, and extract the road surface semantic features;

[0008] Step 3: Construct a multi-level cost map containing physical attribute weights; the multi-level cost map is generated by mapping the physical attribute weights corresponding to the identified road surface material categories to a geometric grid map, and is used to quantify the cost of road passage.

[0009] Step 4: Construct a path planning model based on deep reinforcement learning, transform the multi-level cost map into a state space, and output the optimal action instructions for controlling the mobile robot.

[0010] Step 5: Generate adaptive motion control commands based on the optimal action command and the cost level corresponding to the current path point to control the mobile robot's movement.

[0011] Further, step 1 specifically involves: using the calibrated camera intrinsic parameter matrix to perform distortion correction on the acquired original image to eliminate lens distortion; normalizing the size of the distortion-corrected image to adjust it to the standard input size required by the semantic segmentation network; and using histogram equalization or adaptive gamma correction algorithms to enhance the image and improve image contrast.

[0012] Furthermore, step 2 specifically involves: employing a semantic segmentation network based on the DeepLab architecture, utilizing atrous convolution to expand the receptive field and extract multi-scale features; in the convolutional layers of the semantic segmentation network, the processing formula for the input feature map is as follows:

[0013]

[0014] In the formula, To extract multi-scale road surface semantic features, The input is the preprocessed environmental image features, where i represents the spatial pixel coordinates on the environmental image acquired by the mobile robot's onboard vision sensor. For convolution kernel weights, Here, k is the kernel size, and k is the kernel stride index. This corresponds to the Dilation Rate, which is the hole rate of the DeepLab network architecture.

[0015] Furthermore, step 3 specifically involves: establishing a multi-level cost map that includes physical attribute weights, including:

[0016] A pre-built material attribute database is used, containing friction coefficients and vibration risk coefficients corresponding to different road surface material categories. A weighted fusion formula is used to superimpose the semantic attribute layer, geometric obstacle layer, and expansion risk layer to calculate the comprehensive passage cost for each grid cell. The weighted fusion formula is as follows:

[0017]

[0018] In the formula, cost(x, y) is the comprehensive passage cost for each grid cell, (x, y) are the two-dimensional grid coordinates in the multi-level cost map, and c obs (x, y) represents the cost of the geometric barrier layer at that coordinate. The semantic layer's cost at that coordinate. The value represents the inflation cost, and α, β, and γ are the fusion weight coefficients for each layer.

[0019] Furthermore, step 4 specifically involves constructing a path planning model based on deep reinforcement learning, including:

[0020] Define a state space, which includes a local multi-level cost map centered on the mobile robot, the robot's current velocity vector, and the relative coordinates of the target point.

[0021] Define the Action Space, which includes the combination of linear velocity and angular velocity of the mobile robot;

[0022] Define a reward function, which includes a reward for reaching the target, a collision penalty, and a passage penalty based on the road surface material.

[0023] Furthermore, the calculation formula for the reward function includes a semantic material penalty term, the specific formula of which is:

[0024]

[0025] In the formula, This is a semantic material penalty item. The penalty coefficient is... This is the robot's current location. For the robot at its current location The corresponding comprehensive transit value.

[0026] Furthermore, in step 5, an adaptive motion control command is generated based on the optimal action command and the cost level corresponding to the current path point, including:

[0027] The average road surface cost along the robot's planned path is detected. If the average road surface cost exceeds a preset safety threshold, the linear velocity in the optimal action command is clamped using a velocity correction formula. The velocity correction formula is:

[0028]

[0029] In the formula For the final control command, The original linear velocity output by the deep reinforcement learning model. For average road surface cost, As a safety threshold, is the velocity decay coefficient, and e is the base of the natural logarithm.

[0030] Furthermore, the road surface material categories include cement road surfaces, waterlogged areas, oil-stained areas, gravel areas, and rigid obstacles.

[0031] This invention also discloses a robot navigation method and system based on semantic perception and multi-level cost maps, comprising:

[0032] An image acquisition module is configured to acquire environmental images of the mobile robot's direction of travel using an onboard vision sensor.

[0033] The semantic perception module is configured to process the environmental image using a semantic segmentation network based on the DeepLab architecture, identify the road surface material category, and generate a semantic segmentation map.

[0034] The map building module is configured to build a multi-level cost map containing physical attribute weights, and to map the physical attribute weights corresponding to the identified road surface material categories to the geometric raster map.

[0035] The decision planning module is configured to load a pre-trained deep reinforcement learning model, convert the multi-level cost map into a state space input to the model, and output the optimal action instruction.

[0036] The adaptive control module is configured to generate adaptive motion control commands containing speed constraints based on the optimal action command and the cost level corresponding to the current path point.

[0037] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0038] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention abandons the reliance on precise geometric modeling in traditional hierarchical planning (global planning + local planning), utilizes the powerful multi-scale feature extraction capabilities of the DeepLab network to solve the problem of inaccurate road surface material identification in complex scenarios, and integrates road surface physical properties into the reward function system through deep reinforcement learning algorithms, achieving end-to-end intelligent decision-making from "environmental perception" to "motion control." This method not only automatically learns human-like driving strategies such as "slowing down" or "detouring" when encountering low-adhesion or high-vibration areas, but also significantly improves the robot's adaptability and navigation safety in unknown unstructured environments. Attached Figure Description

[0039] Figure 1 The flowchart shows the robot navigation method and system based on semantic perception and multi-level cost map of the present invention.

[0040] Figure 2A schematic diagram illustrating the principles of semantic segmentation mask extraction and multi-level cost map construction;

[0041] Figure 3 This is a block diagram of the robot navigation method and system based on semantic perception and multi-level cost map of the present invention.

[0042] Figure 4 This is a schematic diagram of the internal structure of a computer device. Detailed Implementation

[0043] The technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this disclosure or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0044] Unless otherwise specified, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this disclosure. It should also be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn to actual scale. Techniques, methods, and apparatus known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and apparatus should be considered part of the specification.

[0045] To address the problem that existing mobile robot path planning technologies rely solely on geometric obstacle information while ignoring road surface physical properties, leading to slippage, vibration, or positioning loss in complex unstructured environments (such as waterlogged, oily, or gravelly surfaces), this disclosure proposes a robot navigation method and system based on semantic perception and a multi-level cost map. This invention extracts environmental semantic features in real time using a deep learning network, constructs a multi-level cost map incorporating physical attribute weights, and utilizes a deep reinforcement learning (DRL) algorithm to achieve end-to-end adaptive path decision-making and motion control, thereby achieving an intelligent balance between "traffic efficiency" and "operational safety."

[0046] Figure 1 This is a flowchart illustrating an embodiment of the robot navigation method and system based on semantic awareness and multi-level cost maps disclosed herein. The execution entity of this method can be the onboard computing unit of a mobile robot (such as an NVIDIA Jetson series industrial computer) or a remote cloud server.

[0047] like Figure 1As shown, in step 1 of the embodiment, environmental images of the mobile robot's driving direction are acquired through an onboard vision sensor and preprocessed.

[0048] It should be noted that the vision sensor is generally mounted on the front of the robot body and can be an RGB monocular camera or an RGB-D depth camera. Since the raw image may be affected by lighting, shaking, or lens distortion, it needs to undergo standardization preprocessing.

[0049] The preprocessing process includes:

[0050] Distortion correction: Eliminate barrel distortion of wide-angle lenses using a calibrated camera intrinsic parameter matrix;

[0051] Size normalization: Adjusts the image resolution to the standard size required for network input (e.g., 512×512 pixels).

[0052] Image enhancement: Histogram equalization (HE) or adaptive gamma correction algorithms are used to improve image contrast in low-light or high-light environments, ensuring the accuracy of subsequent semantic segmentation.

[0053] Step 2 of this embodiment processes the environmental image using a semantic segmentation network based on the DeepLab architecture to identify the road surface material category and generate a pixel-level semantic segmentation map. In some embodiments, step 2 specifically employs the DeepLabV3+ network model. Traditional convolutional networks struggle to simultaneously extract features from both distant small objects (such as puddles) and nearby large objects (such as grass underfoot) due to the perspective effect of "near objects appearing larger than distant objects" from a mobile robot's viewpoint. This embodiment addresses this problem by utilizing DeepLab's unique Atrous Convolution and Spatial Pyramid Pooling (ASPP) modules.

[0054] In the encoder stage of a convolutional neural network, dilated convolution allows for a significant expansion of the receptive field without increasing the number of parameters or computational cost. The formula for dilated convolution in the one-dimensional case is as follows:

[0055]

[0056] In the formula, To extract multi-scale road surface semantic features, The input is the preprocessed environmental image features, where i represents the spatial pixel coordinates on the environmental image acquired by the mobile robot's onboard vision sensor. For convolution kernel weights, Here, k is the kernel size, and k is the kernel stride index. This corresponds to the Dilation Rate in the DeepLab network architecture. When r=1, it is a standard convolution; when r>1, the convolution kernel performs interval sampling on the input feature map, thereby capturing multi-scale contextual semantic information.

[0057] After upsampling by the decoder, the network outputs a semantic mask image of the same size as the original image. The value of each pixel in the mask image represents the specific material category that the location belongs to (e.g., 0-cement road, 1-water accumulation, 2-oil stain, 3-gravel, 4-rigid obstacle).

[0058] Step 3 of the embodiment involves constructing a multi-level cost map containing physical attribute weights and transforming the multi-level cost map into the environment state space of a deep reinforcement learning network.

[0059] It should be noted that the core of this step lies in transforming "semantic information" into "mathematical cost." The multi-level cost map constructed in the example contains three logical layers:

[0060] 1. Static Layer: Generated by LiDAR or depth camera, marking impassable physical obstacles (walls, pillars), denoted as... .

[0061] 2. Semantic Attribute Layer: Based on the identification results of step 2, it is obtained through table lookup mapping. The system has a pre-built material attribute database (such as the "Material-Physical Attribute Weight Table"), for example:

[0062] Cement / asphalt roads: high coefficient of friction, high cost of passage. ;

[0063] Grass / carpet: Slightly higher resistance, higher cost of passage ;

[0064] Gravel roads: high risk of vibration, high cost of passage. ;

[0065] Water / oil stains: extremely high risk of slipping, high cost of passage. ;

[0066] 3. Inflation Layer: Centered on the high-cost region, Gaussian blurring or gradient inflation is applied outward to establish a risk buffer zone.

[0067] The map fusion unit calculates each grid using the following formula. Comprehensive cost of passage :

[0068]

[0069] In the formula, cost(x, y) is the comprehensive passage cost for each grid cell, (x, y) are the two-dimensional grid coordinates in the multi-level cost map, and c obs (x, y) represents the cost of the geometric barrier layer at that coordinate. The semantic layer's cost at that coordinate. The value represents the inflation cost, and α, β, and γ are the fusion weight coefficients for each layer.

[0070] Furthermore, to adapt to reinforcement learning networks, a state space needs to be constructed. A robot-centric local multi-level cost map (e.g., a local window cropped to 80×80; in practical applications, the size of this local window can be dynamically adjusted between 64×64 and 256×256 depending on the robot's computing power and perception range) is used as an image channel and stacked with the robot's current velocity vector and the relative polar coordinates of the target point to form a dimension... state tensor , as input to the DRL network.

[0071] Step 4 of this embodiment involves constructing a path planning model based on deep reinforcement learning (DRL) and outputting the optimal action command based on the current environmental state space. This embodiment abandons the limitations of traditional A* or DWA algorithms that rely on manually designed rules, and uses deep reinforcement learning (DRL) to learn driving strategies for complex road conditions.

[0072] 1. Action Space: Defined as a set of continuous or discrete control instructions. ,in For linear velocity, ω is the angular velocity.

[0073] 2. Reward Function: This is crucial for guiding the robot to learn "profit-seeking and harm-avoiding" behaviors. The reward function designed in this embodiment... as follows:

[0074]

[0075] Among them, the key semantic material penalty item Defined as:

[0076] In the formula, The penalty coefficient is... This is the robot's current location. For the robot at its current location The corresponding comprehensive transit value.

[0077] This formula means that when a robot enters a high-cost area (such as oil slicks), it receives a huge negative reward. After extensive training, the neural network "remembers" this penalty, thus automatically planning a path to avoid oil slicks during the reasoning phase, or choosing a relatively low-cost path when it cannot avoid them.

[0078] 3. Policy Update: The model parameters of the deep reinforcement learning network are updated using the principle of maximizing cumulative reward. For Deep Q-Network (DQN), the iterative update formula for its Q-value is:

[0079]

[0080] In the formula, Q(s,a) represents the state... Next action Action value function, This is the current state. For the current action, To perform the action The instant reward obtained afterward For learning rate, As a discount factor, For the state at the next moment, This is a candidate action for the next moment.

[0081] Step 5 of the embodiment generates adaptive motion control commands based on the optimal action command and the cost level corresponding to the current path point, thereby controlling the mobile robot to move. Although the DRL network can output planning commands, in order to ensure the underlying safety in extreme cases, this step introduces a rule-based safety threshold clamping mechanism.

[0082] The system monitors the average road cost along the robot's planned path in real time. (For example, take the arithmetic mean of the combined travel costs of all grids traversed by the planned path). If this cost exceeds a set safety threshold... (This indicates that the road conditions ahead are poor and speed must be reduced), which affects the linear velocity output by the DRL. Make corrections and generate final control commands. .

[0083]

[0084] In the formula, For the final control command, The original linear velocity output by the deep reinforcement learning model. For average road surface cost, As a safety threshold, This is the speed attenuation coefficient. This formula ensures that the worse the road conditions, the stricter the speed limit, thus physically preventing skidding and severe bumps.

[0085] Figure 3 This is a block diagram of the robot navigation method and system based on semantic perception and multi-level cost mapping disclosed herein. In response to the term "system" mentioned in the title, this embodiment defines the logical architecture of the system in detail. The system runs in the onboard computer of the mobile robot and includes:

[0086] Image acquisition module: Configured to drive the vehicle-mounted vision sensor, capture the ambient RGB image stream in real time, perform distortion correction and synchronization operations, and transmit the preprocessed image to the semantic analysis module.

[0087] Semantic Analysis Module: Configured to load the pre-trained DeepLab semantic segmentation model, perform forward inference on each input frame image, output a semantic segmentation mask containing road surface material classification information, and map it to the robot coordinate system.

[0088] Multi-level cost map construction module: This is the bridge connecting perception and decision-making. It is configured to receive geometric obstacle information and semantic material information, and dynamically update the local multi-level cost map according to a pre-set physical attribute weight table (such as a friction coefficient table) using the aforementioned weighted fusion formula.

[0089] The DRL decision planning module is the core of the system. It is configured to convert the multi-level cost map into a state tensor, input it into a trained deep reinforcement learning policy network, and directly infer the optimal linear and angular velocity probability distributions for the current moment.

[0090] Adaptive Control Module: Configured to perform safety checks on the actions output by the DRL. When it detects that the current or forward path is on a high-cost unstructured surface, it forcibly reduces the linear speed according to the speed correction formula and sends the final command to the robot chassis driver via CAN bus or serial port.

[0091] Based on the same technical solution, this disclosure also relates to a computer device (such as...) Figure 4 (As shown). The device includes a processor, memory, communication interface, and input / output interface connected via a system bus.

[0092] Processors: may include high-performance GPUs (for deep learning inference) and CPUs (for logic control).

[0093] Memory: Stores non-volatile program instructions that, when executed by the processor, implement the path planning method described above.

[0094] Input / output interfaces: connect to cameras, LiDAR sensors, and underlying motor drivers.

[0095] The beneficial effects achieved by this invention are as follows: This invention abandons the limitation of traditional path planning algorithms (such as A* and DWA) that rely solely on geometric location information, and endows the robot with the ability to identify the physical attributes of the road surface ("understanding the road") by utilizing visual semantic segmentation technology; by constructing a multi-level cost map, the abstract "road difficulty" is quantified into specific mathematical costs; and by utilizing the trial-and-error learning mechanism of deep reinforcement learning algorithms, a human-like driving strategy is realized in complex unstructured environments—that is, driving at full speed on flat roads and automatically slowing down or detouring on slippery or gravel roads. Experimental data show that this method significantly reduces the robot's slippage rate and positioning loss rate in complex outdoor environments, and improves the safety and robustness of autonomous navigation.

[0096] The above embodiments are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make several improvements and equivalent substitutions without departing from the principle of the present invention. All such improvements and equivalent substitutions to the claims of the present invention fall within the protection scope of the present invention.

Claims

1. A robot navigation method and system based on semantic perception and multi-level cost maps, characterized in that, Includes the following steps: Step 1: Collect environmental images of the mobile robot's direction of travel using the vehicle-mounted vision sensor and perform preprocessing. Step 2: Use a deep learning-based semantic segmentation network to process the preprocessed image, identify the road surface material category, and extract the road surface semantic features; Step 3: Construct a multi-level cost map containing physical attribute weights; the multi-level cost map is generated by mapping the physical attribute weights corresponding to the identified road surface material categories to a geometric grid map, and is used to quantify the cost of road passage. Step 4: Construct a path planning model based on deep reinforcement learning, transform the multi-level cost map into a state space, and output the optimal action instructions for controlling the mobile robot. Step 5: Generate adaptive motion control commands based on the optimal action command and the cost level corresponding to the current path point to control the mobile robot's movement.

2. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, Step 1 specifically involves: using the calibrated camera intrinsic parameter matrix to perform distortion correction on the acquired original image to eliminate lens distortion; normalizing the size of the distortion-corrected image to adjust it to the standard input size required by the semantic segmentation network; and using histogram equalization or adaptive gamma correction algorithms to enhance the image and improve image contrast.

3. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, Step 2 specifically involves: employing a semantic segmentation network based on the DeepLab architecture, utilizing dilated convolution to expand the receptive field and extract multi-scale features; the processing formula for the input feature map in the convolutional layers of the semantic segmentation network is as follows: ； In the formula, To extract multi-scale road surface semantic features, The input is the preprocessed environmental image features, where i represents the spatial pixel coordinates on the environmental image acquired by the mobile robot's onboard vision sensor. For convolution kernel weights, Here, k is the kernel size, and k is the kernel stride index. This corresponds to the Dilation Rate in the DeepLab network architecture.

4. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, Step 3 specifically involves: establishing a multi-level cost map that includes physical attribute weights, including: A pre-built material attribute database is used, containing friction coefficients and vibration risk coefficients corresponding to different road surface material categories. A weighted fusion formula is used to superimpose the semantic attribute layer, geometric obstacle layer, and expansion risk layer to calculate the comprehensive passage cost for each grid cell. The weighted fusion formula is as follows: ； In the formula, cost(x, y) is the comprehensive passage cost for each grid cell, (x, y) are the two-dimensional grid coordinates in the multi-level cost map, and c obs (x, y) represents the cost of the geometric barrier layer at that coordinate. The semantic layer's cost at that coordinate. The value represents the inflation cost, and α, β, and γ are the fusion weight coefficients for each layer.

5. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, Step 4 specifically involves constructing a path planning model based on deep reinforcement learning, including: Define a state space, which includes a local multi-level cost map centered on the mobile robot, the robot's current velocity vector, and the relative coordinates of the target point. Define the Action Space, which includes the combination of linear velocity and angular velocity of the mobile robot; Define a reward function, which includes a reward for reaching the target, a collision penalty, and a passage penalty based on the road surface material.

6. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 5, characterized in that, The calculation formula for the reward function includes a semantic material penalty term, and the specific formula is as follows: ； In the formula, This is a semantic material penalty item. The penalty coefficient is... This is the robot's current location. For the robot at its current location The corresponding comprehensive transit value.

7. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, In step 5, an adaptive motion control command is generated based on the optimal action command and the cost level corresponding to the current path point, including: The average road surface cost along the robot's planned path is detected. If the average road surface cost exceeds a preset safety threshold, the linear velocity in the optimal action command is clamped using a velocity correction formula. The velocity correction formula is: ； In the formula For the final control command, The original linear velocity output by the deep reinforcement learning model. For average road surface cost, As a safety threshold, is the velocity decay coefficient, and e is the base of the natural logarithm.

8. The robot navigation method and system based on semantic perception and multi-level cost map according to claim 1, characterized in that, The road surface material categories include cement road surfaces, waterlogged areas, oil-stained areas, gravel areas, and rigid obstacles.

9. A robot navigation method and system based on semantic perception and multi-level cost maps, used to implement the method as described in claim 1, characterized in that, include: An image acquisition module is configured to acquire environmental images of the mobile robot's direction of travel using an onboard vision sensor. The semantic perception module is configured to process the environmental image using a deep learning-based semantic segmentation network, identify the road surface material category, and generate a semantic segmentation map. The map building module is configured to build a multi-level cost map containing physical attribute weights, and to map the physical attribute weights corresponding to the identified road surface material categories to the geometric raster map. The decision planning module is configured to load a pre-trained deep reinforcement learning model, convert the multi-level cost map into a state space input to the model, and output the optimal action instruction. The adaptive control module is configured to generate adaptive motion control commands containing speed constraints based on the optimal action command and the cost level corresponding to the current path point.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images