Behavior tree-based mobile robot multi-strategy obstacle avoidance decision-making agent and method

Abstract

The present application relates to the technical field of robot autonomous navigation and intelligent decision-making, and particularly relates to a mobile robot multi-strategy obstacle avoidance decision-making agent and method based on a behavior tree, comprising: receiving sensor original perception data through an obstacle feature extraction module, calculating an obstacle displacement amount to determine its motion state and output structured description information; obtaining the running state of each sensor and environmental conditions through a sensor confidence assessment module, calculating a single-sensor comprehensive confidence score and fusing and outputting; using an obstacle avoidance decision-making engine based on a behavior tree, with a fallback node as a root node, a plurality of strategy branches are connected in series according to priority from high to low, and in recursive traversal, the structured information and the fused confidence score are read by a condition node, when the check fails, the route is automatically degraded through the fallback node until a branch is activated and an abstract action instruction is output; and converting the instruction into chassis bottom layer control commands or a global re-planning request through an obstacle avoidance action execution module.

Description

Technical Field

[0001] This invention relates to the field of robot autonomous navigation and intelligent decision-making technology, and in particular to a behavior tree-based multi-strategy obstacle avoidance decision-making agent and method for mobile robots. Background Technology

[0002] With the development of autonomous driving and industrial automation technologies, various mobile robot platforms requiring autonomous navigation are widely used in a variety of complex scenarios. For example, in agricultural inspection, robots need to deal with temporarily piled-up farm tools and transport vehicles when performing crop inspection tasks in greenhouses or fields; in park security, robots need to deal with dynamic and static obstacles such as pedestrians, temporarily parked vehicles, and construction barriers in real time when performing patrol tasks on park roads; in industrial warehouse inspection, robots often need to autonomously avoid forklifts and pallets in narrow passages with a width of only about 2 meters; in logistics and distribution, delivery robots need to deal with complex environments where people and vehicles mix when performing last-mile delivery in parks or campuses. Currently, these mobile robots typically achieve decision control by modeling obstacle avoidance behavior as a finite state machine, so that after detecting the above obstacles, the robot can explicitly switch between fixed state nodes such as normal driving, deceleration and observation, stopping and waiting, or detouring according to preset transition conditions.

[0003] Traditional obstacle avoidance decision-making schemes based on finite state machines highly couple the transition logic of state nodes with the specific behavioral logic. When the system needs to continuously add and expand new obstacle avoidance decision-making strategies based on different environmental or obstacle characteristics, the number of internal states and the transition conditions between states increases exponentially. Developers have to re-examine and hard-code all existing state transition relationships, resulting in the inability to completely decouple control logic from atomic behavior and extremely poor scalability and maintainability when adding new strategies. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a behavior tree-based multi-strategy obstacle avoidance decision-making agent and method for mobile robots, aiming to improve the problems of inability to completely decouple control logic from atomic behaviors and extremely poor scalability and maintainability when adding new strategies.

[0005] In a first aspect, the present invention provides the following technical solution: a behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots, comprising: The obstacle feature extraction module is used to receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of the obstacle to determine the motion state of the obstacle, and output structured obstacle description information. The sensor confidence assessment module is used to acquire the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of a single sensor and fuse them, and output the fused confidence score. The behavior tree obstacle avoidance decision engine uses a backtrack node as the root node. Under the backtrack node, multiple strategy branches are connected in descending order of priority. The behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, the structured obstacle description information and the fusion confidence score are read through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the backtrack node automatically downgrades the routing to the next priority strategy branch until a certain strategy branch returns a success or running state and outputs an abstract action instruction. The obstacle avoidance action execution module is used to receive the abstract action instruction and convert it into the underlying control command of the mobile robot control chassis, or into a global replanning request for replanning the mobile robot path.

[0006] Preferably, in the obstacle feature extraction module, calculating the displacement of the obstacle includes: Establish a sliding window with a preset number of observation frames, and continuously acquire the center position coordinates of the same obstacle in multiple consecutive frames within the sliding window; Calculate the absolute displacement of the center position coordinates between two adjacent frames within the sliding window, and then sum and average the absolute displacements between all adjacent frames to output the displacement of the obstacle.

[0007] Preferably, in the obstacle feature extraction module, determining the motion state of the obstacle and outputting structured obstacle description information includes: Configure obstacle type enumeration values ​​in the structured obstacle description information; Based on the displacement of the obstacle, combined with the comprehensive confidence score of the single sensor and the current environmental conditions, the motion state of the obstacle is determined as static, dynamic or suspected false detection. The corresponding judgment result is written as the corresponding status flag into the obstacle type enumeration value to complete the configuration and output of the structured obstacle description information.

[0008] Preferably, in the sensor confidence assessment module, the calculation of the overall confidence score for a single sensor includes: Acquire the current operating status and environmental condition data of each sensor on the mobile robot; Based on the current operating status and environmental conditions data, the data frame rate stability component, the data validity ratio component, and the historical detection consistency component are calculated respectively. The data frame rate stability component, the data validity ratio component, and the historical detection consistency component are weighted and summed with the environmental attenuation factor determined based on the current interference density to output a single sensor comprehensive confidence score.

[0009] Preferably, in the sensor confidence assessment module, the overall confidence scores of the single sensors are fused, and the output fused confidence score includes: When multiple heterogeneous sensors are equipped, a weighted geometric average algorithm is used to fuse the comprehensive confidence scores of each individual sensor and output a unified fused confidence score. When any of the sensors is detected to be offline, the default weight of the offline sensor is assigned to the sensor that is currently online.

[0010] Preferably, in the behavior tree obstacle avoidance decision engine, multiple policy branches are chained together under the backoff node in descending order of priority, including: Under the fallback node, the following branches are connected in descending order of priority: autonomous detour strategy branch, false detection handling strategy branch, waiting and fallback strategy branch, and emergency safety strategy branch.

[0011] Preferably, executing the autonomous detour strategy branch includes: The condition nodes are used to sequentially verify the nearest distance to the obstacle, the confidence score, the passable space width, and whether the generated detour waypoints point to the avoidance area; When the verification is successful, the action node calculates the coordinates of the temporary detour target waypoint that deviates from the original travel path based on the center coordinates, circumcircle radius and minimum safe distance of the obstacle in the structured obstacle description information, and outputs the coordinates of the temporary detour target waypoint as an abstract action command.

[0012] Preferably, executing the false detection handling strategy branch and the waiting and fallback strategy branch includes: When the obstacle type enumeration value is a suspected false detection flag, the false detection handling strategy branch is triggered: save the current trigger detection threshold of the current sensor, increase the number of detection points or the detection confidence requirement, and record the current false detection location coordinates; if the false detection signal disappears within the preset monitoring window, restore the threshold and return to the success status; If the false detection signal persists within the preset monitoring window, the routing is downgraded to the waiting and fallback strategy branch, and the child fallback nodes under it are used to try sequentially: The first-level stationary waiting strategy involves cyclically checking the status of obstacles within a preset time. The second-level slow-speed approach strategy reduces the chassis speed and temporarily lowers the sensor sensitivity when the static obstacle distance condition is met. The third-level global replanning request strategy outputs the global replanning request when the first two levels of strategies fail.

[0013] Preferably, executing the emergency security policy branch includes: The nearest distance to the obstacle in the original perception data and the obstacle type enumeration value in the structured obstacle description information are read using the condition node; When the nearest distance to the obstacle is less than a preset safety collision threshold, and the obstacle type enumeration value is a dynamic flag or a static flag, an emergency safety strategy branch is triggered. The highest priority emergency braking interruption command is directly output by the action node as the abstract action command, so that the obstacle avoidance action execution module can control the chassis to perform full-brake emergency stop.

[0014] Secondly, the present invention provides the following technical solution: a multi-strategy obstacle avoidance decision-making method for mobile robots based on behavior trees, the method comprising: S1. Receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of the obstacle to determine the motion state of the obstacle, and output structured obstacle description information. S2. Obtain the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of the single sensor and fuse them, and output the fused confidence score. S3. Using a behavior tree obstacle avoidance decision engine, with a fallback node as the root node, multiple strategy branches are connected in descending order of priority under the fallback node; the behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, the structured obstacle description information and the fusion confidence score are read through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the fallback node automatically downgrades the routing to the next priority strategy branch until a certain strategy branch returns a success or running state and outputs an abstract action instruction. S4. Receive the abstract action instruction and convert it into a low-level control command of the mobile robot control chassis, or convert it into a global replanning request for replanning the mobile robot path.

[0015] The present invention has the following beneficial effects: 1. In this invention, by attaching independent condition nodes and action nodes under the fallback control node of the behavior tree, the control logic and atomic obstacle avoidance behavior are completely decoupled, and new policy branches can be added and used immediately. This solves the problem that in the traditional finite state machine (FSM) scheme, the number of states and transition conditions increases exponentially with the number of policies, making the system extremely difficult to maintain and expand.

[0016] 2. In this invention, through the logical or state transition update mechanism of the Fallback node at the root of the behavior tree, the adaptive sorting and multi-layer dynamic routing switching of the four-level policy priority from the autonomous detour policy branch, through the false detection handling policy branch, the waiting and fallback policy branch, and finally downgraded to the emergency safety policy branch is realized. This solves the problems of traditional obstacle avoidance decision schemes having a single policy, lacking time priority control, and being prone to control oscillations between multiple near-utility behaviors.

[0017] 3. In this invention, the sensor confidence assessment module performs a multi-dimensional weighted comprehensive evaluation of the sensing hardware, and sets up a special false detection tolerance branch between detour and waiting in conjunction with the behavior tree condition nodes. This enables accurate identification and time-series dynamic filtering of low-confidence noise signals caused by environmental interference such as dust, raindrops, and strong backlight. This solves the problem that existing obstacle avoidance decision-making schemes lack sensor data quality perception capabilities and are prone to misjudging environmental noise as real obstacles, resulting in frequent robot stoppages and unnecessary task interruptions. Attached Figure Description

[0018] Figure 1 This is an architecture diagram of the behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots proposed in this invention. Figure 2 This is a flowchart of the data processing for the obstacle feature extraction module of the behavior tree-based mobile robot multi-strategy obstacle avoidance decision-making agent proposed in this invention. Figure 3 This is a graph showing the evaluation factors and weights of the sensor confidence evaluation module for the multi-strategy obstacle avoidance decision-making agent of a mobile robot based on behavior tree proposed in this invention. Figure 4 This is a complete tree structure diagram of the behavior tree obstacle avoidance decision engine for a multi-strategy obstacle avoidance decision-making agent of a mobile robot based on behavior trees, as proposed in this invention. Figure 5 This is a flowchart of a single tick traversal of the behavior tree for a multi-strategy obstacle avoidance decision-making agent for a mobile robot based on behavior trees, as proposed in this invention. Figure 6 This is a schematic diagram comparing the behavior tree and finite state machine (FSM) of the behavior tree-based mobile robot multi-strategy obstacle avoidance decision-making agent proposed in this invention under a strategy switching scenario. Figure 7 This is a schematic diagram of the geometric calculation for generating waypoints for a multi-strategy obstacle avoidance decision-making agent of a mobile robot based on behavior tree proposed in this invention. Figure 8 This is a time sequence diagram of the entire obstacle avoidance decision-making process of the multi-strategy obstacle avoidance decision-making agent for mobile robots based on behavior trees proposed in this invention. Figure 9This is a flowchart of the behavior tree-based multi-strategy obstacle avoidance decision-making method for mobile robots proposed in this invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: In a first embodiment of the present invention, the present invention provides a behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots, such as... Figures 1-8 As shown, it includes the following steps: The obstacle feature extraction module is used to receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of obstacles to determine the motion state of obstacles, and output structured obstacle description information. Furthermore, in the obstacle feature extraction module, the calculation of the obstacle's displacement includes: Establish a sliding window with a preset number of observation frames, and continuously acquire the center position coordinates of the same obstacle in multiple consecutive frames within the sliding window; Calculate the absolute displacement of the center position coordinates between two adjacent frames within the sliding window, and then sum and average the absolute displacements between all adjacent frames to output the displacement of the obstacle.

[0021] Furthermore, in the obstacle feature extraction module, the motion state of the obstacle is determined, and structured obstacle description information is output, including: Configure obstacle type enumeration values ​​in the structured obstacle description information; Based on the displacement of the obstacle, combined with the comprehensive confidence score of the single sensor and the current environmental conditions, the motion state of the obstacle is determined as static, dynamic or suspected false detection. The corresponding judgment result is written as the corresponding status flag into the obstacle type enumeration value to complete the configuration and output of the structured obstacle description information.

[0022] Specifically, the obstacle feature extraction module receives raw perception data from vehicle-mounted environmental perception sensors such as LiDAR and depth cameras. This raw perception data contains 3D point cloud information or depth image pixel coordinates of obstacles in continuous space. To eliminate random noise from single-frame sensor detection and accurately acquire the physical motion characteristics of obstacles, the system constructs a sliding window with a preset number of observation frames N in the obstacle feature extraction module. During the continuous movement of the mobile robot, the system uses a target tracking algorithm to correlate and track the same obstacle in N consecutive frames of data within the sliding window. After each frame of data is processed by a clustering algorithm, the geometric center of the obstacle in the current frame is calculated and mapped to the robot coordinate system of the mobile robot, thereby obtaining the center position coordinates of the obstacle in multiple consecutive frames.

[0023] When calculating the displacement of obstacles, the data processing stream performs inter-frame difference calculations based on the time series within the sliding window. Let the current processing frame within the sliding window be the [number]th frame. The frame, whose corresponding obstacle center position coordinates are The previous frame is the first The coordinates of the obstacle center position corresponding to the frame are The system calculates the Euclidean distance between the center coordinates of these two frames, and uses this as the absolute displacement of the center coordinates between adjacent frames. The calculation formula is as follows: ; in, For the first The x, y, and y coordinates of the obstacle center in the frame. For the first The x, y, and y coordinates of the obstacle center in the frame. Within the entire time period covered by the current sliding window, the system traverses all adjacent frames and calculates the total... The absolute displacements are accumulated and then divided by the number of adjacent intervals. The process involves calculating the mean, ultimately outputting a cumulative average displacement. This cumulative average displacement serves as a quantitative representation of the obstacle's spatial position change. The specific calculation formula is as follows: ; in, The preset number of observation frames for the sliding window is set to a fixed value of 5 here, corresponding to an observation time window of 500ms; For the first The coordinates of the center position of the obstacle in the frame.

[0024] The process of determining the motion state of obstacles and outputting structured obstacle description information relies on the logical mapping of multi-dimensional intermediate parameters and threshold comparison. The system first allocates a structured data storage space in memory to construct the structured obstacle description information. This structured data space contains a fixed-byte obstacle type enumeration field, `ObstacleType`. The system then calculates the displacement of the obstacles... The value is compared with a preset motion detection distance threshold. In this embodiment, the motion detection distance threshold is set to 0.3 meters. If the displacement of the obstacle... If the distance is less than 0.3 meters, the system determines the obstacle's motion state as static. In this case, the system writes the enumerated value representing the static flag into the obstacle type enumeration value field of the structured obstacle description information.

[0025] If the displacement of the obstacle If the distance is greater than or equal to 0.3 meters, the system needs to further incorporate the current moment based on the segmentation judgment logic. Sensor confidence score And perform branch cross-validation on environmental condition data. The system will then use the comprehensive confidence score output by the sensor confidence assessment module at the current moment. Compared with the preset confidence threshold Perform numerical comparisons. Determine the reliability threshold for this procedure. Set to 0.6. If the displacement of the obstacle... Greater than or equal to 0.3 meters, and with the current overall confidence score Greater than or equal to the confidence threshold If the system determines that the obstacle's motion state is dynamic, the system will then write an enumerated value representing the dynamic flag into the structured obstacle description information.

[0026] Conversely, if the above two judgment conditions are not met, the system performs a false detection status verification. When the overall confidence score of the detected condition is... Less than the confidence threshold ; or the number of original feature points of obstacles currently detected within the preset sliding observation window. Below the normal detection point threshold 1.5 times (i.e.) And the characteristic signal disappears within a preset time. The robot disappears within 2 seconds; or the current external operating environment of the mobile robot is detected. When an obstacle is within a pre-defined set of known environmental interference scenarios (such as strong backlight, dusty areas, rainy days, etc.), if any of the above three conditions are met, the system classifies the obstacle's motion state as a suspected false detection. The system then writes an enumerated value representing the suspected false detection flag into the structured obstacle description information to complete the configuration and output of the structured obstacle description information. This structured data is then transmitted via a data interface to the downstream behavior tree obstacle avoidance decision engine, serving as the verification input for condition nodes and participating in subsequent policy degradation routing.

[0027] The sensor confidence assessment module is used to acquire the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of a single sensor and fuse them, and output the fused confidence score. Furthermore, in the sensor confidence assessment module, the calculation of the overall confidence score for a single sensor includes: Acquire the current operating status and environmental condition data of each sensor on the mobile robot; Based on the current operating status and environmental conditions, the data frame rate stability component, the data validity ratio component, and the historical detection consistency component are calculated respectively. The data frame rate stability component, data validity ratio component, and historical detection consistency component are weighted and summed with the environmental attenuation factor determined based on the current interference density to output a single sensor comprehensive confidence score.

[0028] Furthermore, in the sensor confidence assessment module, the overall confidence scores of individual sensors are fused, and the output fused confidence score includes: When equipped with multiple heterogeneous sensors, the weighted geometric average algorithm is used to fuse the comprehensive confidence scores of each individual sensor and output a unified fused confidence score. When any of the sensors is detected to be offline, the default weight of the offline sensor is assigned to the sensor that is currently online.

[0029] Specifically, the sensor confidence assessment module is responsible for continuously evaluating the data quality of each onboard sensing hardware. This module acquires real-time operational status data and external environmental condition data from each sensor on the mobile robot through the underlying data interface. The operational status data includes the underlying communication heartbeat, real-time data throughput, and timestamps of point cloud or image data frames; the environmental condition data includes particle concentration, light level, and precipitation status obtained through onboard environmental sensors or global information.

[0030] In the process of calculating the comprehensive confidence score of a single sensor, the data processing flow first constructs sub-features of the perceived data within the time series window, and calculates the data frame rate stability component, the data validity ratio component, and the historical detection consistency component respectively.

[0031] Data frame rate stability component This is used to evaluate the temporal continuity of sensor data streams. The system constructs a first sliding window of length 10 seconds and performs statistical analysis on the time intervals of the received sensor data frames within this window. Let the average time interval between adjacent frames within this window be denoted as _____. The standard deviation is The formula for calculating the inter-frame stability component is as follows: ; Taking LiDAR as an example, its nominal frame rate is 10Hz, and the average frame interval under normal conditions is... The time interval should be stable at 100 milliseconds. When the actual received frame rate decreases due to insufficient computing resources or communication link jitter in the industrial control computer, the standard deviation of the time interval is... As the value increases, the value of this component decreases linearly.

[0032] Data validity proportion component This system is used to quantitatively characterize the data quality of a single frame of point cloud or image. It constructs a second sliding window with a length of 5 seconds, and counts the total number of points in each frame of point cloud within the window. and the number of valid points. Real-time statistics are performed on the behavior. A valid point is a measurement point whose depth value is neither zero nor null (NaN). The formula for calculating the valid point ratio component is as follows: ; When the lidar is subjected to severe dust or raindrop interference, a large number of laser beams cannot return or produce ineffective reflections, resulting in a decrease in the number of effective points. The value of this component decreased sharply.

[0033] Historical detection consistency component Used to filter out sporadic temporal noise. The system constructs a third sliding window of 10 seconds to count the number of frames in which the obstacle exists continuously and consistently throughout the entire lifecycle of the target. Represents the total number of frames in this window The proportion. The formula for calculating the historical consistency component is: ; If the detected signal undergoes irregular jumps or abrupt changes within a very short period of time, this component will output a lower confidence score.

[0034] Environmental degradation factor This system is used to model the macroscopic negative impacts of harsh external weather or extreme physical environments on the physical detection capabilities of sensors. Based on the acquired environmental condition data, the system extracts the current density of interfering objects. (e.g., dust concentration or rainfall level). The formula for calculating the environmental degradation factor is: ; in, The preset environmental attenuation coefficient is pre-calibrated based on the physical characteristics of the sensing hardware. When the mobile robot enters a non-interfering normal area, the density of interfering objects... Approaching zero, environmental degradation factor Approaching 1.

[0035] The system performs a weighted summation of the three components mentioned above with an environmental attenuation factor, and outputs a single sensor value. At the present moment Single sensor overall confidence score The specific calculation formula is as follows: ; in, These are the weight coefficients for each component, satisfying the constraint that the sum of all weights equals 1. In this embodiment, the fixed weight coefficients are set as follows: .

[0036] When a mobile robot is equipped with multiple heterogeneous sensors (such as simultaneously carrying LiDAR and a depth camera), the system fuses the comprehensive confidence scores of each individual sensor downstream of the data processing stream, and outputs a unified fused confidence score. The system uses a weighted geometric average algorithm to perform fusion calculations, and its specific calculation formula is as follows: ; in, This represents the total number of heterogeneous sensors currently online. The overall confidence score for the lidar is given. The default weights for LiDAR; A comprehensive confidence score for the depth camera. This represents the default weights for the depth camera. In this embodiment, the initial weight allocation value is set to... And the sum of the weights remains 1.

[0037] The system continuously runs an underlying hardware online status monitoring thread within the fusion processing stream. When any sensor is detected to be offline due to hardware power failure, communication bus malfunction, or physical damage, the system initiates dynamic weight reallocation logic. At this time, the system proportionally adds all the default weights of the offline sensors to the currently online sensors to maintain the boundary condition that the sum of the weights of the remaining effective sensors remains 1. For example, when a depth camera is determined to be offline due to a disconnection, its default weight of 0.45 is automatically transferred and added to the currently online LiDAR, making the LiDAR's real-time fusion weight 1.0, thereby increasing the fusion confidence score. Dynamic degradation adjustment is based on the confidence score of a single lidar. Independent decision. The final output is a unified fusion confidence score. The data will be written to the shared memory cache in floating-point form, and then read and compared with the threshold by the condition nodes in the subsequent behavior tree decision engine to drive the branch degradation routing of the obstacle avoidance strategy.

[0038] The behavior tree obstacle avoidance decision engine uses a backtrack node as the root node. Under the backtrack node, multiple policy branches are linked together in descending order of priority. The behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, it reads the structured obstacle description information and fuses the confidence score through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the backtrack node automatically degrades the routing to the next priority policy branch until a policy branch returns a success or running state and outputs an abstract action instruction. Furthermore, in the behavior tree obstacle avoidance decision engine, multiple policy branches are linked together under the backoff node in descending order of priority, including: Under the rollback node, the following branches are connected in descending order of priority: autonomous detour strategy branch, false detection handling strategy branch, waiting and fallback strategy branch, and emergency safety strategy branch.

[0039] Furthermore, the branch for implementing the autonomous detour strategy includes: The condition nodes are used to sequentially verify the nearest distance to the obstacle, the confidence score, the passable space width, and whether the generated detour waypoints point to the avoidance area; When the verification is successful, the action node calculates the coordinates of the temporary detour target waypoint that deviates from the original travel path based on the center coordinates, circumcircle radius and minimum safe distance of the obstacle in the structured obstacle description information, and outputs the coordinates of the temporary detour target waypoint as an abstract action command.

[0040] Furthermore, the branches for implementing false positive handling strategies and waiting / fallback strategies include: When the obstacle type enumeration value is a suspected false detection flag, the false detection handling strategy branch is triggered: save the current trigger detection threshold of the current sensor, increase the number of detection points or the detection confidence requirement, and record the coordinates of the current false detection location; if the false detection signal disappears within the preset monitoring window, restore the threshold and return to the success status; If a false alarm signal persists within the preset monitoring window, the routing is downgraded to the waiting and fallback strategy branch, and the child fallback nodes under it are used to try in sequence: The first-level stationary waiting strategy involves cyclically checking the status of obstacles within a preset time. The second-level slow-speed approach strategy reduces the chassis speed and temporarily lowers the sensor sensitivity when the static obstacle distance condition is met. The third-level global replanning request strategy outputs a global replanning request when the first two strategies fail.

[0041] Furthermore, the implementation of emergency security strategies includes the following branches: The nearest distance to obstacles in the raw perception data and the obstacle type enumeration value in the structured obstacle description information are read using condition nodes; When the nearest distance to an obstacle is less than a preset safety collision threshold, and the obstacle type enumeration value is a dynamic or static flag, the emergency safety strategy branch is triggered. The highest priority emergency braking interruption command is directly output by the action node as an abstract action command, so that the obstacle avoidance action execution module can control the chassis to perform full-brake emergency stop.

[0042] Specifically, the behavior tree obstacle avoidance decision engine, as the core component, is responsible for orchestrating and scheduling multi-level obstacle avoidance strategies in real time on a millisecond-level timescale. The decision engine uses a fallback node as the root node of the entire behavior tree, with multiple policy branches attached to the fallback node, and these branches are organized in descending order of priority. During operation, the behavior tree obstacle avoidance decision engine performs recursive traversal at a preset tick frequency, which is the fixed-period clock step. Clock step The time interval is set to 100 milliseconds, corresponding to a 10Hz spatiotemporal update cycle. In each recursive traversal triggered by a clock step, the engine performs a depth-first search from top to bottom.

[0043] The internal control nodes of the behavior tree update routes according to their specific internal state transition mechanisms. For fallback nodes, their control logic satisfies a logical OR relationship, that is, each child node is evaluated sequentially from left to right in priority order. According to the fallback node state transition rules, its update logic is expressed by the following formula: ; In the formula, For the first The return status of the control node during the next traversal; children is the set of all child nodes of the current control node; For the first The return status of each child node. Once a child branch returns a success or running status, the fallback node immediately stops traversing downwards and returns the same status, thus locking the guidance strategy; only when the current child branch returns a failure status will the fallback node automatically downgrade the route to the next priority strategy branch. In each recursive traversal, the engine reads the structured obstacle description information and fused confidence score from memory in real time through the condition nodes in the leaf nodes. When the condition node fails to verify and returns a failure status, the control logic of the fallback node triggers automatic downgrade routing until the action node of a certain strategy branch is activated and outputs abstract action instructions to the downstream execution-level module.

[0044] Among the multiple policy branches attached to the fallback node, the system connects them in order of priority from high to low as follows: autonomous detour policy branch (Branch1), false detection handling policy branch (Branch2), waiting and fallback policy branch (Branch3), and emergency safety policy branch (Branch4).

[0045] When executing the autonomous detour strategy branch, the data processing flow first enters the sequence node at the very beginning of the branch. Parallel condition nodes then sequentially verify whether the current environmental characteristics meet the detour access boundary. The condition nodes employ the sequence node state transition rules, which satisfy a logical AND relationship: ; Within each branch, the nearest distance (Distance) between the obstacle and the mobile robot is checked sequentially to see if it is greater than 3 meters, and the fusion confidence score is calculated. The sequence node checks whether the obstacle's width (PassableGapWidth) is greater than or equal to 0.6, whether the passable space width on either side of the obstacle is greater than 1.5 meters, and whether the generated waypoint points to the avoidance area specified in the task constraints. When all the above prerequisite nodes pass verification and return a success status, the sequence node activates its lower-level action node. This action node reads the x-coordinate of the obstacle's center from the structured obstacle description information. and ordinate According to the circumcircle radius of the obstacle and the preset minimum safe distance Perform geometric vector projection calculations in the map coordinate system to calculate the coordinates of the temporary detour target waypoints that deviate from the original travel path. The specific calculation formula is as follows: ; ; in, The coordinates of the target waypoint to be bypassed; The coordinates of the obstacle's center in the map coordinate system; Let the radius be the circumcircle radius of the obstacle. To ensure the minimum safe distance, its value is set to meet the following requirements. ; This represents the tangent direction angle of the current global path at the obstacle's location. The sign in the formula is determined by the passable space width, and the direction selection constraint rules follow the formula's limitations: ; Calculated coordinates of the temporary detour target waypoint It will be output as an abstract action command to the chassis control guide line.

[0046] When an obstacle type enumeration value in the structured obstacle description information is detected as a suspected false detection flag, the autonomous detour strategy branch fails to validate, the route automatically degrades, and a false detection handling strategy branch is triggered. The false detection handling strategy branch first saves the current trigger detection threshold configured in memory for the current sensor as a baseline parameter through the action node. Then, it temporarily increases the detection threshold of the sensor that triggers the detection via hardware to reduce perception sensitivity, either by increasing the LiDAR point count threshold to 1.5 times the baseline value or by increasing the camera detection confidence requirement to 0.85. Simultaneously, it records the current false detection location coordinates to the local knowledge base.

[0047] Subsequently, the false detection handling strategy branch initiates a preset monitoring window, which lasts for 3 seconds, corresponding to 30 behavior tree traversal cycles. Within the preset monitoring window, the condition node continuously monitors whether the noise feature disappears. If the noise signal disappears within the 3-second monitoring window, the action node immediately restores the sensor parameters to the original baseline value and returns to a success state, allowing the mobile robot to complete noise filtering without interrupting its navigation mission. If the false detection signal persists for more than 3 seconds within the preset monitoring window, the action node returns to a failure state, triggering the control logic of the fallback node, upgrading the obstacle type to a low-confidence real obstacle, and automatically downgrading the route to the waiting and fallback strategy branch.

[0048] After the routing degrades to the waiting and fallback strategy branch, the system utilizes its lower-level configured sub-backoff nodes to sequentially attempt three levels of sub-strategies in a cascaded order from left to right. The first level is the stationary waiting strategy, which checks whether the obstacle is static and whether the distance is greater than 1 meter through a condition node. When the conditions are met, the chassis braking action is triggered. After the chassis stops, the system starts an internal timer, which cyclically checks the obstacle's movement status within a preset maximum waiting time. If the obstacle is still present after the waiting timeout, the first-level strategy returns to failure, and the sub-backoff node automatically switches to the second-level slow-speed passage strategy. This strategy is triggered in scenarios where static obstacles are far away or low-speed dynamic obstacles are encountered. When the obstacle distance is greater than 5 meters and the movement speed is less than 0.5 m / s, the action node reduces the chassis cruise engine speed to 300 RPM and temporarily increases the LiDAR detection point threshold to 1.5 times the baseline value to adjust the perception sensitivity, controlling the mobile robot to cautiously and slowly pass through the obstacle area. If the slow pass strategy returns to a failure state again, the child backtracking node will be downgraded to the third-level global replanning request strategy, which will directly send a global replanning request to the upper layer. The request data includes the current obstacle position and attribute information, and an alarm notification will be pushed to the operator at the same time.

[0049] When a mobile robot suddenly encounters an obstacle at extremely close range, causing all preconditions of the high-priority policy branches to fail, the behavior tree obstacle avoidance decision engine downgrades the route to the emergency safety policy branch as a last resort. When executing this emergency safety policy branch, the condition node first reads the nearest obstacle distance from the vehicle's raw perception data and the obstacle type enumeration value from the structured obstacle description information. The branch is triggered immediately when the nearest obstacle distance is less than a preset safe distance and the obstacle type enumeration value is determined to be either a dynamic or static flag. The action node of the emergency safety policy branch directly outputs the highest-priority emergency braking interruption command as an abstract action command. Upon receiving this command, the obstacle avoidance action execution module directly sends a zero-speed hard control command to the chassis, switching the system mode to manual obstacle avoidance. Subsequently, it controls the chassis to reverse at a low speed of 200 RPM for 0.5 meters and pushes an emergency obstacle avoidance notification to the operator, thus completing active safety protection for the robot's hardware.

[0050] The obstacle avoidance action execution module is used to receive abstract action instructions and convert them into low-level control commands for the mobile robot control chassis, or into global replanning requests for path replanning of the mobile robot.

[0051] Specifically, the obstacle avoidance action execution module is located at the very end of the entire agent decision-making pipeline. Its task is to convert the software-level, abstract data instructions output by the upstream behavior tree obstacle avoidance decision engine into hardware-level physical action control signals or high-level logic requests. This module monitors the decision engine's instruction buffer in real time. Once it receives a specific abstract action instruction, it converts it into a low-level control command for the mobile robot's control chassis, or a global replanning request submitted to the high-level planning system, based on the content of the abstract action instruction and the current dynamic state of the system.

[0052] During the process of converting and outputting low-level control commands, the data processing stream performs precise parameter calculations and hardware-level protocol encapsulation based on the business logic of different abstract action instructions.

[0053] When the received abstract action command is an emergency braking command (Act_EmergencyStop), it indicates that the system has degraded to a safety fallback state. At this time, the control thread within the obstacle avoidance action execution module will directly call the chassis's underlying bus interface function. The specific command conversion logic is represented by the following formula: ; This action directly generates a set of fixed-byte low-level control commands with both the target linear velocity and the target angular velocity being zero, and sends them to the chassis controller via serial port or industrial bus to control the chassis to perform full-brake emergency stop, while simultaneously switching the system mode to manual obstacle avoidance.

[0054] When the received abstract action command is a low-speed reverse command (Act_BackOff), the data processing stream, after detecting that the chassis has safely come to a stop, initiates a timing state machine with a step size of 500 milliseconds. Within this timing window, the module encodes the reverse command into reverse motion control parameters and outputs specific control commands via the bus. ; The underlying control command drives the left and right servo motor drivers to control the chassis to smoothly reverse 0.5 meters at a fixed reverse speed of 200 RPM. After 500 milliseconds, the braking interruption is automatically triggered to bring the chassis to a stop.

[0055] When the received abstract action instruction is a timeout wait instruction (Act_WaitWithTimeout), the module first sends a low-level control command to the motor brake controller to reset the linear and angular velocities to zero, keeping the mobile robot stationary. Simultaneously, the module's internal timer continuously checks the obstacle status. If the obstacle's features disappear within the maximum waiting time, the chassis resumes its original path. If the obstacle still exists after the timer count reaches the maximum waiting time, the thread sends a timeout failure signal to the behavior tree decision engine to drive automatic route degradation in subsequent policy branches.

[0056] When the received abstract action command is a slow pass command (Act_SlowPass), the module initiates the dynamic parameter reset pipeline of the cruise chassis, overwriting the current base cruise speed parameter CruiseEngine.BaseSpeed ​​of the main drive engine of the cruise chassis with a low-speed limit value, which is set to a fixed parameter of 300 RPM. At the same time, in order to reduce false detection triggers during passage, the module establishes a reverse adjustment loop, temporarily multiplying the real-time detection sensitivity parameter of the LiDAR (i.e., the trigger point threshold LiDAR.PointThreshold) by a coefficient of 0.7, that is, reducing it to 70% of the original baseline parameter. After carefully passing through the obstacle area, each parameter is then reset to the original baseline state.

[0057] When the received abstract action commands are a temporary waypoint generation command (Act_GenerateCircumventWaypoint) and a detour execution command (Act_ExecuteCircumvent), the obstacle avoidance action execution module performs spatial coordinate transformation and dynamic path queue injection. First, based on the current high-precision differential positioning origin coordinates and absolute heading angle, combined with the relative physical orientation of the obstacle relative to the robot, the module calculates the absolute coordinates of the temporary detour target waypoint in the world coordinate system. Then, the module's path management thread injects the calculated temporary target waypoint coordinates into the head of the current task path waypoint queue in a hard-insertion manner, and notifies the cruise engine to reload the waypoint list, driving the chassis control system to track the new local guidance curve and head towards the temporary waypoint.

[0058] When the received abstract action instruction is a replanning request instruction (Act_RequestReplan), the obstacle avoidance action execution module switches to performing cross-system-level protocol data stream serialization conversion. The module sends a replanning event request containing the current obstacle's position and size attributes to the higher-level task planning engine (TaskAgentEngine) through a distributed communication framework. This triggers the higher-level system to call a global path planning algorithm to regenerate a global path that completely bypasses the currently obstructed path, thus completing the closed-open control flow.

[0059] Example 2: When a mobile robot travels along a predetermined path through dusty or rainy areas, the discontinuous physical scattering of the laser beam emitted by the lidar by suspended dust particles or raindrops in the external air causes the sensor confidence assessment module to detect a decrease in the validity ratio of lidar data to 0.45, and more than 50% of the points in the data point cloud appear as invalid depth values. Due to this environmental noise interference, the obstacle feature extraction module continuously detects high-frequency flashing obstacle signals within a very short period of time, causing the system to treat these signals as real obstacles. This leads to the direct triggering of the chassis's emergency stop control flow in industrial control environments where there are no physical obstacles ahead, resulting in unnecessary task interruptions and frequent erroneous robot stops. To solve the above problems, this invention employs a behavior tree-based multi-strategy obstacle avoidance decision-making method for mobile robots, the structure of which is as follows: Figure 9 As shown. The specific implementation process of this method is as follows: S1. Receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of obstacles to determine the motion state of obstacles, and output structured obstacle description information. S2. Acquire the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of the single sensor and fuse them, and output the fused confidence score; S3. Using the behavior tree obstacle avoidance decision engine, with the backoff node as the root node, multiple strategy branches are connected in descending order of priority under the backoff node. The behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, the structured obstacle description information and confidence score are read through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the backoff node automatically degrades the routing to the next priority strategy branch until a certain strategy branch returns a success or running state and outputs an abstract action instruction. S4. Receive abstract action instructions and convert them into low-level control commands for the mobile robot control chassis, or into global replanning requests for mobile robot path replanning.

[0060] Specifically, after receiving the raw sensing data collected by the sensing hardware, if it is affected by physical scattering interference from dust or raindrops, when determining the physical motion state of the obstacle, the system obtains the center position coordinates of multiple frames through a sliding window and verifies them in conjunction with the unified fusion confidence score calculated at the same time. Since the flickering signal meets the temporal characteristic of disappearing within a preset time and is currently in a known environmental interference scenario, the system determines that the motion state of the obstacle is a suspected false detection state. Subsequently, it configures structured obstacle description information in memory, sets the obstacle type enumeration value as a suspected false detection flag, and outputs it.

[0061] During this process, the current operating status and environmental conditions of each sensing hardware are acquired in real time. When the actual data validity ratio component of the LiDAR is detected to drop to 0.45 (i.e., more than 50% of the points in the point cloud are invalid depth values), a weighted sum is calculated by combining the data frame rate stability component, historical detection consistency component, and environmental attenuation factor from the last 10 seconds, to obtain the single-sensor comprehensive confidence score of the LiDAR. Subsequently, a weighted geometric average algorithm is used to fuse the single-sensor comprehensive confidence scores of the LiDAR and camera, outputting a unified fused confidence score. Due to severe data quality degradation, the final calculated unified fused confidence score is lower than the preset confidence threshold of 0.6.

[0062] Next, a depth-first recursive traversal is initiated from the root backdoor node at a preset beat frequency of 100 milliseconds. The structured description information and fusion confidence score are read through condition nodes. The depth-first traversal flow first enters the highest-priority autonomous detour strategy branch. However, because the current obstacle type enumeration value is identified as a suspected false detection (i.e., a fusion confidence score less than 0.6), it fails the non-false detection check and returns a failure status, causing the entire autonomous detour strategy branch to return to a failure status. Upon receiving the failure status, the backdoor node automatically downgrades the route to the next-priority false detection handling strategy branch according to the control logic.

[0063] The condition node at the front end of this branch passes the false detection check, activating the false detection handling sequence: First, the trigger detection threshold of the currently triggered sensor is saved as a baseline value in memory. Then, the detection threshold of the sensor is temporarily increased to reduce the sensing sensitivity (the LiDAR point count threshold is increased to 1.5 times the baseline value), and the GPS coordinates and timestamp of the current noise location are recorded in the local knowledge base. After the adjustment is completed, the control process enters a 3-second preset monitoring window. If the vehicle leaves the interference area and the noise signal disappears within 3 seconds, the detection threshold will be restored to the previously saved baseline value after the false detection signal is cleared and the system will return to the success state. This will output the abstract action command for normal cruise, keep the existing low-level control commands continuously issued, and complete the timing filtering of environmental noise without interrupting the cruise mission.

[0064] If the suspected false alarm signal persists for more than 3 seconds within the preset monitoring window, a failure status is reported to the sequence node, upgrading the obstacle type to a low-confidence real obstacle. After the false alarm handling strategy branch returns to a failure status, the control logic of the fallback node automatically downgrades the route to the next priority waiting and fallback strategy branch, using its sub-fallback nodes to attempt the route sequentially, and outputs abstract action instructions to indicate stopping, slowing down, or replanning.

[0065] Finally, the abstract action command output above is received and converted into a low-level control command by the internal communication protocol encoder to control the chassis to slow down to 300 RPM and reduce the speed of the drive motor; or after a failed attempt to pass slowly, it is converted into a global replanning request containing the current obstacle position and attribute information for sending to the higher level, thereby performing refined hierarchical vehicle control and emergency response.

[0066] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots, characterized in that, include: The obstacle feature extraction module is used to receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of the obstacle to determine the motion state of the obstacle, and output structured obstacle description information. The sensor confidence assessment module is used to acquire the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of a single sensor and fuse them, and output the fused confidence score. The behavior tree obstacle avoidance decision engine uses a backtrack node as the root node. Under the backtrack node, multiple strategy branches are connected in descending order of priority. The behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, the structured obstacle description information and the fusion confidence score are read through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the backtrack node automatically downgrades the routing to the next priority strategy branch until a certain strategy branch returns a success or running state and outputs an abstract action instruction. The obstacle avoidance action execution module is used to receive the abstract action instruction and convert it into the underlying control command of the mobile robot control chassis, or into a global replanning request for replanning the mobile robot path.

2. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 1, characterized in that, In the obstacle feature extraction module, the calculation of the obstacle's displacement includes: Establish a sliding window with a preset number of observation frames, and continuously acquire the center position coordinates of the same obstacle in multiple consecutive frames within the sliding window; Calculate the absolute displacement of the center position coordinates between two adjacent frames within the sliding window, and then sum and average the absolute displacements between all adjacent frames to output the displacement of the obstacle.

3. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 1, characterized in that, In the obstacle feature extraction module, the motion state of the obstacle is determined, and structured obstacle description information is output, including: Configure obstacle type enumeration values ​​in the structured obstacle description information; Based on the displacement of the obstacle, combined with the comprehensive confidence score of the single sensor and the current environmental conditions, the motion state of the obstacle is determined as static, dynamic or suspected false detection. The corresponding judgment result is written as the corresponding status flag into the obstacle type enumeration value to complete the configuration and output of the structured obstacle description information.

4. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 1, characterized in that, In the sensor confidence assessment module, the calculation of the overall confidence score for a single sensor includes: Acquire the current operating status and environmental condition data of each sensor on the mobile robot; Based on the current operating status and environmental conditions data, the data frame rate stability component, the data validity ratio component, and the historical detection consistency component are calculated respectively. The data frame rate stability component, the data validity ratio component, and the historical detection consistency component are weighted and summed with the environmental attenuation factor determined based on the current interference density to output a single sensor comprehensive confidence score.

5. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 1, characterized in that, In the sensor confidence assessment module, the overall confidence scores of individual sensors are fused, and the output fused confidence score includes: When multiple heterogeneous sensors are equipped, a weighted geometric average algorithm is used to fuse the comprehensive confidence scores of each individual sensor and output a unified fused confidence score. When any of the sensors is detected to be offline, the default weight of the offline sensor is assigned to the sensor that is currently online.

6. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 1, characterized in that, In the behavior tree obstacle avoidance decision engine, multiple policy branches are linked together under the backoff node in descending order of priority, including: Under the fallback node, the following branches are connected in descending order of priority: autonomous detour strategy branch, false detection handling strategy branch, waiting and fallback strategy branch, and emergency safety strategy branch.

7. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 6, characterized in that, Executing the autonomous detour strategy branch includes: The condition nodes are used to sequentially verify the nearest distance to the obstacle, the confidence score, the passable space width, and whether the generated detour waypoints point to the avoidance area; When the verification is successful, the action node calculates the coordinates of the temporary detour target waypoint that deviates from the original travel path based on the center coordinates, circumcircle radius and minimum safe distance of the obstacle in the structured obstacle description information, and outputs the coordinates of the temporary detour target waypoint as an abstract action command.

8. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 6, characterized in that, Executing the false detection handling strategy branch and the waiting and fallback strategy branch includes: When the obstacle type enumeration value is a suspected false detection flag, the false detection handling strategy branch is triggered: save the current trigger detection threshold of the current sensor, increase the number of detection points or the detection confidence requirement, and record the current false detection location coordinates; if the false detection signal disappears within the preset monitoring window, restore the threshold and return to the success status; If the false detection signal persists within the preset monitoring window, the routing is downgraded to the waiting and fallback strategy branch, and the child fallback nodes under it are used to try sequentially: The first-level stationary waiting strategy involves cyclically checking the status of obstacles within a preset time. The second-level slow-speed approach strategy reduces the chassis speed and temporarily lowers the sensor sensitivity when the static obstacle distance condition is met. The third-level global replanning request strategy outputs the global replanning request when the first two levels of strategies fail.

9. The behavior tree-based multi-strategy obstacle avoidance decision-making agent for mobile robots according to claim 6, characterized in that, Executing the aforementioned emergency security policy branch includes: The nearest distance to the obstacle in the original perception data and the obstacle type enumeration value in the structured obstacle description information are read using the condition node; When the nearest distance to the obstacle is less than a preset safety collision threshold, and the obstacle type enumeration value is a dynamic flag or a static flag, an emergency safety strategy branch is triggered. The highest priority emergency braking interruption command is directly output by the action node as the abstract action command, so that the obstacle avoidance action execution module can control the chassis to perform full-brake emergency stop.

10. A multi-strategy obstacle avoidance decision-making method for mobile robots based on behavior trees, characterized in that, The method for a behavior tree-based multi-strategy obstacle avoidance decision-making agent for a mobile robot according to any one of claims 1-9 includes: S1. Receive raw perception data collected by various sensors on the mobile robot, calculate the displacement of the obstacle to determine the motion state of the obstacle, and output structured obstacle description information. S2. Obtain the operating status and environmental condition data of each sensor, calculate the comprehensive confidence score of the single sensor and fuse them, and output the fused confidence score. S3. Using a behavior tree obstacle avoidance decision engine, with a fallback node as the root node, multiple strategy branches are connected in descending order of priority under the fallback node; the behavior tree obstacle avoidance decision engine performs recursive traversal at a preset beat frequency. In each traversal, the structured obstacle description information and the fusion confidence score are read through the condition node of the behavior tree. When the condition node returns a failure state, the control logic of the fallback node automatically downgrades the routing to the next priority strategy branch until a certain strategy branch returns a success or running state and outputs an abstract action instruction. S4. Receive the abstract action instruction and convert it into a low-level control command of the mobile robot control chassis, or convert it into a global replanning request for replanning the mobile robot path.

Images