A Robot Obstacle Avoidance Method Based on Dynamic Feature Perception

Abstract

This invention discloses a robot obstacle avoidance method based on dynamic feature perception, aiming to solve the obstacle avoidance problem of robots in complex environments where unknown static and dynamic obstacles coexist. The method first acquires sensor scan data and the robot's own pose data, and then performs feature decoupling: a multi-scale residual convolutional network is used to extract static environment features, while a long short-term memory network is combined to learn the motion trend features of dynamic obstacles. Subsequently, an attention mechanism layer fuses the static environment features, dynamic obstacle motion trend features, and the robot's own pose data into a comprehensive perception feature vector, which is then input into a deep reinforcement learning model to output the optimal obstacle avoidance action. This invention can effectively distinguish between static and dynamic obstacles, and by predicting dynamic trends, it significantly improves the accuracy, safety, and efficiency of obstacle avoidance decisions.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to a method for dynamic feature perception and obstacle avoidance of mobile robots based on deep reinforcement learning, applicable to complex environments containing unknown static and dynamic obstacles. Background Technology

[0002] With the increasing prevalence of mobile robots in everyday scenarios such as shopping malls, factories, and hotels, obstacle avoidance plays a crucial role in robot autonomous navigation. Traditional obstacle avoidance algorithms, such as artificial potential field methods or vector field histograms, can achieve basic obstacle avoidance planning in simple, known environments. However, in environments with both unknown static and dynamic obstacles, these traditional algorithms struggle to cope with the dynamic nature and uncertainty of the environment, as well as the complexity of human-robot interaction.

[0003] In recent years, deep reinforcement learning has demonstrated tremendous potential in the field of robotics. Deep reinforcement learning-based methods can enable robots to autonomously navigate and avoid obstacles in dynamic environments by designing reward functions. However, existing advanced deep reinforcement learning methods still have limitations when applied to environments with both unknown static and dynamic obstacles. For example, while dynamic warning zones or crowds and obstacles avoidance with DRL methods can achieve good obstacle avoidance results, they lack effective spatiotemporal modeling of dynamic information, making it difficult to distinguish between static and dynamic obstacles. This leads to uncertainty in obstacle avoidance decisions, affecting the robot's efficiency and safety, and also reducing the comfort experience for pedestrians in the surrounding area.

[0004] Current methods generally suffer from insufficient perception of static and dynamic obstacle information in the environment, meaning that robots are at significant risk of misjudging obstacles and cannot guarantee optimal obstacle avoidance performance. Therefore, designing an obstacle avoidance method that can effectively extract and distinguish the features of static and dynamic obstacles, and make accurate, safe, and efficient decisions accordingly, is a pressing technical problem to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and propose a robot obstacle avoidance method based on dynamic feature perception. This method can accurately perceive and distinguish static and dynamic obstacles in the environment, effectively extract spatiotemporal dynamic information, thereby achieving more efficient, stable and safe obstacle avoidance and improving the comfort experience of pedestrians.

[0006] To achieve the above objectives, the technical solution provided by this invention is: a robot obstacle avoidance method based on dynamic feature perception, comprising the following steps:

[0007] S1: Acquire sensor scan data of the robot in the current environment and the robot's own pose data;

[0008] S2: Decouple the sensor scan data by extracting static environment features and dynamic obstacle movement trend features respectively;

[0009] S3: The static environment features, dynamic obstacle movement trend features, and robot pose data are fused to generate a unified comprehensive perception feature vector.

[0010] S4: Input the integrated perception feature vector into a pre-trained deep reinforcement learning model, and the deep reinforcement learning model outputs the optimal obstacle avoidance action command to control the robot's movement.

[0011] Furthermore, in step S2, the process of extracting static environmental features includes:

[0012] S21: For the sensor scan data set {d1,d2,d3,...,d...} i ,...,d N Each sensor scan data in} is preprocessed, where d i This represents the original distance of the i-th sensor scan data, which is also the original distance of the i-th scan point. N represents the total number of sensor scan data. The original distance of the i-th scan point is converted into a two-dimensional coordinate point (x, y) through Cartesian coordinate transformation. i ,y i ):

[0013]

[0014] In the formula, d represents the distance of the i-th preprocessed scan point. max θ represents the preset maximum detection distance of the sensor. i x represents the angle of the i-th laser beam. i and y i The original distance d represents the i-th scan point. i The corresponding x and y coordinate values ​​after conversion;

[0015] S22: Transfer the two-dimensional coordinates (x, y) generated in step S21 to... i ,y i Mapping to a 2D raster map based on raster resolution In the middle, the two-dimensional coordinate point (x i ,y i raster index The calculation method is as follows:

[0016]

[0017] In the formula, and Represents a two-dimensional coordinate point (x i ,y i The row and column index values ​​of the corresponding raster map are used to uniquely determine the position of a raster within the entire raster map. min and y min The origin of the raster map is defined, and the transformation datum from the two-dimensional coordinate points to the row and column indices of the raster map is defined, where r represents the raster resolution.

[0018] S23: Transfer the two-dimensional raster map The input is fed into a multi-scale residual convolutional network for learning, and static environment features are extracted and output. Each time t has corresponding static environmental characteristics.

[0019] Furthermore, in step S2, the process of extracting the motion trend features of dynamic obstacles includes:

[0020] S31: Raster map for consecutive times t and t-1 and Based on the changes in the robot's own pose data (Δx, Δy, Δθ), where Δx and Δy represent the displacement changes of the robot at consecutive times t and t-1, and Δθ represents the change in the robot's orientation angle, for... The point in the middle corresponds to the original two-dimensional coordinate point (x). i ,y i Motion compensation is performed to obtain the new raster index after motion compensation.

[0021]

[0022] In the formula, and Representative to The two-dimensional coordinates (x) of the point in the middle. i ,y i The new row and column index values ​​of the raster map after motion compensation, where r represents the raster resolution and x... min and y min The origin of the raster map is defined, and the transformation datum for the row and column indices of the raster map from two-dimensional coordinate points is defined, δ. x and δ y This is the introduced disturbance bias;

[0023] S32: After obtaining a new raster map at time t-1 through motion compensation, the dynamic feature raster is identified using the inter-frame spatiotemporal difference method on this new raster map and the raster map at time t.

[0024]

[0025] In the formula, A dynamic feature grid representing time t. For indicator functions, The raster map representing time t in the raster index The value at that location, The raster map after motion compensation at time t-1 is indexed in the new raster. The value at;

[0026] S33: From the dynamic feature grid Extracting the position data of dynamic obstacles at time t and the location data The sensor scan data set {d1,d2,d3,...,d...} at the current time t i ,...,d N After fusion, the data is fed into a Long Short-Term Memory (LSTM) network for temporal feature learning, extracting and outputting the dynamic obstacle movement trend features at the current time t.

[0027] Furthermore, before motion compensation, a continuous planar structure filtering step is included to eliminate interference from continuous static structures in the sensor scan data. Specifically, this involves dynamically adjusting the window length L to detect continuous points and filtering out points with a distance difference less than a threshold τ. diff The continuous static structure, where the window length L is calculated using the following formula:

[0028]

[0029] In the formula, N base Adjust the base for the distance window, d i This represents the original distance of the i-th scan point being processed during the filtering calculation.

[0030] Furthermore, in step S3, an attention mechanism layer is used to analyze the static environment features at the current time t. Dynamic obstacle movement trend characteristics And the robot's own pose data P t =[x t ,y t ,θ t Weighted fusion is performed to generate a comprehensive perceptual feature vector. This enables deep reinforcement learning models to autonomously focus on features that are more important for decision-making; where x t y t θ represents the horizontal and vertical coordinates of the robot at the current time t in a standard two-dimensional Cartesian coordinate system. t This represents the robot's orientation angle at the current time t.

[0031] Furthermore, in step S4, the deep reinforcement learning model is a duel dual deep Q network based on an attention mechanism.

[0032] Furthermore, the sensor scanning data originates from a two-dimensional lidar.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0034] 1. Feature decoupling and accurate perception: This invention can effectively separate and extract static environmental features and dynamic obstacle movement trend features from sensor scan data, which significantly reduces the risk of robots misjudging the environment.

[0035] 2. Efficient use of spatiotemporal information: By introducing a long short-term memory network to process temporal features, this method can obtain the dynamic trend features of obstacles, thereby making more predictive obstacle avoidance decisions rather than simple reactive obstacle avoidance.

[0036] 3. Improved decision-making performance: By combining the attention mechanism layer with an advanced dual-depth Q-network, the method of this invention can more intelligently integrate static environment features, dynamic obstacle motion trend features, and the robot's own pose data, focusing on key information, thereby achieving a higher success rate, a lower collision rate, and a more efficient path in dual unknown obstacle environments where both unknown static and dynamic obstacles exist simultaneously.

[0037] 4. Improve human-computer interaction experience: By accurately identifying and predicting the movement trends of dynamic obstacles, such as the movement of pedestrians, robots can take obstacle avoidance behaviors that are more in line with social norms, reduce the intrusion on pedestrians, and significantly improve the comfort experience of pedestrians. Attached Figure Description

[0038] Figure 1 This is a structural framework diagram of the deep reinforcement learning model of the method of the present invention.

[0039] Figure 2 This is a schematic diagram comparing the effects of filtering a continuous planar structure before and after using the method of the present invention.

[0040] Figure 3 This is a comparison chart of the target point distance convergence curves of the method of this invention and other methods during the training process.

[0041] Figure 4 This is a comparison diagram of the navigation trajectories of the method of the present invention and other methods in static and dynamic test environments. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0043] like Figure 1 As shown in the figure, this embodiment discloses a robot obstacle avoidance method based on dynamic feature perception, and its specific implementation process is as follows:

[0044] 1) Data acquisition and preprocessing.

[0045] The robot is equipped with a two-dimensional LiDAR as its main sensor, which acquires a set of 360-degree sensor scan data L in real time at the current time t. t ={d1,d2,d3...d i ...d 360 This is also two-dimensional LiDAR distance data. Simultaneously, the robot's pose data P at the current time t is obtained through positioning systems such as odometry. t =[x t ,y t ,θ t ].

[0046] 2) Extraction of static environmental features and dynamic obstacle movement trend features, i.e., feature decoupling.

[0047] This step is one of the core aspects of the present invention, which aims to decompose the mixed environmental information into independent static environmental features and dynamic obstacle movement trend features for processing.

[0048] 2.1) Acquisition of Static Environmental Characteristics

[0049] Extract static information from the environment, such as walls and fixed obstacles.

[0050] First, the two-dimensional lidar range data L at the current time t. t Processing will result in the sensor exceeding its maximum detection range (d). max For example, the data for 8.0 meters is truncated. Then, L... t Convert to two-dimensional coordinates (x) i ,y i ):

[0051]

[0052] In the formula, d represents the distance after the i-th preprocessing step. max This represents the preset maximum detection distance, such as 8.0 meters, θ. i x represents the angle of the i-th laser beam. i and y i d represents the distance data of the i-th two-dimensional lidar. i The corresponding x and y coordinates after conversion.

[0053] The one-dimensional distance data from the two-dimensional lidar is converted into two-dimensional coordinate points, and these two-dimensional coordinate points are mapped onto a two-dimensional grid map at the current time t. This grid map visually represents the relative positions of static obstacles around the robot.

[0054] Finally, the generated raster map is input into a multi-scale residual convolutional network. This network effectively extracts multi-scale spatial features from the raster map, ultimately compressing them into a static environmental feature representation at the current time t.

[0055] 2.2) Obtaining the dynamic obstacle movement trend characteristics

[0056] It focuses on capturing information about dynamic obstacles, such as pedestrians, from temporal changes.

[0057] First, to eliminate background changes caused by the robot's own motion, motion compensation is needed for the grid map at consecutive times t and t-1. This is achieved by utilizing the changes in the robot's pose data (Δx, Δy, Δθ), where Δx and Δy represent the displacement changes at consecutive times t and t-1, and Δθ represents the change in the robot's orientation angle. The grid map at the previous time t-1 is then used for this purpose. Align with the raster map at the current time t.

[0058] Before compensation, to avoid pseudo-dynamic obstacle movement trend characteristics caused by positioning or radar errors in large static objects such as walls during motion compensation, a continuous planar structure filtering mechanism was introduced. For example... Figure 2 As shown, the formula for calculating the window length L is:

[0059]

[0060] In the formula, N base Adjust the base for the distance window, d i This represents the original distance to the i-th data point being processed during the filtering calculation. By continuously and dynamically adjusting the window length L, continuous planar structures are identified and filtered out, such as... Figure 2 (b) is the original scan. Figure 2 In the middle (c) scan, it can be seen that pedestrian points are retained and wall points are filtered out.

[0061] By comparison and after compensation and filtering The inter-frame spatiotemporal difference method can be used to identify changing grids; these changing regions are called dynamic feature grids.

[0062] Extract the dynamic obstacle information at the current time t Distance data L from two-dimensional lidar t The data is pieced together to form a spatiotemporal joint information system containing both background and dynamic objects. This joint information is then input into a Long Short-Term Memory (LSTM) network, which learns the temporal dependencies of the joint information, thereby extracting the motion trend features of the dynamic obstacles.

[0063] 3) Multi-feature fusion and decision-making

[0064] The static environment characteristics obtained at the current time t Dynamic obstacle movement trend characteristics And the robot's own pose data P t The features are then concatenated. The concatenated feature vector is input into an attention mechanism layer. This layer dynamically weights the features based on their importance to the current decision, generating a more comprehensive perceptual feature vector with higher information density.

[0065] Finally, the integrated perceptual feature vector is input into a dueling dual-depth Q-network based on an attention mechanism. This dual-depth Q-network can evaluate the Q-value of taking different actions in the current state, such as moving forward, backward, turning left, and turning right. The robot selects the action with the highest Q-value to execute according to a policy such as ε-greedy, thereby achieving intelligent obstacle avoidance.

[0066] 4) Experimental verification and results

[0067] To verify the effectiveness of the robot obstacle avoidance method based on dynamic feature perception, a large number of experiments were conducted in the Gazebo simulation environment. The experimental environment included different numbers of static obstacles and dynamic pedestrians moving at different speeds, such as 0.3 m / s to 1.2 m / s.

[0068] Training process comparison: such as Figure 3 As shown, the robot obstacle avoidance method based on dynamic feature perception provided by this invention, and the dynamic warning zones method and crowds and obstacles avoidance with DRL method which use the method of this invention to extract static environmental features and dynamic obstacle motion trend features, converge significantly faster than the dynamic warning zones method and crowds and obstacles avoidance with DRL method which do not use the method of this invention. This indicates that the feature extraction method of this invention is more suitable for navigation and obstacle avoidance tasks related to deep reinforcement learning.

[0069] Success rate measures a robot's ability to successfully navigate from a starting point to a predetermined target point within a given number of tests. The success rate is calculated as follows:

[0070]

[0071] In the formula, SR represents the success rate, T represents the total number of tests, and g i It is a Boolean value; if the robot successfully reaches the target point in the i-th test, then g... i =1; otherwise 0;

[0072] Collision rate quantifies the average number of collisions a robot makes with obstacles during navigation. It is a key indicator of robot navigation safety; a low collision rate means the obstacle avoidance strategy of the algorithm is safer and more effective. The collision rate is calculated as follows:

[0073]

[0074] In the formula, CR represents the collision rate, T represents the total number of tests, and c i This represents the total number of collisions that occurred to the robot in the i-th test;

[0075] Average steps refer to the average number of decisions a robot makes to reach its destination from the starting point in all successfully completed navigation tasks. Fewer decisions indicate better path planning and higher navigation efficiency. The average steps are calculated as follows:

[0076]

[0077] In the formula, MS represents the average number of steps, T represents the total number of tests, and k i This represents the total number of decisions the robot makes to reach its destination in the i-th test.

[0078] The Comfort Erosion Index is specifically designed to quantify the disturbance or discomfort a robot causes to surrounding dynamic obstacles, such as pedestrians, during navigation. It assesses the social acceptability of the robot's behavior, which is crucial in human-robot coexistence scenarios. A Comfort Erosion Index closer to 0 indicates less disturbance to pedestrians and a better pedestrian experience. The calculation of the Comfort Erosion Index involves two steps. First, the Comfort Erosion Index for a single pedestrian is calculated as follows:

[0079]

[0080] In the formula, This represents the comfort erosion index of the robot for the j-th pedestrian in the i-th test; α is an indicator function used to determine whether the robot has entered the personal space of the j-th pedestrian in the direction of movement during the i-th test. If it has entered, the value is 1; if it has not entered, the value is 0. α is a constant that defines how severe the basic penalty is for violating the pedestrian's dynamic area. This represents the Euclidean distance between the robot and the j-th pedestrian at some moment in the i-th test; The design means that the farther the robot is from the pedestrian, the smaller the penalty value; δ is a constant that defines how close a distance is considered too close; β is an indicator function used to determine whether the Euclidean distance between the robot and the j-th pedestrian is less than a preset safety threshold δ in the i-th test. If it is less than the threshold, the value is 1; if it is not less than the threshold, the value is 0. β is a constant that determines the rate at which the penalty increases as the distance shrinks after entering the safety threshold δ range. The design indicates that the closer the objects are, the greater the difference and the heavier the penalty.

[0081] Then, the comfort erosion scores caused by all pedestrians in all test rounds are summed, and the average is calculated to obtain the final comfort erosion index. The calculation method is as follows:

[0082]

[0083] In the formula, CEI represents the comfort erosion index, T represents the total number of tests, and N represents the comfort erosion index. * This represents the total number of people in the i-th test. This represents iterating through each round of testing and accumulating the comfort erosion index for each pedestrian in that round of testing.

[0084] Performance Comparison: As shown in Table 1, in a dynamic testing environment, the robot obstacle avoidance method based on dynamic feature perception provided by this invention achieves a success rate improvement of up to 31.4% and a collision rate reduction of 72.9% compared to the dynamic warning zones method and the crowds and obstacles avoidance with DRL method. In particular, the method of this invention scores significantly lower than other methods on the comfort erosion index, which is crucial for pedestrian experience, demonstrating that the method of this invention can better consider pedestrian comfort while achieving safe obstacle avoidance.

[0085] Table 1: Performance Comparison Evaluation in Static and Dynamic Environments

[0086]

[0087]

[0088] Qualitative analysis: such as Figure 4The trajectory diagrams shown illustrate that, in dynamic environments, the dynamic warning zones method exhibits hesitant behavior, failing to distinguish between static and dynamic obstacles, while the crowds and obstacles avoidance with DRL method tends to ignore dynamic obstacles and proceed in a straight line. In contrast, the trajectory provided by the method of this invention is smooth and efficient, effectively avoiding all obstacles, demonstrating that the method of this invention can effectively identify dynamic features and make stable decisions.

[0089] Robustness analysis: Experiments by changing the number of static obstacles and the average speed of dynamic obstacles show that the method of the present invention can still maintain a high success rate and a low collision rate under different environmental configurations, and has good robustness and generalization ability.

[0090] In summary, the method of this invention, through unique decoupling and perception design of static environmental features and dynamic obstacle motion trend features, combined with an advanced deep reinforcement learning framework, can effectively explore the spatiotemporal correlations in the environment and significantly improve the obstacle avoidance performance, safety, and social acceptability of robots in unknown dynamic environments.

[0091] Experimental Conclusions: Existing obstacle avoidance methods for robots in complex environments containing both unknown static and dynamic obstacles suffer from insufficient perception of static and dynamic information and difficulty in effectively distinguishing between the two types of obstacles, leading to decision uncertainty and impacting obstacle avoidance efficiency and safety. This invention proposes a robot obstacle avoidance method based on dynamic feature perception. Extensive experimental evaluation in the Gazebo simulation environment shows that the proposed method, through unique decoupling of static environment features and dynamic obstacle motion trend features, a long short-term memory network, and a duel-based dual-depth Q-network based on an attention mechanism, significantly outperforms existing state-of-the-art methods in key performance indicators such as obstacle avoidance success rate, collision rate, and the comfort erosion index, which is crucial for pedestrian experience. The experimental results fully demonstrate that the proposed method can effectively extract and distinguish between static and dynamic features in the environment, and make accurate, safe, and efficient decisions accordingly. Future research will further explore the deployment and optimization of this method on real physical robot platforms, as well as its application in more complex group interaction scenarios, demonstrating broad application prospects and promotional value.

[0092] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A robot obstacle avoidance method based on dynamic feature perception, characterized in that, Includes the following steps: S1: Acquire sensor scan data of the robot in the current environment and the robot's own pose data; S2: Decouple the sensor scan data by extracting static environment features and dynamic obstacle movement trend features respectively; The process of extracting static environment features includes: S21: Sensor scan data set Each sensor scan data in the process is preprocessed, where Representing the The sensor scan data is also the first The original distance of each scan point Represents the total number of sensor scan data, and transforms the first data into Cartesian coordinates. The original distance of each scan point is converted into two-dimensional coordinates. : ； In the formula, Representing the Distance between preprocessed scan points This represents the preset maximum detection distance of the sensor. Representing the The angle of the laser beam and Representing the Original distance of each scan point The corresponding x and y coordinate values ​​after conversion; S22: Transfer the two-dimensional coordinate points generated in step S21 Mapped to a 2D raster map based on raster resolution. In the middle, two-dimensional coordinate points raster index The calculation method is as follows: ； In the formula, and Represents two-dimensional coordinate points The row and column index values ​​of the corresponding raster map are used to uniquely determine the position of a raster within the entire raster map. and The origin of the raster map is defined, and a transformation datum for the row and column indices of the raster map from two-dimensional coordinate points is defined. Represents raster resolution; S23: Transfer the two-dimensional raster map The input is fed into a multi-scale residual convolutional network for learning, and static environment features are extracted and output. Every moment Each has corresponding static environment characteristics ; The process of extracting dynamic obstacle motion trend features includes: For continuous time and raster map and Based on the changes in the robot's own pose data ,in and Representing consecutive moments and The amount of displacement change of the robot. The amount representing the change in the robot's orientation angle, for The point in the middle corresponds to the original two-dimensional coordinate point. Perform motion compensation to obtain the new raster index after motion compensation. : ； In the formula, and Representative to The two-dimensional coordinates of the point in the middle New row and column index values ​​of the raster map after motion compensation. Represents raster resolution. and The origin of the raster map is defined, and a transformation datum for the row and column indices of the raster map from two-dimensional coordinate points is defined. and This is the introduced disturbance bias; At the moment when motion compensation is achieved After creating the new raster map, the new raster map and time... The raster map uses inter-frame spatiotemporal difference to identify dynamic feature raster. : ； In the formula, Representative moment Dynamic feature grid, For indicator functions, Representative moment Raster map in raster index The value at that location, Representative moment The raster map after motion compensation is indexed in the new raster index. The value at; From the dynamic feature grid Extracting time Location data of dynamic obstacles and the location data With the current moment Sensor scan data set After fusion, the data is fed into a Long Short-Term Memory (LSTM) network for temporal feature learning, extracting and outputting the current time step. Dynamic obstacle movement trend characteristics ; Before motion compensation, a continuous planar structure filtering step is included to eliminate interference from continuous static structures in the sensor scan data. Specifically, this involves dynamically adjusting the window length. To detect continuous points and filter out those with a distance difference less than a threshold. A continuous static structure, where the window length The calculation formula is: ； In the formula, Adjust the base value for the distance window. This represents the current processing step during the filtering calculation. The original distance of each scan point; S3: The static environment features, dynamic obstacle motion trend features, and the robot's own pose data are fused to generate a unified comprehensive perception feature vector, as follows: Employ attention mechanism layer for the current time step Static environment characteristics Dynamic obstacle movement trend characteristics and the robot's own pose data Weighted fusion is performed to generate a comprehensive perceptual feature vector. This enables deep reinforcement learning models to autonomously focus on features that are more important for decision-making; among them, , This indicates the robot's current moment in a standard two-dimensional Cartesian coordinate system. The x and y coordinate values, Indicates the robot's current state. Direction angle; S4: Input the integrated perception feature vector into a pre-trained deep reinforcement learning model, and the deep reinforcement learning model outputs the optimal obstacle avoidance action command to control the robot's movement.

2. The robot obstacle avoidance method based on dynamic feature perception according to claim 1, characterized in that, In step S4, the deep reinforcement learning model is a duel dual deep Q network based on an attention mechanism.

3. The robot obstacle avoidance method based on dynamic feature perception according to claim 1, characterized in that, The sensor scanning data comes from a two-dimensional lidar.

Images