Body robot obstacle avoidance review control method and device and body robot system

Abstract

The present application relates to the technical field of robot autonomous navigation and motion control, in particular to a body robot obstacle avoidance look-back control method, device and body robot system; comprising: collecting real-time kinematics data and forward environment point cloud data through a forward sensor, and keeping a side rear sensor in a basic monitoring state; performing timestamp synchronization, pose solution, constructing a local static map and extracting multi-degree-of-freedom rigid body physical boundary parameters; combining a rear half nonholonomic constraint model to construct a rear half safety envelope line, calculating a rear trajectory nonlinear offset and a rear blind area pose fragile entropy; comparing the fragile entropy with a dynamics danger threshold value, triggering look-back control, reducing speed and acquiring dynamic tail slip data when the threshold value is exceeded, generating independent obstacle avoidance fine tuning parameters and issuing a chassis controller, otherwise maintaining the current forward propulsion speed; the collision probability of the rear half tail can be reduced.

Description

Technical Field

[0001] This invention relates to the field of robot autonomous navigation and motion control technology, specifically to a method, device, and system for obstacle avoidance and retrospective control of an embodied robot. Background Technology

[0002] Obstacle avoidance control technology for embodied robots is a common method to ensure the safe passage of robots in complex environments. It can usually acquire environmental information based on forward sensors and combine it with the chassis motion state to perform path planning, obstacle detection and motion control, so that the robot can complete forward movement, turning and local obstacle avoidance tasks, meet the real-time requirements of the system in mobile operation scenarios and have practical applicability. With the development of security patrols, chemical plant inspections, and narrow passage operations, embodied robots have gradually evolved from obstacle avoidance methods that only focus on forward passability to collaborative control methods that need to consider the space occupied by the robotic arm, rear load compartment, and towing module. However, most existing obstacle avoidance control technologies focus on whether the front path is passable, lacking effective assessment of rear trajectory swing, rear blind spot risks, and dynamic tail-following slip interference. This can easily lead to situations where the front has passed through but the rear swings out and collides during turns, edge crossings, or high-inertia operations, resulting in the inability to meet preset requirements for passage safety and control reliability in complex environments. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides a method, apparatus, and system for obstacle avoidance and retrospective control of an embodied robot. Specifically, the technical solution of this invention is as follows: The obstacle avoidance and replay control method for embodied robots includes: Real-time kinematic data and forward environmental point cloud data are collected by the forward-facing sensor on the embodied robot, while the side and rear sensors on the embodied robot are kept in a basic monitoring state. The real-time kinematic data and the forward environment point cloud data are time-stamped and synchronized. The pose is calculated based on the body coordinate system of the embodied robot to construct a local static map. Multi-degree-of-freedom rigid body physical boundary parameters are extracted from the local static map. Based on the physical boundary parameters of the multi-degree-of-freedom rigid body and the local static map, and combined with the nonholonomic constraint model of the rear half of the embodied robot, a rear half safety envelope is constructed. The nonlinear offset of the rear trajectory between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope is obtained. The rear blind zone pose vulnerability entropy is calculated using the rear trajectory nonlinear offset. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half deviates from the rear half safety envelope. The rear blind zone pose vulnerability entropy is compared with a preset dynamic danger threshold to generate corresponding control mode commands. When the rear blind zone pose vulnerability entropy is greater than or equal to the dynamic danger threshold, the control mode command is to trigger the look-back control mechanism, perform deceleration operation and acquire dynamic tail-following slip data based on the side and rear sensors, generate independent obstacle avoidance fine-tuning parameters based on the dynamic tail-following slip data, and convert the independent obstacle avoidance fine-tuning parameters into a low-level control law and send it to the chassis controller of the robot to adjust the rear half movement trajectory for independent obstacle avoidance fine-tuning. When the rear blind zone pose vulnerability entropy is less than the dynamic danger threshold, the control mode command is to maintain the current forward propulsion speed.

[0004] Optionally, the steps of synchronizing the real-time kinematic data and the forward environmental point cloud data with timestamps, performing pose calculation based on the body coordinate system of the embodied robot, constructing a local static map, obtaining the structural outline data of the embodied robot, and extracting multi-degree-of-freedom rigid body physical boundary parameters from the local static map in combination with the structural outline data include: The distortion correction process is performed on the forward environmental point cloud data to obtain standard point cloud data; The standard point cloud data and the real-time kinematic data are time-stamped and synchronized, and pose calculation and data fusion are performed based on the body coordinate system of the embodied robot to construct a local static map; The front movement trajectory of the embodied robot is extracted from the local static map, and the structural outline data of the embodied robot is combined to calculate the safety margin reserved for the rear half of the body. Based on the front motion trajectory and the safety margin reserved for the rear body, the spatial boundary mapping of the structural outline data is performed to generate the physical boundary parameters of the multi-degree-of-freedom rigid body.

[0005] Optionally, the step of constructing a rear half safety envelope by combining the rear half nonholonomic constraint model of the embodied robot, obtaining the rear trajectory nonlinear offset between the actual rear half trajectory of the embodied robot and the rear half safety envelope, and calculating the rear blind zone pose vulnerability entropy using the rear trajectory nonlinear offset includes: Obtain the nonholonomic constraint model of the rear half of the embodied robot; Combining the nonholonomic constraint model of the latter half and the physical boundary parameters of the multi-degree-of-freedom rigid body, the safety envelope of the latter half is constructed, and the difference between the actual trajectory of the latter half and the safety envelope of the latter half is extracted as the nonlinear offset of the latter trajectory. The nonlinear offset of the rear trajectory is input as an independent variable into a probability distribution function configured with a positive correlation mapping relationship for calculation, thereby obtaining the rear blind zone pose vulnerability entropy that characterizes the magnitude of the deviation probability.

[0006] Optionally, the steps of performing deceleration and acquiring dynamic trailing slip data based on the side and rear sensors, and generating independent obstacle avoidance fine-tuning parameters based on the dynamic trailing slip data, include: After performing the deceleration operation, obtain the available computing power resources of the system controller configured for the android; The available computing resources are allocated to the algorithm module that processes data from the rear-side sensors to perform a backward scan and generate dynamic tail-following slip data. Reconstruct the rear body safety envelope based on the dynamic tail-following slip data; Independent obstacle avoidance fine-tuning parameters are generated based on the reconstructed rear body safety envelope, and independent obstacle avoidance fine-tuning is performed.

[0007] Optionally, the step of generating independent obstacle avoidance fine-tuning parameters based on the reconstructed rear-body safety envelope includes: Extract the independent control degrees of freedom of the mobile chassis and actuators of the embodied robot; Based on the reconstructed rear half safety envelope and combined with the independent control degrees of freedom, the independent obstacle avoidance fine-tuning parameters of the mobile chassis and the actuator are calculated using the control obstacle function with the obstacle avoidance safety distance as the boundary constraint. The independent obstacle avoidance fine-tuning parameters are converted into underlying control laws and sent to the chassis controller to perform independent obstacle avoidance fine-tuning.

[0008] Optionally, before comparing the rear blind zone pose vulnerability entropy with a preset dynamic hazard threshold, the method further includes: Obtain the initial exploration trajectory output by the deep reinforcement learning obstacle avoidance model without smoothing filtering; The initial exploration trajectory is input into a rigid body dynamics safety boundary model constrained by a collision avoidance safety distance for collision verification; The physical boundaries of obstacles in the local static map are updated based on the collision verification results.

[0009] Optionally, the step of acquiring dynamic tail-following slip data based on the side-rear sensor includes: The physical boundary of the rear obstacle at the current moment is obtained based on the side and rear sensor, and the physical boundary of the rear obstacle saved at the previous moment is extracted as the historical physical boundary of the obstacle, and the change of the two is monitored. When the change is greater than or equal to a preset slip tolerance threshold, dynamic tail slip interference is determined to have occurred. The change is used as dynamic tail slip data, and the threshold decay is calculated based on the product of the change and a preset penalty coefficient. The dynamic danger threshold is dynamically lowered by subtracting the threshold decay from the current dynamic danger threshold. When the change is less than the preset slip tolerance threshold, it is determined that no dynamic tail slip interference has occurred, and the dynamic danger threshold is kept unchanged.

[0010] The obstacle avoidance and replay control device for the embodied robot includes: The data acquisition module is used to collect real-time kinematic data and forward environmental point cloud data through the forward sensors mounted on the embodied robot, and to keep the side and rear sensors mounted on the embodied robot in a basic monitoring state. The boundary extraction module is used to timestamp and synchronize the real-time kinematic data and the forward environmental point cloud data, perform pose calculation based on the body coordinate system of the embodied robot, construct a local static map, and obtain the structural outline data of the embodied robot. Combined with the structural outline data, the module extracts the multi-degree-of-freedom rigid body physical boundary parameters from the local static map. The vulnerability entropy calculation module is used to construct a rear half safety envelope based on the multi-degree-of-freedom rigid body physical boundary parameters and the local static map, combined with the rear half nonholonomic constraint model of the embodied robot, to obtain the rear trajectory nonlinear offset between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope, and to calculate the rear blind zone pose vulnerability entropy using the rear trajectory nonlinear offset. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half deviates from the rear half safety envelope. The instruction generation module is used to compare the rear blind zone pose vulnerability entropy with a preset dynamic danger threshold and generate corresponding control mode instructions. The adaptive control module is used to trigger a lookback control mechanism when the rear blind zone pose vulnerability entropy is greater than or equal to the dynamic danger threshold. This mechanism performs deceleration and acquires dynamic tail-following slip data based on the side and rear sensors. It then generates independent obstacle avoidance fine-tuning parameters based on the dynamic tail-following slip data and converts these parameters into a low-level control law, which is then sent to the chassis controller of the robot to adjust the rear half's motion trajectory for independent obstacle avoidance fine-tuning. When the rear blind zone pose vulnerability entropy is less than the dynamic danger threshold, the module maintains the current forward propulsion speed.

[0011] Body-bound robotic systems include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the obstacle avoidance and retrospective control method of the embodied robot according to any one of claims 1 to 7.

[0012] Compared with the prior art, the present invention has the following beneficial effects: In the above scheme, real-time kinematic data and forward environmental point cloud data are collected by the forward-facing sensors on the embodied robot, and the side and rear sensors are kept in basic monitoring state under normal conditions. The synchronized data is then used to calculate the pose to construct a local static map. Multi-degree-of-freedom rigid body physical boundary parameters containing the front motion trajectory and the safety margin reserved for the rear half are extracted. The rear half safety envelope is constructed by combining the rear half nonholonomic constraint model. The rear blind zone pose vulnerability entropy is calculated by using the nonlinear offset of the rear trajectory between the actual rear half trajectory and the rear half safety envelope. When the rear blind zone pose vulnerability entropy reaches the dynamic danger threshold, the look-back control mechanism is triggered to perform deceleration, active side and rear scanning, dynamic tail-following slip recognition, and independent obstacle avoidance fine-tuning. This scheme is particularly suitable for daily patrols in the security industry, safety inspections in chemical plant areas, and narrow passage cornering scenarios. It has the advantages of effectively assessing rear blind zone risks, timely responding to dynamic tail-following slip interference from obstacles, reducing the probability of rearward collisions after the front has passed, and improving traffic safety and control reliability in complex environments. Attached Figure Description

[0013] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the obstacle avoidance and retrospective control method for the embodied robot of the present invention; Figure 2 This is a structural diagram of the embodied robot system of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0015] Example 1: Please see Figure 1 The obstacle avoidance and re-viewing control method for embodied robots includes: Real-time kinematic data and forward environmental point cloud data are collected by the forward sensors mounted on the embodied robot, while the side and rear sensors mounted on the embodied robot are kept in a basic monitoring state. The real-time kinematic data and forward environment point cloud data are time-stamped and synchronized. The pose is calculated based on the body coordinate system of the embodied robot, a local static map is constructed, and the structural outline data of the embodied robot is obtained. Combined with the structural outline data, the physical boundary parameters of the multi-degree-of-freedom rigid body are extracted from the local static map. Based on the physical boundary parameters of the multi-degree-of-freedom rigid body and the local static map, and combined with the nonholonomic constraint model of the rear half of the embodied robot, a rear half safety envelope is constructed. The nonlinear offset of the rear trajectory between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope is obtained. The rear blind zone pose vulnerability entropy is calculated using the rear trajectory nonlinear offset. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half will deviate from the rear half safety envelope. The pose vulnerability entropy of the rear blind zone is compared with the preset dynamic danger threshold to generate corresponding control mode commands. When the pose vulnerability entropy of the rear blind zone is greater than or equal to the dynamic danger threshold, the control mode command is to trigger the look-back control mechanism, perform deceleration operation, and acquire dynamic tail-following slip data based on the side and rear sensors. Based on the dynamic tail-following slip data, independent obstacle avoidance fine-tuning parameters are generated, and the independent obstacle avoidance fine-tuning parameters are converted into the underlying control law and sent to the chassis controller of the embodied robot to adjust the movement trajectory of the rear half of the body for independent obstacle avoidance fine-tuning. When the pose vulnerability entropy of the rear blind zone is less than the dynamic danger threshold, the control mode command is to maintain the current forward propulsion speed.

[0016] This embodiment provides an overall mechanism for obstacle avoidance and retrospective control of embodied robots. Specifically, this mechanism is applicable to scenarios where embodied robots for security patrols or chemical plant inspections perform patrol tasks within complex factory areas: the robot has a tracked or wheeled mobile chassis, a swingable robotic arm, a rear-mounted load compartment, and an optional towing module. The load compartment carries security monitoring equipment or environmental monitoring instruments. The environment contains intersecting pipelines, equipment fences, moving debris, and narrow corners. Due to dust and dense smoke obstructing the view, the robot's forward sensors can detect obstacles ahead, but the rear half of the robot and the end of the robotic arm pose a blind spot risk after turning, which can easily lead to situations where the front has already passed through while the rear experiences a tail-end collision. Specifically, the robot collects real-time kinematic data and forward environmental point cloud data through forward-facing LiDAR, forward-facing depth camera, inertial measurement unit, wheel speedometer, and joint encoder. The real-time kinematic data can include chassis linear velocity, angular velocity, heading angle, robotic arm joint angle, attitude of the rear payload compartment relative to the chassis, and timestamps. The forward environmental point cloud data can include 3D points of the front wall, rubble pile, passage boundary, and passable area. The side and rear sensors can include side and rear short-range LiDAR, ultrasonic array, millimeter-wave radar, or rotatable depth camera. During normal driving, the side and rear sensors do not perform high-frame-rate full scans, but operate in a basic monitoring state, such as retaining only low-frame-rate distance alarms, local sector occupancy detection, or event trigger cache, to reduce computing power and energy consumption. The system synchronizes real-time kinematic data with forward environmental point cloud data according to timestamps and transforms them uniformly into the robot's body coordinate system. The robot's body coordinate system can have the front axis center or chassis geometric center as the origin, the forward direction as the longitudinal axis, the lateral direction as the lateral axis, and the vertical direction as the height axis. Through pose calculation, the system obtains the robot's continuous pose within a preset time window and projects the forward point cloud into this window to form a local static map. This local static map is not a global high-precision map, but a local occupancy representation used for obstacle avoidance decisions within the robot's current control cycle. It records obstacle boundaries, passable corridors, local narrow sections, and safety margins near the robot's rigid body contour. After constructing the local static map, the system extracts multi-degree-of-freedom rigid body physical boundary parameters from the map. These parameters describe not only the chassis outer contour but also the space occupancy of the robotic arm, rear load compartment, or towing module in the current posture. For example, in a simplified two-dimensional plane, the robot can be approximated as a combination of a front half rectangle and a rear half rectangle. The width of the front half is 1.2 meters, the width of the rear half is 1.4 meters, the length of the rear payload compartment is 0.8 meters, and the robotic arm extends an additional 0.15 meters to the right in the current storage posture. If the local static map shows that the distance between the right wall and the center line of the chassis is 0.85 meters, then the available margin on the right side of the rear half is 0.85 meters minus 0.7 meters and then minus 0.15 meters, which is 0 meters, indicating that the rear half is already in a critical state of touching the edge. Based on this, the system constructs the rear half safety envelope by combining the rear half nonholonomic constraint model. The nonholonomic constraint model is used to reflect the motion characteristics of the heavy chassis and rear mechanism, which cannot move laterally instantaneously and cannot arbitrarily change the tail trajectory. For example, when the chassis is moving forward at 0.8 m / s and turning left at 0.35 radians / s, the front center trajectory may bypass the right-side gravel, but the rear half's right rear corner will sweep over the obstacle edge along a more outer trajectory. The system calculates the space zone that the rear half may occupy in the next few control cycles based on the current speed, angular velocity, wheelbase, rear outline, and robotic arm extension. The outer edge of this space zone is the rear half safety envelope. The system acquires the nonlinear offset of the rear trajectory between the actual trajectory of the rear half and the safety envelope of the rear half. This offset is not a simple straight-line distance, but a difference value that takes into account the changes in turning radius, rear swing, robot arm attitude lag and load inertia. To facilitate engineering implementation, multiple key points of the rear half can be selected in each control cycle, such as the right rear corner, left rear corner, robot arm end projection point and the center point of the load compartment tail end. The minimum distance difference between these key points and the safety envelope is calculated respectively, and the point with the greatest risk is taken as the offset of this cycle. As an example, suppose the robot selects three key rear points at a certain moment, denoted as the rear right corner, rear left corner, and outer edge of the robotic arm. The safety envelope requires these points to be at least 0.20 meters away from the obstacle boundary. The measured or predicted distances between the three points and the obstacle are 0.24 meters, 0.31 meters, and 0.14 meters, respectively. The corresponding safety margin differences are +0.04 meters, +0.11 meters, and -0.06 meters, respectively. The system can normalize the 0.06-meter shortfall corresponding to the negative value into the nonlinear offset of the rear trajectory in this cycle. It can also amplify this offset by combining the steering angular velocity, for example, by multiplying it by 1.5 in a high angular velocity state to obtain a risk offset of 0.09 meters. In other words, in this embodiment, the safety margin difference can be expressed as positive or negative to indicate excess or deficiency, but the nonlinear offset of the rear trajectory actually used for subsequent vulnerability entropy calculation is uniformly expressed as a non-negative risk quantity, and the positive margin is not directly written into the probability mapping calculation; the system further uses this offset to calculate the pose vulnerability entropy of the rear blind zone; this evaluation value is used to characterize the probability or risk intensity of the actual trajectory of the rear half deviating from the safety envelope; in implementation, the offset can be input into a monotonically increasing probability mapping function; for example, when the offset is 0 meters, the vulnerability entropy is 0.2; when the offset is 0.05 meters, the vulnerability entropy is 0.55; when the offset is 0.10 meters, the vulnerability entropy is 0.82; this mapping is not required to be limited to a specific formula, as long as the larger the offset, the higher the evaluation value; The system can also incorporate speed, load mass, and turning angular velocity as weighting factors to make heavy-load high-speed turning a higher risk value than low-speed straight-line driving; then, the system compares the rear blind zone pose vulnerability entropy with the preset dynamic hazard threshold; the dynamic hazard threshold can be set according to robot size, load level, environmental risk level, and task mode. For example, the threshold can be set to 0.75 in normal inspection mode, 0.60 in dangerous goods transportation mode, and 0.50 in high-risk collapse areas. When the rear blind zone pose vulnerability entropy is greater than or equal to this threshold, the system generates a control mode command to trigger the lookback control mechanism. When the rear blind zone pose vulnerability entropy is less than this threshold, the system generates a control mode command to maintain the current forward propulsion speed. When the lookback control mechanism is triggered, the system prioritizes deceleration, such as reducing the forward speed from 0.8 m / s to 0.35 m / s, or smoothly reducing the speed according to the maximum allowable deceleration to avoid impact on the items in the load compartment due to sudden stop. After deceleration, the side and rear sensors switch from basic monitoring to active detection to acquire dynamic tail-following slip data. This data represents the amount of sliding, tilting, or intrusion of the obstacle boundary relative to the historical map after the robot passes over it. Based on this data, the system generates independent obstacle avoidance fine-tuning parameters, such as reducing the differential speed of the right rear wheel, decreasing the speed of the left track, retracting the robotic arm inward by 0.12 meters, and fine-tuning the chassis heading angle by 2 degrees. These parameters are then converted into low-level control laws and sent to the chassis controller to converge the rear half of the robot's trajectory towards a safe area. If the rear blind spot pose vulnerability entropy is lower than the dynamic danger threshold, the system does not trigger backtracking control but maintains the current forward propulsion speed while continuing basic monitoring by the side and rear sensors. This avoids frequent deceleration and backtracking when there is sufficient safety margin, ensuring the stability of the rescue mission's progress. In abnormal situations, such as timestamp synchronization failure, missing frames in the forward point cloud, or abnormal values ​​returned by the side and rear sensors, the system will not directly use abnormal data to generate high-speed passage instructions. A data validity flag can be set: if critical data is unavailable for two consecutive control cycles, the rear blind zone pose vulnerability entropy will be temporarily increased to a preset safety avoidance value, such as 0.8, and a low-speed review or safe stop will be triggered. If only a single sensor fails, the most recent effective map, wheel speedometer trajectory, and inertial measurement data will be used for short-term extrapolation, while limiting the maximum speed and maximum steering angular velocity. For boundary cases where the vulnerability entropy is exactly equal to the dynamic danger threshold, the system will process it as a high-risk branch to reduce misjudgments caused by critical jitter. In the security patrol corridor of a chemical plant, the robot navigates around complex pipe corridors. Forward-facing lidar identifies a pile of debris to its right and plans a left turn to avoid it. After the front chassis safely passes, the system predicts, based on the length of the rear load compartment and the robotic arm's storage position, that the right rear corner will approach the edge of the debris. For ease of engineering implementation, a formula is used... The calculated nonlinear offset of the rear trajectory is 0.08 meters, corresponding to a pose vulnerability entropy of 0.72 in the rear blind zone. Since the current mission is in the chemical plant area inspection mode, the dynamic danger threshold is 0.60. The system immediately reduces the speed and starts the side and rear back scan. It is found that the edge of the accumulation has slid into the channel relative to the historical map by 0.06 meters. Therefore, the fine-tuning control law for suppressing the right rear outward swing is issued, so that the rear half passes through with a smaller outward swing radius. The purpose of this step is to expand the traditional obstacle avoidance control, which only focuses on whether the path in front is passable, into a closed-loop control that simultaneously manages the risk of rear space occupancy. This enables the robot to automatically switch between forward speed and rear running safety based on the pose vulnerability entropy of the rear blind spot under conditions of high inertia, narrow passage and blind spot interference, thereby reducing the probability of tail swing-out collision. Furthermore, to ensure consistency in terminology throughout the document, in subsequent embodiments of this specification, the rear blind zone pose vulnerability entropy can be simply referred to as vulnerability entropy. Both refer to the same assessment value used to characterize the probability or risk intensity of the actual trajectory of the rear half deviating from the rear half safety envelope. The rear half safety envelope can be simply referred to as the safety envelope, but unless otherwise specified, it refers to the safety boundary established facing the rear half sweep area. The back-look control mechanism, side-rear back scan, backward scan, active scan, and active detection state, unless otherwise ambiguous, all refer to the same control stage where the side-rear sensor switches from basic monitoring to active sensing after the vulnerability entropy reaches the danger threshold. The standard correspondence is still based on the trigger back-look control mechanism and the dynamic tailing slip data obtained from the side-rear sensor in the embodiments. By defining the terminology above, it is possible to avoid misunderstanding different expressions as independent control objects or different algorithm processes. The steps of synchronizing real-time kinematic data and forward environment point cloud data with timestamps, calculating pose based on the embodied robot's body coordinate system, constructing a local static map, and extracting multi-degree-of-freedom rigid body physical boundary parameters from the local static map include: Standard point cloud data is obtained by performing distortion correction processing on the forward environmental point cloud data; The standard point cloud data and real-time kinematic data are time-stamped and synchronized. The pose calculation and data fusion are performed based on the body coordinate system of the embodied robot to construct a local static map. The front movement trajectory of the embodied robot is extracted from a local static map, and the safety margin for the rear half is calculated by combining the structural outline data of the embodied robot. Based on the front motion trajectory and the safety margin reserved for the rear half, the spatial boundary mapping of the structural outline data is performed to generate multi-degree-of-freedom rigid body physical boundary parameters.

[0017] This embodiment provides a mechanism for constructing a local static map and extracting physical boundary parameters. Specifically, when there is dust, vibration, and high-speed turning in the patrol passage of a chemical plant, directly superimposing the original point cloud collected by the forward sensor onto the map will cause the wall boundary to be stretched and the obstacle outline to appear as a ghost image, resulting in an incorrect estimate of the rear safety margin. The aforementioned overall solution can complete the basic review trigger when the point cloud is not corrected, but when the robot turns at high speed close to the wall, the distortion of the original point cloud will cause the safety envelope to produce a positive or negative deviation that exceeds the preset error range. Therefore, this embodiment introduces point cloud distortion correction, timestamp synchronization, and boundary parameter extraction processes. Specifically, the system first performs distortion correction on the forward environment point cloud data to obtain standard point cloud data. Distortion sources can include scanning time differences caused by robot movement, point position offsets caused by sensor vibration, and outliers generated by the depth camera in dust. During correction, the system can use the sampling time of each point of the LiDAR, combined with the short-term pose changes provided by the inertial measurement unit and wheel speedometer, to uniformly compensate the points in the same frame of point cloud to the same reference time. For example, if a frame of point cloud is continuously collected for 100 milliseconds, during which the robot moves forward 0.08 meters and turns 2 degrees to the left, without correction, the right wall will appear as a slanted shadow on the map. After correction, the wall points are uniformly transformed to the end time of the frame or the middle time of the frame to form standard point cloud data. The system timestamps the standard point cloud data with the real-time kinematic data. Synchronization can be achieved using nearest neighbor time matching, interpolation matching, or buffer queue matching. For ease of understanding, assuming the timestamp of the standard point cloud frame is 10.00 seconds and the wheel velocity meter data timestamps are 9.98 seconds and 10.02 seconds, the system can perform linear interpolation on the two sets of wheel velocity meter data to obtain the linear velocity and angular velocity corresponding to 10.00 seconds. If the inertial measurement unit has multiple high-frequency data points around 10.00 seconds, the average or integral result of a short window containing that moment can be used to reduce instantaneous noise. After synchronization, the system performs pose calculation and data fusion based on the embodied robot's body coordinate system to construct a local static map. The local static map can take the form of a grid map, a point cloud sub-map, or an occupancy distance field. As a simplified example, the area 3 meters in front of the robot and 1.5 meters to the left and right can be divided into a grid of 3 rows and 5 columns. A grid value of 1 indicates that the area is occupied by an obstacle, 0 indicates that the robot is passable, and an unknown area is represented by 0.5. If the standard point cloud shows that there are loose rocks in the two squares to the right and in front, the map can be represented as: the first row second line The third line The robot chooses to travel on the left side of the map, but the range of its rear swing needs to be further determined in conjunction with the rear safety margin. The system extracts the front motion trajectory and the rear safety margin from the local static map. The front motion trajectory can be a sequence of trajectories of the front center point of the chassis or the center point of the front wheel axle over several future control cycles, for example, with coordinates all located in the robot's body coordinate system. The safety margin reserved for the rear half can be the remaining distance from the key point of the rear half to the nearest obstacle boundary, such as 0.18 meters for the right rear corner, 0.42 meters for the left rear corner, 0.12 meters for the outer edge of the robotic arm, and 0.20 meters for the tail load compartment. The system uses the front motion trajectory and the safety margin reserved for the rear half together as the physical boundary parameters of the multi-degree-of-freedom rigid body, so that the subsequent risk assessment not only depends on the passable area in front, but also reflects the spatial constraints of the rigid body structure in the rear half. As a fault-tolerance mechanism, if there are still outliers exceeding a preset threshold after point cloud distortion correction, the system can set an outlier threshold. For example, if more than 30% of the points in the standard point cloud have abrupt changes in distance from their surrounding points, the frame is marked as a low-confidence frame. Instead of directly updating the obstacle boundary, the system uses the previous frame's reliable map and reduces the speed. If the timestamp difference exceeds the allowed synchronization window, for example, if the difference between the point cloud time and the kinematic data time exceeds 80 milliseconds, the system will not perform forced fusion to avoid incorrectly binding the pose at different times to the obstacle position. If the calculated result of the rear half safety margin is negative, it indicates that the current predicted trajectory has invaded the obstacle boundary. The system will directly enter the replay or stop branch without waiting for the pose vulnerability entropy in the rear blind area to accumulate further. The robot makes a left turn around a section of intersecting pipelines in the security inspection corridor of a chemical plant. Due to chassis vibration, the pipeline boundaries in the original point cloud of the forward radar show multiple false boundary outlines. The system performs distortion correction on the point cloud based on the inertial attitude and wheel speed interpolation over 10.00 seconds, merging the ghosting into a stable boundary. The system extracts the front trajectory, showing that the front wheels will pass through the left side of the pipeline. At the same time, it calculates that the safety margin reserved at the right rear corner of the rear half is only 0.11 meters, and the safety margin reserved at the outer edge of the robotic arm is 0.09 meters. Therefore, these values ​​are included in the subsequent vulnerability entropy calculation. The purpose of this step is to reduce the boundary errors caused by point cloud trailing and time mismatch by fusing standard point cloud, synchronous kinematics, and body coordinate system. This allows the safety margin reserved at the rear half to truly reflect the passability of a multi-degree-of-freedom rigid body in local space, thereby improving the reliability of the retrospective trigger judgment. The steps for constructing a rear-body safety envelope by combining the rear-body nonholonomic constraint model of the embodied robot, obtaining the nonlinear offset of the rear trajectory between the actual trajectory of the rear body and the rear-body safety envelope, and calculating the pose vulnerability entropy of the rear blind zone using the nonlinear offset of the rear trajectory include: Obtain the nonholonomic constraint model of the latter half of the embodied robot; By combining the nonholonomic constraint model of the latter half and the physical boundary parameters of the multi-degree-of-freedom rigid body, a safety envelope of the latter half is constructed, and the difference between the actual trajectory of the latter half and the safety envelope of the latter half is extracted as the nonlinear offset of the latter trajectory. The nonlinear offset of the rear trajectory is input as an independent variable into a probability distribution function configured with a positive correlation mapping relationship for calculation, thereby obtaining the rear blind zone pose vulnerability entropy that characterizes the magnitude of the deviation probability.

[0018] This embodiment provides a mechanism for constructing the rear half safety envelope and calculating the pose vulnerability entropy of the rear blind zone. Specifically, the aforementioned local static map can provide obstacle boundaries and safety margins, but if the collision is judged solely based on the robot's outer contour at the current moment, the lag outward swing and non-lateral movement constraints of the rear half of the heavy robot during the turning process will be ignored. Especially in narrow corners, the front trajectory and the rear trajectory do not coincide, and the passage of the front does not necessarily mean that the rear will pass. Therefore, this embodiment introduces a rear-part nonholonomic constraint model to enable the prediction of rear trajectory risks based on the motion state. Specifically, the system acquires the robot's rear-part nonholonomic constraint model. This model can be composed of chassis geometric parameters, rear axle position, load compartment size, towing articulation point, robotic arm posture, and wheel-ground constraints. For differential chassis, the model can describe the inability of the chassis to move laterally, and the speed of the key points of the rear body is determined by the chassis's linear velocity and angular velocity. For robots with towing modules, the model can also describe the upper limit of the towing module's swing angle change relative to the chassis. For robots with robotic arms, the model can also describe the relative outward expansion boundary of the robotic arm when the chassis turns. The system combines the rear-part nonholonomic constraint model and the physical boundary parameters of the multi-degree-of-freedom rigid body to construct the rear-part safety envelope. In practice, the predicted positions of the key points of the rear body can be generated within a prediction window of 1 or 2 seconds with a fixed control cycle, and the outer edges of these key points can be connected to form a spatial envelope. As a simplified example, the robot moves at 0.6 m / s with an angular velocity of 0.3 radians / s, a prediction window of 1 second, and sampling points at 0, 0.5, and 1 second. The system predicts that the distances of the right rear corner relative to the obstacle boundary in the body coordinate system are 0.22 m, 0.16 m, and 0.10 m, respectively. However, the required safe distance is 0.18 m, resulting in insufficient distances at 0.5 and 1 seconds. After using the boundary line that meets the safe distance as the safe envelope, the predicted trajectory of the right rear corner deviates outward from the safe envelope at the last two sampling points. The system extracts the difference between the actual trajectory of the rear half and the safe envelope of the rear half as the nonlinear offset of the rear trajectory. The actual trajectory can be obtained from recent odometry playback, kinematic extrapolation of rear key points, or observation from side and rear sensors. The difference can be combined according to key points, time points, and directions. For example, the right rear corner is 0.02 meters less than the safety line at 0.5 seconds and 0.08 meters less at 1 second; the outer edge of the robotic arm is 0.04 meters less at 1 second. The system can select the maximum intrusion of 0.08 meters as the offset, or it can weight the offset according to the importance of key points. For example, the rear of the load compartment has a higher weight than ordinary shell points, and the end of the robotic arm has a higher weight than the chassis side plate. After obtaining the offset, the system uses the offset as an independent variable to calculate the rear blind zone pose vulnerability entropy using a probability distribution function configured with a positive correlation mapping relationship. This probability mapping function can use a lookup table function, a piecewise linear function, or a smooth increasing function. As an example, the following mapping can be set: when the offset is less than 0.02 meters, output 0.30; when the offset is 0.05 meters, output 0.55; when the offset is 0.08 meters, output 0.72; when the offset is greater than 0.12 meters, output 0.90 or higher; if the chassis angular velocity is detected to exceed 0.4 radians / second, the output value can be increased by 0.05 to 0.10; if the load is of a high sensitivity level, a safety factor can be further increased; in this way, the vulnerability entropy can reflect the combined effects of insufficient distance and dynamic instability trend. In abnormal situations, if the parameters of the non-holonomic constraint model in the latter half are missing, such as the towed module not reporting the hinge angle, the system can use a conservative model as a substitute, setting the towed module's swing angle to the maximum allowable swing angle to avoid underestimating the risk; if the offset crosses the danger threshold repeatedly within multiple cycles, the system can set a hysteresis interval, for example, a trigger threshold of 0.60 and a release threshold of 0.50, and only exit the review state after being continuously below 0.50 for several cycles, to prevent control jitter caused by frequent switching; if the input of the probability mapping function exceeds the preset range, the output will be limited to between 0 and 1; if the offset is an invalid value, it will be treated as a high-risk value and the speed will be limited. Within the right-angled passageway of the chemical plant area, the robot's front section avoids the debris on the right side by turning left. The local map shows that the center trajectory of the front section has a margin of 0.35 meters, but the model predicts that the rear right corner will swing outward to the right after 1 second. The system constructs a safety envelope based on the width of the rear load compartment and the current angular velocity, and finds that the predicted trajectory of the rear right corner is 0.08 meters off from the safety line. After inputting this offset into the positive correlation mapping function, the rear blind zone pose vulnerability entropy is 0.72, which exceeds the threshold in the security patrol mode, and therefore enters the retrospective control process. The purpose of this mechanism is to separate the incomplete motion characteristics of the rear half from the static outer contour judgment, enabling the system to identify the risk of a tail-flip that is passable in the front but not in the rear in advance, and to provide a unified trigger basis for subsequent deceleration, backsweep and fine-tuning through the vulnerability entropy value. Furthermore, in order to avoid understanding the nonlinear offset of the rear trajectory and the vulnerability entropy of the rear blind zone pose as abstract quantities lacking clear extraction rules, the extraction rules of the offset can be further structured in this embodiment as follows: first, select a set of key points of the rear half within the prediction window, and then calculate the insufficiency of each key point relative to the safety envelope; a deficiency of zero indicates that the key point is still within the safety envelope, and a positive deficiency indicates that the key point has entered the risk area. During project implementation, the system can be based on the first The key point is the actual or predicted distance from the obstacle boundary at the corresponding sampling time. And the corresponding safety distance requirements for this key point. By calculating the non-negative maximum value of the difference between the required safety distance and the actual or predicted distance at each key point, the offset for this period is extracted uniformly. That is, adopt Perform calculations; among which, The sequence number of the selected key points in the posterior body, for example ,in, The total number of key points. This indicates all selected key points. Take the maximum value. This indicates taking the larger value among the terms within the parentheses; the constant 0 in the formula represents the non-negative threshold of the difference between the safety distance and the distance variable, and its dimension is consistent with that of the distance parameter; that is, when When the threshold is reached, it means that the critical point has not intruded into the safety envelope area, and the local offset risk of the point is recorded as 0. Through this rule, the offset always represents the degree of intrusion of the most dangerous critical point into the safety envelope, avoiding the mixing of safety margin and risk offset; further, the aforementioned actual trajectory of the rear half can be understood as the actual effective trajectory used for control determination within the current control cycle. This actual effective trajectory can be either the playback trajectory obtained from recent odometer playback and observation by the side and rear sensors, or the short-term execution trajectory extrapolated from the current kinematic state when the prediction window has not yet been fully completed at the current moment. In other words, during the risk prediction phase before the trigger, the system allows the use of a combination of recent real trajectory and short-term kinematic extrapolation trajectory as the actual trajectory of the latter half of the body, so as to ensure that real-time evaluation can be completed within the current control cycle, rather than waiting for the entire physical motion to end before making a post-event judgment; furthermore, the aforementioned probability distribution function is preferably implemented in engineering by using a preset calibration table or piecewise interpolation method; Specifically, a one-to-one correspondence between offset intervals and vulnerability entropy intervals can be pre-saved. For example, 0 meters to 0.02 meters corresponds to a low-risk segment, 0.02 meters to 0.08 meters corresponds to a medium-risk segment, and above 0.08 meters corresponds to a high-risk segment. When the input offset is between adjacent calibration points, the system uses linear interpolation to obtain continuous output. This preserves the openness of the positive correlation mapping relationship in the embodiment and also makes the calculation rules have a clear data flow process. If weighting factors such as angular velocity and load level are introduced at the same time, it is preferable to perform bounded correction after completing the basic mapping, and then restrict the output range to between 0 and 1 again after correction to avoid out-of-range results after multiple weighting. Furthermore, to ensure consistency with the examples in other implementations, whenever the text describes a negative safety margin difference as a certain amount of intrusion or outward deviation from the safety line, unless otherwise specified, it should be understood as an intermediate description used to obtain the same nonlinear offset of the rear trajectory. The quantity that is actually input to the probability distribution function is uniformly the non-negative offset extracted according to the above rules. Thus, the nonlinear offset of the rear trajectory in the text always uniquely corresponds to the degree of risk intrusion of the rear key point to the safety envelope in a physical sense, and no longer refers to both the signed distance difference and the unsigned risk quantity at the same time, thereby avoiding the same term having multiple meanings in different paragraphs. The steps for performing deceleration and acquiring dynamic trailing slip data based on side and rear sensors, and generating independent obstacle avoidance fine-tuning parameters based on the dynamic trailing slip data, include: After performing the deceleration operation, obtain the available computing power resources of the system controller configured for the embodied robot; Allocate available computing resources to the algorithm module that processes data from the rear-side sensors to perform a backward scan and generate dynamic tail-following slip data. Reconstructing the rear half safety envelope based on dynamic tail-following slip data; Independent obstacle avoidance fine-tuning parameters are generated based on the reconstructed rear body safety envelope, and independent obstacle avoidance fine-tuning is performed.

[0019] This embodiment provides a mechanism for allocating computing power, backward scanning, and reconstructing the safety envelope after triggering a lookback. Specifically, the aforementioned vulnerability entropy calculation can determine whether the risk of the rear half has increased. However, in chemical plants or security patrol scenarios, the source of increased risk may not be the robot's own trajectory planning error, but rather the dynamic tailing and slippage of obstacles after the robot passes, such as the slide of accumulated materials, the tilting of material edges, or pipelines bouncing back into the channel. If the static map previously collected by the forward sensor is still used, the rear half safety envelope will be generated based on outdated boundaries and cannot reflect the changed physical environment in the blind spot behind. Therefore, in this embodiment, computing power is reallocated after deceleration to actively scan the side and rear regions. Specifically, after the deceleration operation is performed, the system obtains the available computing power resources of the controller. These computing power resources can be represented by processor utilization, available inference queue of graphics processing unit, sensor bus bandwidth, or real-time task scheduling window. The deceleration operation itself will reduce the frequency pressure of forward path planning, enabling some resources to switch from high-frequency forward exploration to side and rear perception. For example, before deceleration, forward point cloud processing occupies 70% of computing resources, chassis control occupies 15%, side and rear basic monitoring occupies 5%, and the remaining 10% is used; after deceleration, forward point cloud processing can be reduced to 45%, freeing up 25% of resources for side and rear point cloud reconstruction and obstacle boundary tracking; the system allocates available computing resources to the algorithm module that processes side and rear sensor data to perform backward scanning; backward scanning may include increasing the frame rate of side and rear lidar, adjusting the orientation of the gimbal depth camera, expanding the ultrasonic array sampling sector, or enabling near-range target tracking of millimeter-wave radar; the scan results are compared with historical obstacle boundaries to generate dynamic trailing slip data; As a simplified example, in the historical map, the distance between the right rear edge of the debris and the robot's centerline is 0.82 meters. The current side-rear scan shows this distance to be 0.74 meters, indicating the obstacle intrudes 0.08 meters into the robot's path. This 0.08 meters can be used as part of the dynamic tail-following slip data, recording the slip direction as lateral intrusion, the occurrence area as the right rear edge, and a confidence level of 0.86. Based on the dynamic tail-following slip data, the system reconstructs the rear half safety envelope. During reconstruction, the system does not need to completely rebuild the global map, but rather updates the obstacle boundary and safety distance within the predicted sweep area of ​​the rear half. For example, if the original safety envelope allowed the right rear corner to pass at a position 0.70 meters from the centerline, and the obstacle boundary slips from 0.82 meters to 0.74 meters, if a safety distance of 0.18 meters is still required, the right rear corner's allowable boundary shrinks to 0.56 meters. Based on this, the system obtains a rear half safety envelope that shrinks inward and has a larger collision avoidance margin, and recalculates the offset of the rear half key points. The system generates independent obstacle avoidance fine-tuning parameters based on the reconstructed rear half safety envelope. These parameters do not involve replanning a completely different global route. Instead, they make local corrections to the chassis, rear wheels, track differentials, robotic arm retraction posture, or towing angle while maintaining the overall direction of the rescue mission. For example, if an obstacle intrudes into the right rear, the system can generate differential parameters of 0.30 m / s for the left track and 0.24 m / s for the right track to reduce tail swing. Simultaneously, the robotic arm retracts inward by 0.10 meters, and the rear payload compartment remains locked in position. If the fine-tuning still fails to meet the safety envelope requirements, the system enters a more forceful stopping or reversing strategy. As a fault-tolerant mechanism, if the available computing power is still insufficient after deceleration, for example, when the controller is handling an emergency stabilization control task, the system prioritizes chassis attitude control and braking safety and does not forcibly start high-load retrace. At this time, it can adopt low-speed conservative passage or stop in place to wait for computing power to recover. If the reliability of the side and rear scan results is lower than the preset value, for example, dust causes the depth camera to have a point cloud missing area with a size exceeding the preset area threshold, the system can fuse the distance alarms of ultrasonic and millimeter-wave radar and update the rear half safety envelope according to the minimum distance principle. If the dynamic tail-following slip data differs too much from the historical map, for example, a sudden environmental structural change that exceeds the limit occurs, the system will not perform minor adjustments, but will trigger an emergency stop and report to the task control center. After the robot completed obstacle avoidance at the front of the pipe corridor in the chemical plant, the system calculated that the pose vulnerability entropy of the rear blind zone exceeded the threshold, so the speed was reduced from 0.75 m / s to 0.30 m / s. After the speed reduction, the controller released about 20% of the computing power, the side and rear depth camera turned to the right rear and increased the sampling frequency. The scan results showed that the pile of gravel that had just been swept by the front of the robot slid into the passage by 0.07 meters, intruding into the reserved margin of the rear load compartment. Based on this, the system reconstructed the rear half safety envelope and generated fine-tuning parameters to reduce the tail swing and the inward retraction of the robotic arm by the chassis differential. The purpose of this mechanism is to switch computing resources from forward speed-first mode to backward survival safety-first mode after a risk is triggered. By actively scanning and correcting outdated maps from the side and rear, the rear half safety envelope can reflect the real spatial boundary after dynamic tailing and sliding, thereby improving the effectiveness of independent obstacle avoidance fine-tuning. The steps for generating independent obstacle avoidance fine-tuning parameters based on the reconstructed rear-body safety envelope include: Extract the independent control degrees of freedom of the embodied robot's mobile chassis and actuators; Based on the reconstructed rear half safety envelope and combined with the independent control degrees of freedom, the independent obstacle avoidance fine-tuning parameters of the mobile chassis and actuators are calculated using the control obstacle function with the obstacle avoidance safety distance as the boundary constraint. The independent obstacle avoidance fine-tuning parameters are converted into underlying control laws and sent to the chassis controller to execute independent obstacle avoidance fine-tuning.

[0020] This embodiment provides an obstacle avoidance fine-tuning mechanism based on independent control degrees of freedom and control obstacle function. Specifically, the aforementioned scheme can reconstruct the safety envelope of the rear half of the robot by backscanning. However, if the speed of the entire vehicle is reduced uniformly, it may not be able to effectively change the outward swing trajectory of the rear half in a narrow space. For example, if the robot has entered the middle of a corner, simple deceleration can only delay the collision and cannot move the outer edge of the robotic arm or the tail corner of the load compartment away from the obstacle. Therefore, this embodiment separates the controllable degrees of freedom of the mobile chassis and the actuator, and uses a control obstacle function constrained by the obstacle avoidance safety distance to calculate fine-tuning parameters, enabling different mechanisms to make local responses to rear-body risks. Specifically, the system extracts the independent control degrees of freedom of the robot's mobile chassis and actuator. The degrees of freedom of the mobile chassis may include the speed of the left wheel, the speed of the right wheel, the linear velocity of the chassis, the angular velocity of the chassis, the differential speed of the tracks, or the angle of the steering wheel. The degrees of freedom of the actuator may include the shoulder joint, elbow joint, wrist joint of the robotic arm, the end-effector retraction posture, the locking angle of the rear load compartment, or the articulation damping of the towing module. The system determines which degrees of freedom can participate in obstacle avoidance based on the current task status. For example, if the robotic arm is holding critical equipment and cannot swing significantly, it is only allowed to retract the robotic arm inward by 5 degrees. If the towing module has a sensitive load, the articulation angular velocity is limited to avoid rapid swinging. Based on the reconstructed rear half safety envelope, the system calculates independent obstacle avoidance fine-tuning parameters using a control obstacle function constrained by the obstacle avoidance safety distance. The control obstacle function is configured to satisfy the following constraint: that is, within the current control cycle, the sum of the Lie derivatives generated by each independent control degree of freedom of the moving chassis and actuator is greater than or equal to the negative gradient of the risk margin function formed by the difference between the current actual distance and the obstacle avoidance safety distance, thereby ensuring that the generated fine-tuning parameters are strictly limited to the feasible space where no collision occurs. Specifically, the control barrier function satisfies the following mathematical constraints: ,in, The risk margin function is constructed from the difference between the actual distance and the obstacle avoidance safe distance. and These are the system's uncontrolled drift vector field and controlled input vector field with respect to... Li Daoshu, These are independent obstacle avoidance fine-tuning parameters that include each independent control degree of freedom. For monotonically increasing extended class K functions; As an example, assuming the reconstructed safety distance requirement is 0.18 meters, the current predicted distance to the right rear corner is 0.14 meters, the predicted distance to the outer edge of the robotic arm is 0.16 meters, and the left-side space margin is 0.35 meters; the system extracts three controllable quantities: chassis angular velocity correction, right track speed correction, and robotic arm retraction angle; the obstacle control function determines that the right-side distance is insufficient, so it restricts the chassis from continuing to swing outward to the right, correcting the angular velocity from 0.30 radians / second to 0.22 radians / second, the right track speed from 0.32 meters / second to 0.28 meters / second, and retracting the robotic arm inward by 6 degrees; if these fine-tunings increase the predicted right rear corner distance to 0.19 meters and the robotic arm outer edge distance to 0.20 meters, then this set of parameters is accepted; the system converts the independent obstacle avoidance fine-tuning parameters into a low-level control law and sends it to the chassis controller for execution; the conversion process may include speed command limiting, acceleration smoothing, joint angle interpolation, control cycle alignment, and safety verification; For example, the high-level fine-tuning parameters are: chassis linear velocity 0.30 m / s, angular velocity 0.22 radians / s, and the second joint of the robotic arm retracting by 6 degrees. The low-level control law can be converted to a left track target velocity of 0.34 m / s, a right track target velocity of 0.26 m / s, and the second joint target angle adjusted from the current 35 degrees to 29 degrees with a maximum angular velocity not exceeding 10 degrees / s. The chassis controller and joint controller execute according to a unified time reference, so that the chassis trajectory and the robotic arm retraction action are completed in coordination. As a fault-tolerant mechanism, if the control obstacle function cannot find feasible fine-tuning parameters that satisfy the safe distance, such as excessive intrusion of the obstacle on the right and a wall on the left, the system will not output a control law with a safety margin lower than the preset lower limit. Instead, it will generate instructions for safe stopping, short-distance reversal, or requesting remote confirmation. If an actuator malfunctions, such as the robotic arm failing to retract, the system will remove that degree of freedom from the controllable set and recalculate the fine-tuning parameters that depend only on the chassis. If recalculation is still not feasible, the system will enter the stopping branch. If the fine-tuning parameters exceed the speed, torque, or acceleration limits of the underlying actuator, the system will trim the actuator according to its upper limit and recheck whether the trimmed motion still satisfies the safety boundary. At a narrow corner during a security patrol at a chemical plant, the robot detected a sliding debris behind its right rear. The reconstructed envelope showed that both the outer edge of the robotic arm and the right corner of the rear load compartment were close to obstacles. The system extracted three controllable degrees of freedom: chassis differential speed, robotic arm shoulder joint adduction, and load compartment locking angle, and verified them in accordance with... The constraint-based obstacle function calculation yields a combination of right track deceleration, left track slight maintenance, and robotic arm retraction of 8 degrees. After execution by the underlying controller, the robot's tail swing radius decreases, and the outer edge of the robotic arm avoids the gravel boundary. The purpose of this mechanism is to transform the spatial risks obtained after review into executable underlying control actions, so that the mobile chassis and actuator are no longer regarded as an inseparable whole, but rather compress the rear half sweeping area together through their independent degrees of freedom, thereby achieving local obstacle avoidance fine-tuning without complete shutdown. Before comparing the pose vulnerability entropy of the rear blind zone with the preset dynamic hazard threshold, the following steps are also included: Obtain the initial exploration trajectory output by the deep reinforcement learning obstacle avoidance model without smoothing filtering; The initial exploration trajectory is input into a rigid body dynamics safety boundary model constrained by a collision avoidance safety distance for collision verification; Update the physical boundaries of obstacles in the local static map based on the collision verification results.

[0021] This embodiment provides a mechanism for rigid body safety verification of the initial exploration trajectory before vulnerability entropy comparison. Specifically, the aforementioned scheme can trigger review and fine-tuning when the risk of the rear half increases. However, in highly dynamic rescue missions, the forward obstacle avoidance trajectory may be generated by a deep reinforcement learning obstacle avoidance model. The deep reinforcement learning obstacle avoidance model is a neural network built based on Markov decision process. Its state space input includes at least the occupied grid data of the local static map, the current linear velocity and angular velocity of the embodied robot, and its action space output is the target linear acceleration and angular acceleration of the embodied robot. The deep reinforcement learning obstacle avoidance model uses a reward function that includes a collision penalty term, a target arrival reward term, and a control action smoothness penalty term for parameter iterative training. This type of model has the advantage of rapid exploration and adaptation to unknown environments, but its output initial exploration trajectory may exhibit high dynamic curvature and a safety margin close to the preset lower limit. Furthermore, without smoothing filtering, it may exhibit a heading angle change rate exceeding the preset threshold, approach obstacles along the edge of the minimum collision avoidance safety distance, or have a curvature change rate exceeding the limit in a short period of time. If this trajectory is directly used for rear-body vulnerability entropy calculation, the obstacle boundaries in the local map may not be fully verified, leading to a lag in risk assessment. Therefore, this embodiment incorporates a rigid body dynamics safety boundary model for collision verification before threshold comparison. Specifically, the system acquires the initial exploration trajectory output by the deep reinforcement learning obstacle avoidance model. This trajectory can consist of several discrete pose points, each including position, heading, velocity, and angular velocity. The lack of smoothing filtering means that the trajectory still retains the original control trend generated by the model for rapid obstacle avoidance, such as rapidly turning left and then right within a short distance, or grazing along the obstacle boundary. The system does not immediately reject this trajectory because, in urgent rescue situations, an aggressive trajectory may lead to higher passage efficiency. However, the system needs to first check whether the trajectory violates the rigid body physical safety boundary. The system inputs the initial exploration trajectory into a rigid body dynamics safety boundary model constrained by a collision avoidance safety distance for collision verification. This model can perform spatial sweep checks on each sampling point on the trajectory based on the multi-rigid body contours of the robot chassis, robotic arm, and payload compartment. As an example, assume the initial exploration trajectory includes three sampling poses P1, P2, and P3, the obstacle boundary is located 0.80 meters to the right, the outer contour of the robot's rear right side is 0.68 meters, and the required collision avoidance safety distance is 0.15 meters. If the predicted distance at P1 is 0.20 meters, the predicted distance at P2 is 0.13 meters, and the predicted distance at P3 is 0.18 meters, then P2 does not meet the safety distance requirement. The system can mark the physical boundary of the obstacle near P2 as a high-risk boundary, or mark the trajectory segment as requiring replanning, deceleration, or review confirmation. Based on the collision verification results, the system updates the physical boundary of the obstacle in the local static map. The update does not necessarily mean that the obstacle has actually moved; it can also represent the system's correction of the boundary's danger level. For example, if the initial exploration trajectory is too close to a wall at a certain point, the system can safely expand the boundary of the wall in the local static map, extending the obstacle boundary to the passable area by 0.05 meters to 0.15 meters, making the subsequent vulnerability entropy calculation more conservative; if the collision check finds that the outer edge of the robotic arm may have swept over the target inclined beam, the effective boundary of the inclined beam in the map is expanded to the range of the robotic arm's height projection, instead of judging it only by the chassis height; In abnormal situations, if the deep reinforcement learning obstacle avoidance model does not output a valid trajectory, the system can use a traditional local planner to generate a conservative trajectory and limit the forward propulsion speed to a low range. If there is an excessive curvature change in the initial exploration trajectory, such as the change in angular velocity between adjacent sampling points exceeding the chassis's physical limit, the system first determines that the trajectory segment is dynamically infeasible and does not enter the collision verification branch. If all collision verifications pass, but the safe distance is close to the lower limit, the system can still increase the initial value of vulnerability entropy or reduce the subsequent danger threshold to reflect the rear half sensitivity caused by passing close to the edge. If the collision verification result deviates from the sensor map data by more than the preset fault tolerance threshold, such as the model believing there is an obstacle but the point cloud is empty, the system retains a more conservative obstacle boundary until the side, rear or forward sensors reconfirm it. In a chemical plant, a robot needs to quickly traverse a narrow, winding passage through an inspection tunnel. A deep reinforcement learning obstacle avoidance model outputs an unsmoothed initial exploration trajectory, which approaches the right-side wall at the second turning point. A rigid body dynamics safety boundary model projects the rear load chamber and the outer edge of the robotic arm onto this trajectory, discovering that at point P2, the outer edge of the robotic arm is only 0.12 meters from the wall, less than the 0.15-meter collision avoidance safety distance. Based on this, the system safely expands the right-side wall boundary in a local static map and calculates the subsequent rear blind zone pose vulnerability entropy based on the updated boundary. The purpose of this mechanism is to improve passage efficiency by utilizing a highly dynamic exploration trajectory while physically verifying the aggressive trajectory through a rigid body dynamics safety boundary model, thus correcting obstacle boundaries in the local static map in advance and reducing the underestimation of rear-side risk caused by the unsmoothed trajectory approaching the edge. Furthermore, to avoid misinterpreting the updating of the physical boundaries of obstacles in the local static map as the system's determination that the obstacles have actually moved, the update in this embodiment is preferably understood as a correction of the effective physical boundaries of obstacles used for obstacle avoidance decision-making; that is, what is updated at the map level is the boundary representation in a control sense, rather than a factual rewriting of the objective world obstacle morphology without sensor confirmation; the effective physical boundary of obstacles can be either equivalent to the actual boundary directly observed by the sensor, or a safety expansion boundary formed based on the actual boundary and the rigid body sweep verification results; Furthermore, the following structured rules can be adopted when updating the boundary of the collision verification result: If the minimum distance between the multi-rigid body contour and the obstacle boundary at a certain sampling pose is greater than or equal to the anti-collision safety distance, the effective boundary of the obstacle at that location remains unchanged; if the minimum distance is less than the anti-collision safety distance, the insufficient amount is written into the local map region corresponding to the pose, and the corresponding obstacle boundary is expanded locally towards the passable area; if multiple adjacent sampling poses do not meet the safety distance, the expansion value corresponding to the largest insufficient amount is taken to uniformly correct the continuous boundary segment; in this way, the update action and the collision verification result of the initial exploration trajectory form a one-to-one correspondence, avoiding the situation where all obstacle boundaries are expanded without basis simply because the trajectory is aggressive; Furthermore, the above update process can be limited to a local area related to the initial exploration trajectory, rather than making a global modification to the entire local static map; preferably, only the obstacle boundaries within the initial exploration trajectory sweep corridor and its adjacent safe distance zone are corrected; for areas unrelated to the trajectory, the original map boundaries are retained; this ensures that the map on which the subsequent vulnerability entropy calculation is based contains both the conservative verification results of the aggressive trajectory and avoids the erroneous compression of the passability of unrelated areas due to over-expansion; Furthermore, if the collision verification results show that a certain trajectory segment is passable at the chassis level, but the outer edge of the robotic arm, the rear end of the load compartment, or the towing module does not meet the safe distance in the height projection or outward swing projection, the system prefers to update the effective boundary layer of obstacles related to the corresponding rigid body parts, rather than only updating the chassis planar layer boundary. As a result, the boundary correction in the local static map can be consistent with the physical boundary parameters of the multi-degree-of-freedom rigid body, so that the subsequent calculation of the rear blind zone pose vulnerability entropy is aimed at the map boundary after rigid body consistency verification, rather than the simplified boundary obtained only based on the two-dimensional contour of the chassis. The steps for acquiring dynamic tail-following slip data based on side and rear sensors include: The physical boundary of the rear obstacle at the current moment is obtained based on the side and rear sensors, and the physical boundary of the rear obstacle saved at the previous moment is extracted as the historical physical boundary of the obstacle. The changes between the two are monitored. When the change is greater than or equal to the preset slip tolerance threshold, dynamic tail slip interference is determined to have occurred. The change is used as dynamic tail slip data, and the threshold decay is calculated based on the product of the change and the preset penalty coefficient. The dynamic danger threshold is dynamically lowered by subtracting the threshold decay from the current dynamic danger threshold. When the change is less than the preset slip tolerance threshold, it is determined that no dynamic tail slip disturbance has occurred, and the dynamic danger threshold remains unchanged.

[0022] This embodiment provides a mechanism for dynamic tail-following slip recognition and threshold adaptive adjustment. Specifically, the aforementioned scheme can obtain changes in the rear environment through side-rear scanning. However, if all boundary displacements with amplitudes within the sensing noise range are considered dangerous slips, it will cause the robot to frequently slow down and look back, affecting the rescue timeliness. Conversely, if it is not sensitive to changes in rear obstacles, the intrusion of gravel or wall edges smaller than the slip tolerance threshold may also cause the rear safety margin to fail instantly. Therefore, this embodiment determines whether the change exceeds the slip tolerance threshold by comparing the physical boundary of the rear obstacle at the current moment with that at the previous moment, and adjusts the dynamic danger threshold according to the determination result. Specifically, the system acquires the physical boundary of the rear obstacle at the current moment based on the side and rear sensors; this boundary can be represented by point cloud clustering, distance profile, or local occupancy grid; the system also extracts the physical boundary of the rear obstacle saved in the previous moment as the physical boundary of the historical obstacle; the previous moment here can be the previous control cycle, the previous back scan cycle, or the boundary recorded by the forward sensor when the front of the robot just passed the obstacle; the system aligns the two in the robot body coordinate system or the local map coordinate system and calculates the change; the change can include the lateral intrusion distance, the change in the boundary tilt angle, the number of newly added occupancy grids, or the reduction in the distance to the nearest obstacle; As an example, assuming the distance between the right-side debris and the robot's centerline in the historical obstacle boundary is 0.82 meters, and the current distance measured by the rear-side sensor is 0.76 meters, then the lateral change is 0.06 meters. If the preset slip tolerance threshold is 0.04 meters, then 0.06 meters is greater than the threshold, the system determines that dynamic tailing slip interference has occurred, and uses 0.06 meters as dynamic tailing slip data. If the historical distance of another wall is 0.95 meters, and the current distance is 0.93 meters, the change is 0.02 meters, which is less than 0.04 meters. This change can be considered as sensor noise or slight jitter, and does not trigger slip interference judgment. When the change is greater than or equal to the slip tolerance threshold, the system determines that dynamic tailing slip interference has occurred, uses the change as dynamic tailing slip data, and lowers the dynamic danger threshold. Lowering the threshold allows the system to trigger a review or maintain a review state earlier when the rear environment is unstable. It should be noted that, in order to ensure the consistency of dimensions in the calculation process, the preset penalty coefficient has the dimension of the reciprocal of distance. This allows the product of the change with the dimension of distance and the penalty coefficient to obtain a dimensionless pure value as the threshold attenuation amount, thus maintaining a consistent numerical dimension with the dynamic hazard threshold. For example, the standard danger threshold is 0.60. When a slip of 0.06 meters is detected, the threshold can be lowered to 0.52. If slip is detected for two consecutive cycles, the threshold can be further lowered to 0.48. In this way, even if the pose vulnerability entropy of the rear blind zone has not yet reached the original threshold, it will enter conservative control in advance because the environment is invading the rear half of the body envelope. When the change is less than the slip tolerance threshold, the system determines that no dynamic tail slip interference has occurred and maintains the dynamic danger threshold unchanged. Maintaining the threshold unchanged can prevent sensor noise from causing excessive conservative control. For example, if the side-rear radar experiences random distance fluctuations of 0.01 meters to 0.02 meters in a dusty environment, and the slip tolerance threshold is 0.04 meters, the system will not change the threshold, nor will it write the change into the dynamic tail slip data. Instead, it will continue to determine whether a backtracking is needed based on the current vulnerability entropy. As a fault-tolerance mechanism, if the current boundary and the historical boundary cannot be reliably aligned, for example, if the robot's pose estimation error is too large, the system does not directly regard all the differences as slippage, but first increases the positioning reliability requirement or uses multi-sensor cross-validation; if the change is exactly equal to the slippage tolerance threshold, the system treats it as slippage to ensure safety in the critical state; if the change is detected multiple times in succession but shows a monotonically accumulating trend, for example, each intrusion is 0.02 meters, and the cumulative change is 0.06 meters three times, the system can set a cumulative slippage judgment, and trigger slippage interference when the cumulative change exceeds the threshold; if the change direction is that the obstacle is moving away from the robot, the system can record the change but does not immediately increase the speed, and needs to wait for several stable periods before restoring the original danger threshold; The robot's front section had just passed a pile of rubble against the right wall, and the historical map recorded that the edge of the rubble was 0.84 meters from the centerline. Half a second later, the side and rear radar, in retrospect, measured the obstacle's edge to be 0.77 meters from the centerline, a change of 0.07 meters, exceeding the slip tolerance threshold of 0.04 meters. The system determined that the rubble had undergone dynamic tail slippage, recorded 0.07 meters in the dynamic tail slippage data, and lowered the dynamic danger threshold from 0.60 to 0.50, allowing the robot to continue maintaining low speed and fine-tuning its rear half. If the subsequent three scans all showed a change of less than 0.01 meters, the system would gradually restore the threshold. The purpose of this mechanism is to transform the small but dangerous intrusion of the obstacle boundary in the rear blind spot into quantifiable dynamic trailing slip data, and to improve the system’s sensitivity to rear-side risks by lowering the dynamic hazard threshold, thereby avoiding making full-speed forward decisions based on outdated maps. Furthermore, to reduce false slip misjudgments caused by the robot's own movement, the physical boundary of the obstacle behind at the current moment is preferably compensated for based on the change in robot pose between the two moments before comparing the physical boundary of the obstacle behind at the current moment with the physical boundary of the obstacle behind at the historical moment. That is, the system first uses the wheel speedometer, inertial measurement unit or short-time pose calculation results to transform the physical boundary of the historical obstacle to the robot body coordinate system corresponding to the current moment, and then compares the overlapping area with the current retrace result. Through this compensation step, the apparent displacement caused by the robot's forward movement, turning and slight posture changes can be eliminated from the boundary difference, making the change closer to the actual sliding, tilting or intrusion of the obstacle relative to the robot's channel. Furthermore, the aforementioned changes are preferably extracted in the unfavorable direction toward the robot's sweeping area; specifically, when the current boundary contracts toward the inside of the robot's passage relative to the historical boundary, it is counted as an effective intrusion change; when the boundary moves away from the robot relative to the historical boundary or expands only in the non-sweeping direction, it can be recorded as environmental release information, but the dynamic danger threshold is not immediately increased accordingly; this ensures that the threshold adaptive adjustment is mainly driven by changes that are unfavorable to the rear half safety envelope, and does not cause the system to relax its vigilance prematurely due to occasional changes away from the robot. Furthermore, in multi-sensor scenarios, the system can first perform consistency verification on the boundary results of the side-rear lidar, millimeter-wave radar, ultrasonic array, or rotatable depth camera within the same rear sector, and then output the final change amount for threshold adjustment. If multiple sensors have the same judgment direction for the same boundary change, the reliability of the change amount is improved. If there is a significant conflict between different sensors, the boundary result corresponding to the smaller distance is adopted according to the conservative principle, or the current low-speed state is maintained and rescanning is continued without immediately restoring a higher threshold. Through this processing, the impact of dust obstruction, echo loss, or local voids on the slip determination results can be reduced. Furthermore, the aforementioned cumulative slip determination is preferably limited to the same obstacle cluster or the same sweep area; that is, the system only accumulates these small changes when several consecutive detected changes come from the same orientation, similar boundary segments, or the same cluster of obstacles; if the source of the change jumps between different areas, it is preferable to judge them separately as independent events; this can avoid simply superimposing multiple unrelated sensor noises into a single misjudged over-limit slip event, thereby ensuring that the basis for lowering the dynamic hazard threshold corresponds to the specific rear risk location; Example 2: The android obstacle avoidance and replay control device applies the android obstacle avoidance and replay control method as described in any one of the embodiments, including: The data acquisition module is used to collect real-time kinematic data and forward environmental point cloud data through the forward sensors mounted on the embodied robot, and to keep the side and rear sensors mounted on the embodied robot in a basic monitoring state. The boundary extraction module is used to timestamp and synchronize real-time kinematic data and forward environmental point cloud data, perform pose calculation based on the body coordinate system of the embodied robot, construct a local static map, and obtain the structural outline data of the embodied robot. Combined with the structural outline data, the multi-degree-of-freedom rigid body physical boundary parameters are extracted from the local static map. The vulnerability entropy calculation module is used to construct the rear half safety envelope of the embodied robot based on the physical boundary parameters of the multi-degree-of-freedom rigid body and the local static map, combined with the nonholonomic constraint model of the rear half of the embodied robot. It obtains the nonlinear offset of the rear trajectory between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope, and uses the nonlinear offset of the rear trajectory to calculate the rear blind zone pose vulnerability entropy. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half will deviate from the rear half safety envelope. The instruction generation module is used to compare the pose vulnerability entropy of the rear blind zone with the preset dynamic hazard threshold and generate corresponding control mode instructions. The adaptive control module triggers a lookback control mechanism when the rear blind zone pose vulnerability entropy is greater than or equal to the dynamic danger threshold. This mechanism performs deceleration and acquires dynamic tail-following slip data based on the side and rear sensors. It then generates independent obstacle avoidance fine-tuning parameters based on the dynamic tail-following slip data and converts these parameters into a low-level control law, which is then sent to the chassis controller of the embodied robot to adjust the rear half's motion trajectory for independent obstacle avoidance fine-tuning. When the rear blind zone pose vulnerability entropy is less than the dynamic danger threshold, the module maintains the current forward propulsion speed.

[0023] This embodiment provides a modular implementation mechanism for an obstacle avoidance and replay control device for an embodied robot. Specifically, the aforementioned method can be implemented by multiple software modules, hardware interface modules, or software-hardware collaborative units in the robot's onboard controller. For security patrol and chemical plant inspection robots, sensor data processing, motion control, local map construction, and robotic arm control are often run in parallel by different control tasks. If there is a lack of clear module division, the timing of replay triggering, back-side scanning, and low-level control distribution may be disordered. Therefore, this embodiment divides the obstacle avoidance and replay control device into a data acquisition module, a boundary extraction module, a vulnerability entropy calculation module, an instruction generation module, and an adaptive control module. Specifically, the data acquisition module connects to the forward-facing LiDAR, depth camera, inertial measurement unit, wheel speedometer, robotic arm encoder, and side-rear sensors. This module is responsible for caching data according to a unified time reference and keeping the side-rear sensors in a basic monitoring state under normal conditions. For example, the data acquisition module receives the forward point cloud every 50 milliseconds, the inertial data every 10 milliseconds, and the low frame rate distance alarm from the side-rear sensors every 100 milliseconds. If a review trigger signal is received from the instruction generation module, the module switches the side-rear sensors to a high frame rate or active scanning state. The boundary extraction module obtains standardized data from the data acquisition module, timestamps and synchronizes the real-time kinematic data and the forward environmental point cloud data, completes the pose calculation in the body coordinate system, constructs a local static map, and extracts the physical boundary parameters of the multi-degree-of-freedom rigid body. This module can maintain a short-term map buffer, such as saving the local map and obstacle boundaries within the last 2 seconds. If the robot turns left in a narrow passage, this module will output the front motion trajectory, the safety margin reserved for the rear half, the outward expansion boundary of the robotic arm, and the space occupied at the rear end of the payload compartment. The vulnerability entropy calculation module receives the physical boundary parameters and local static map output by the boundary extraction module, and calls the rear half nonholonomic constraint model to construct the rear half safety envelope. This module calculates the nonlinear offset of the rear trajectory between the actual rear half trajectory and the safety envelope, and uses positive correlation mapping to obtain the rear blind zone pose vulnerability entropy. As a simplified data flow example, the boundary extraction module outputs a right rear corner margin of 0.12 meters, a robotic arm outer edge margin of 0.10 meters, and a current angular velocity of 0.35 radians / second; the vulnerability entropy calculation module calculates the maximum risk offset of 0.07 meters and outputs a vulnerability entropy of 0.68; the instruction generation module compares the evaluation value output by the vulnerability entropy calculation module with the dynamic hazard threshold and generates control mode instructions; if the evaluation value is greater than or equal to the threshold, the module outputs a review control instruction; if it is less than the threshold, it outputs a speed maintenance instruction; the instruction generation module can also receive dynamic tail slip judgment results and dynamically adjust the hazard threshold; For example, after a slip interference occurs, the threshold drops from 0.60 to 0.52, and a vulnerability entropy of 0.55 can trigger a lookback control. The adaptive control module receives control mode instructions. If lookback control is entered, the module executes a deceleration strategy, requests high-frame-rate data from the side and rear from the data acquisition module, generates independent obstacle avoidance fine-tuning parameters based on dynamic tail slip data, and converts the fine-tuning parameters into underlying control laws and sends them to the chassis controller and actuator controller. If the current forward propulsion speed is maintained, the module still retains basic monitoring and vulnerability entropy periodic updates without additionally consuming high-computing power backscanning resources. As a fault-tolerance mechanism, if a module experiences data delay or anomaly, the device can set a health status flag between modules; if the data acquisition module reports that the side and rear sensors are offline, the adaptive control module will prohibit high-speed edge-grabbing and limit the maximum speed to low speed; if the boundary extraction module outputs a map with low confidence, the vulnerability entropy calculation module will adopt a conservative boundary; if the instruction generation module receives contradictory inputs in a continuous cycle, the adaptive control module will execute a hysteresis strategy or a safe stop; inter-module communication can use a timestamp message queue, and if a message exceeds the valid time window, it will not participate in the control calculation for this cycle. During security patrols at the chemical plant, the data acquisition module continuously receives forward point cloud and chassis motion data; the boundary extraction module corrects the point cloud and constructs a local map, discovering that the margin between the right-side debris and the rear load compartment is lower than the preset safety distance limit; the vulnerability entropy calculation module predicts the tail swing based on nonholonomic constraints and outputs a rear blind zone pose vulnerability entropy of 0.70; the instruction generation module compares it with a threshold of 0.60 and issues a retrospective control instruction; the adaptive control module reduces the vehicle speed, initiates a side and rear scan, discovers that the debris continues to intrude, generates track differential and robotic arm retraction control laws, and issues them for execution. The purpose of this device is to achieve closed-loop coordination between data acquisition, boundary extraction, vulnerability entropy calculation, instruction generation and adaptive control through a modular approach, so that obstacle avoidance and look-back control can be stably deployed in the vehicle controller and have a clear degradation processing path when sensors malfunction, map confidence decreases or obstacles slip behind. Example 3: Please see Figure 2 Embossed robotic systems include: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of any of the embodiments of the obstacle avoidance and retrospective control method for the android.

[0024] This embodiment provides an implementation mechanism for an embodied robot system. Specifically, the system can be installed on an embodied robot for security patrols or chemical plant inspections, and includes a memory, a processor, a sensor interface, a chassis control interface, an actuator control interface, and a power management unit. The memory stores computer-executable instructions, robot geometric parameters, a rear half nonholonomic constraint model, hazard threshold configuration, sensor calibration parameters, and a local map cache. The processor executes the above instructions, enabling the robot to complete forward perception, local map construction, rear half safety envelope generation, vulnerability entropy calculation, back-view triggering, side and rear scanning, and low-level control distribution. Specifically, the memory can include non-volatile memory and running memory; the non-volatile memory stores the robot's factory calibration data, such as chassis wheelbase, body width, length of each joint of the robotic arm, dimensions of the payload compartment, installation angle of the side and rear sensors, and initial table of dynamic hazard thresholds; the running memory stores real-time data, such as the most recent frame point cloud, kinematic state, local static map, historical rear obstacle boundaries, dynamic tail-following slip data, and current control mode; the processor can be an on-board central computing unit, a real-time controller, or a heterogeneous computing platform composed of multiple processing units; when the processor executes computer-executable instructions, it first reads the forward point cloud and kinematic data from the sensor interface and sets the side and rear sensors to the basic monitoring state; The processor performs timestamp synchronization, pose calculation, and map construction; then, based on the rear half nonholonomic constraint model, it calculates the rear half safety envelope and the pose vulnerability entropy of the rear blind zone; when the evaluation value reaches the dynamic danger threshold, the processor outputs a deceleration command to the chassis control interface and an active scanning command to the side and rear sensor interface; after obtaining dynamic tail-following slip data, the processor generates independent obstacle avoidance fine-tuning parameters and converts them into low-level control laws, which are then sent to the chassis controller and actuator controller. As an example of the data flow, the danger threshold stored in the memory is 0.60, and the safe distance for the rear half of the body is 0.18 meters. In a certain control cycle, the processor reads the predicted distance of the right rear corner as 0.13 meters and the predicted distance of the outer edge of the robotic arm as 0.16 meters, calculates the offset as 0.05 meters, and maps it to the rear blind zone pose vulnerability entropy of 0.62. Since 0.62 is greater than 0.60, the processor executes a deceleration command to reduce the target speed from 0.70 m / s to 0.30 m / s, and at the same time starts the side and rear scan. The scan result shows that the obstacle boundary intrudes 0.05 meters, the processor lowers the danger threshold to 0.52, and issues a control law for the right track to decelerate and the robotic arm to retract. As a fault-tolerant mechanism, if the geometric parameters in the memory fail to be verified, such as the missing length parameter of the robotic arm or the damage to the sensor mounting parameters, the processor will not execute the high-speed obstacle avoidance process, but will instead load the default conservative outline parameters and limit the speed. If the processor load exceeds the preset computing power limit, the system can degrade according to task priority: chassis stability control and braking control have the highest priority, followed by rear-view scanning, and forward high-speed exploration has the lowest priority. If the processor detects that the underlying controller has not confirmed the execution of the deceleration or fine-tuning command, the system will send the speed limit command again and enter a safe stop after the timeout. If the power supply is insufficient to support high frame rate retracement and robotic arm movements, the system will prioritize maintaining low-speed stability of the chassis and safe storage of the robotic arm. When performing patrol and security operations in complex areas of a chemical plant, the embodied robot system's memory pre-stores the dimensions of the rear half of the vehicle body, the robotic arm's storage boundary, and the security patrol mode threshold. After the processor executes the control program, it reads the forward point cloud in real time and constructs a local map, calculating the high rear blind spot pose vulnerability entropy at corners. The system then reduces the chassis speed, calls the side and rear sensors to check for gravel slippage, and sends the fine-tuning control law to the chassis controller, enabling the robot to avoid rear-end collisions with the rear load compartment in narrow passages. The purpose of this system is to enable the embodied robot to complete closed-loop control from perception, evaluation, triggering to execution on the vehicle end by carrying the obstacle avoidance review control process through memory and processor. Under the conditions of post-disaster dynamic environment, rear blind spot and multi-degree-of-freedom rigid body constraints, it can maintain interpretable and degradable obstacle avoidance control capability.

[0025] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for obstacle avoidance and re-viewing control of an embodied robot, characterized in that, include: Real-time kinematic data and forward environmental point cloud data are collected by the forward-facing sensor on the embodied robot, while the side and rear sensors on the embodied robot are kept in a basic monitoring state. The real-time kinematic data and the forward environmental point cloud data are time-stamped and synchronized. The pose is calculated based on the body coordinate system of the embodied robot to construct a local static map and obtain the structural outline data of the embodied robot. In combination with the structural outline data, the multi-degree-of-freedom rigid body physical boundary parameters are extracted from the local static map. Based on the physical boundary parameters of the multi-degree-of-freedom rigid body and the local static map, and combined with the nonholonomic constraint model of the rear half of the embodied robot, a rear half safety envelope is constructed. The nonlinear offset of the rear trajectory between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope is obtained. The rear blind zone pose vulnerability entropy is calculated using the rear trajectory nonlinear offset. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half deviates from the rear half safety envelope. The rear blind zone pose vulnerability entropy is compared with a preset dynamic danger threshold to generate corresponding control mode commands. When the rear blind zone pose vulnerability entropy is greater than or equal to the dynamic danger threshold, the control mode command is to trigger the look-back control mechanism, perform deceleration operation and acquire dynamic tail-following slip data based on the side and rear sensors, generate independent obstacle avoidance fine-tuning parameters based on the dynamic tail-following slip data, and convert the independent obstacle avoidance fine-tuning parameters into a low-level control law and send it to the chassis controller of the robot to adjust the rear half movement trajectory for independent obstacle avoidance fine-tuning. When the rear blind zone pose vulnerability entropy is less than the dynamic danger threshold, the control mode command is to maintain the current forward propulsion speed.

2. The obstacle avoidance and rewind control method for an embodied robot as described in claim 1, characterized in that, The steps of synchronizing the real-time kinematic data and the forward environment point cloud data with timestamps, performing pose calculation based on the embodied robot body coordinate system, constructing a local static map, and extracting multi-degree-of-freedom rigid body physical boundary parameters from the local static map include: The distortion correction process is performed on the forward environmental point cloud data to obtain standard point cloud data; The standard point cloud data and the real-time kinematic data are time-stamped and synchronized, and pose calculation and data fusion are performed based on the body coordinate system of the embodied robot to construct a local static map; The front movement trajectory of the embodied robot is extracted from the local static map, and the structural outline data of the embodied robot is combined to calculate the safety margin reserved for the rear half of the body. Based on the front motion trajectory and the safety margin reserved for the rear body, the spatial boundary mapping of the structural outline data is performed to generate the physical boundary parameters of the multi-degree-of-freedom rigid body.

3. The obstacle avoidance and retrospective control method for an embodied robot as described in claim 2, characterized in that, The steps of constructing a rear half safety envelope based on the rear half nonholonomic constraint model of the embodied robot, obtaining the rear trajectory nonlinear offset between the actual rear half trajectory of the embodied robot and the rear half safety envelope, and calculating the rear blind zone pose vulnerability entropy using the rear trajectory nonlinear offset include: Obtain the nonholonomic constraint model of the rear half of the embodied robot; Combining the nonholonomic constraint model of the latter half and the physical boundary parameters of the multi-degree-of-freedom rigid body, the safety envelope of the latter half is constructed, and the difference between the actual trajectory of the latter half and the safety envelope of the latter half is extracted as the nonlinear offset of the latter trajectory. The nonlinear offset of the rear trajectory is input as an independent variable into a probability distribution function configured with a positive correlation mapping relationship for calculation, thereby obtaining the rear blind zone pose vulnerability entropy that characterizes the magnitude of the deviation probability.

4. The obstacle avoidance and retrospective control method for an embodied robot as described in claim 3, characterized in that, The steps of performing deceleration and acquiring dynamic trailing slip data based on the side and rear sensors, and generating independent obstacle avoidance fine-tuning parameters based on the dynamic trailing slip data for independent obstacle avoidance fine-tuning include: After performing the deceleration operation, obtain the available computing power resources of the system controller configured for the android; The available computing resources are allocated to the algorithm module that processes data from the rear-side sensors to perform a backward scan and generate dynamic tail-following slip data. Reconstruct the rear body safety envelope based on the dynamic tail-following slip data; Independent obstacle avoidance fine-tuning parameters are generated based on the reconstructed rear body safety envelope, and independent obstacle avoidance fine-tuning is performed.

5. The obstacle avoidance and retrospective control method for an embodied robot as described in claim 4, characterized in that, The steps for generating independent obstacle avoidance fine-tuning parameters based on the reconstructed rear-body safety envelope include: Extract the independent control degrees of freedom of the mobile chassis and actuators of the embodied robot; Based on the reconstructed rear half safety envelope and combined with the independent control degrees of freedom, the independent obstacle avoidance fine-tuning parameters of the mobile chassis and the actuator are calculated using the control obstacle function with the obstacle avoidance safety distance as the boundary constraint. The independent obstacle avoidance fine-tuning parameters are converted into underlying control laws and sent to the chassis controller to perform independent obstacle avoidance fine-tuning.

6. The obstacle avoidance and retrospective control method for an embodied robot as described in claim 5, characterized in that, Before comparing the rear blind zone pose vulnerability entropy with a preset dynamic hazard threshold, the method further includes: Obtain the initial exploration trajectory output by the deep reinforcement learning obstacle avoidance model without smoothing filtering; The initial exploration trajectory is input into a rigid body dynamics safety boundary model constrained by a collision avoidance safety distance for collision verification; The physical boundaries of obstacles in the local static map are updated based on the collision verification results.

7. The obstacle avoidance and retrospective control method for an embodied robot as described in any one of claims 1 to 6, characterized in that, The steps for acquiring dynamic tail-following slip data based on the side-rear sensor include: The physical boundary of the rear obstacle at the current moment is obtained based on the side and rear sensor, and the physical boundary of the rear obstacle saved at the previous moment is extracted as the historical physical boundary of the obstacle, and the change of the two is monitored. When the change is greater than or equal to a preset slip tolerance threshold, dynamic tail slip interference is determined to have occurred. The change is used as dynamic tail slip data, and the threshold decay is calculated based on the product of the change and a preset penalty coefficient. The dynamic danger threshold is dynamically lowered by subtracting the threshold decay from the current dynamic danger threshold. When the change is less than the preset slip tolerance threshold, it is determined that no dynamic tail slip interference has occurred, and the dynamic danger threshold is kept unchanged.

8. A robot obstacle avoidance and replay control device, employing the robot obstacle avoidance and replay control method as described in any one of claims 1 to 7, characterized in that, include: The data acquisition module is used to collect real-time kinematic data and forward environmental point cloud data through the forward sensors mounted on the embodied robot, and to keep the side and rear sensors mounted on the embodied robot in a basic monitoring state. The boundary extraction module is used to timestamp and synchronize the real-time kinematic data and the forward environmental point cloud data, perform pose calculation based on the body coordinate system of the embodied robot, construct a local static map, and obtain the structural outline data of the embodied robot. Combined with the structural outline data, the module extracts the multi-degree-of-freedom rigid body physical boundary parameters from the local static map. The vulnerability entropy calculation module is used to construct a rear half safety envelope based on the multi-degree-of-freedom rigid body physical boundary parameters and the local static map, combined with the rear half nonholonomic constraint model of the embodied robot, to obtain the rear trajectory nonlinear offset between the actual trajectory of the rear half of the embodied robot and the rear half safety envelope, and to calculate the rear blind zone pose vulnerability entropy using the rear trajectory nonlinear offset. The rear blind zone pose vulnerability entropy is an evaluation value used to characterize the probability that the actual trajectory of the rear half deviates from the rear half safety envelope. The instruction generation module is used to compare the rear blind zone pose vulnerability entropy with a preset dynamic danger threshold and generate corresponding control mode instructions. The adaptive control module is used to trigger a lookback control mechanism when the rear blind zone pose vulnerability entropy is greater than or equal to the dynamic danger threshold. This mechanism performs deceleration and acquires dynamic tail-following slip data based on the side and rear sensors. It then generates independent obstacle avoidance fine-tuning parameters based on the dynamic tail-following slip data and converts these parameters into a low-level control law, which is then sent to the chassis controller of the robot to adjust the rear half's motion trajectory for independent obstacle avoidance fine-tuning. When the rear blind zone pose vulnerability entropy is less than the dynamic danger threshold, the module maintains the current forward propulsion speed.

9. An embodied robotic system, including: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the obstacle avoidance and retrospective control method of the embodied robot according to any one of claims 1 to 7.

Images