Method for autonomous obstacle avoidance and task rescheduling of robot dog for high-risk area of cement plant

Abstract

This invention relates to the field of embodied intelligence technology, and discloses a method for autonomous obstacle avoidance and task rescheduling for robotic dogs operating in high-risk areas of cement plants. This method employs multimodal sensor fusion to achieve real-time environmental perception, and constructs a dynamic risk field based on an offline-calibrated 2.5D grid and a lookup table method. Then, it uses hierarchical path planning—global A*-guided low-risk paths and local DWA real-time obstacle avoidance—for task path planning and execution. Combined with asynchronous task rescheduling, task sequence optimization is placed in an independent thread, supporting timeout returns of suboptimal solutions, and directly triggering preset templates for emergency events, improving the timeliness of response to sudden emergencies. This invention ensures the operational safety of robotic dogs in high-risk environments such as cement plants, improves the navigation robustness and task response of robotic dogs in dynamic high-risk environments, and enhances overall inspection efficiency.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligence technology, particularly to the movement and task planning technology of intelligent robots, and more specifically to a method for autonomous obstacle avoidance and task rescheduling for a robot dog in high-risk areas of a cement plant. Background Technology

[0002] Many high-risk areas exist in process industries such as cement, metallurgy, and chemicals, including the vicinity of critical equipment such as rotary kilns, preheaters, grate coolers, large fans, and high-altitude corridors. Regular inspections of these areas are crucial for ensuring production safety and preventing accidents. Traditional manual inspection methods suffer from inherent drawbacks such as high safety risks, high labor intensity, and subjective and discontinuous data. In recent years, with the development of mobile robot technology, especially quadruped robots (robot dogs) due to their superior terrain-crossing capabilities, they have begun to be used to replace or assist humans in performing these high-risk inspection tasks.

[0003] However, existing robotic dogs and their control systems used in industrial inspection still have significant shortcomings in navigation and task planning capabilities. For example, traditional navigation algorithms rely on preset static maps and fixed paths, lacking real-time response capabilities to common dynamic changes in cement plants (such as temporary material accumulation, vehicle passage, and equipment maintenance obstructing access). When encountering unexpected obstacles, they often have to stop and wait or passively detour, leading to frequent task interruptions. Inspection task sequences are usually pre-programmed fixed processes that cannot be dynamically adjusted based on real-time perceived environmental risks (such as sudden high temperatures or hazardous gas leaks) or equipment status (such as emergency shutdowns), lacking task-level intelligent decision-making capabilities. In obstacle avoidance and path replanning, existing methods mostly consider geometric reachability and do not integrate cement plant-specific process safety knowledge (such as high-temperature radiation safety distances, dust explosion risk areas, and confined space access rules), making it difficult to make safe and efficient navigation and task decisions in complex and high-risk environments. Furthermore, when the plant area is large and there are many task points, existing systems lack the ability to coordinate task allocation and path planning for multiple robotic dogs, easily leading to path conflicts or inspection blind spots. Summary of the Invention

[0004] In view of the defects and shortcomings of the existing technology, the purpose of this invention is to provide a method for autonomous obstacle avoidance and task rescheduling of robot dogs in high-risk areas of cement plants. This method can quantify industrial safety risks as core optimization targets and achieve deep collaboration between the navigation layer and the task layer, thereby improving the robustness and intelligent response capability of inspection in high-risk dynamic environments and enabling efficient and timely response to equipment anomalies and safety risks.

[0005] According to a first aspect of the present invention, a method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant is proposed, comprising the following steps:

[0006] Step 1: Load the plant area semantic map, process safety knowledge base, and inspection task list; and construct the plant area's basic risk layer based on the static objects in the semantic map according to the process safety knowledge base;

[0007] Step 2: Construct a 2.5D grid map centered on the robot dog, with H meters in front and D meters to the left and right. Each grid cell stores a risk value of 0 to 1 to represent passability. The initial value is 0, indicating that it is passable, and 1 indicates that it is prohibited from passing.

[0008] Step 3: Utilize the multi-sensor fusion perception system on the robot dog to obtain the robot dog's real-time location and the distribution of static obstacles, dynamic objects, high-temperature areas, and gas concentrations in the surrounding area, and generate a dynamic risk layer regarding high-temperature risk, gas risk, and moving object risk.

[0009] Step 4: According to the preset planning cycle, superimpose the basic risk layer and the dynamic risk layer to obtain the current risk field R(x, y); where, if the height of the step exceeds the threshold, the risk is set to 1;

[0010] Step 5: Using the task head point of the task queue as the current task target point, and starting from the robot dog's current position, use the global planner to perform a global search in the grid area of ​​non-static obstacles to plan a global reference path from the current position to the target task point.

[0011] Step 6: Guided by the global reference path, perform local path planning through DWA in each control cycle, and execute the local path planning at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking movement.

[0012] Step 7: Continuously monitor the robot dog's operating status and external input, and when path blocking events, environmental risk events, device status events, and task-specific abnormal events are detected, write the event information to shared memory and set a rescheduling request flag; and

[0013] Step 8: In response to the detection of the rescheduling request flag, the robot dog stores the current state in shared memory and continues to execute the current local path planning. At the same time, it starts a low-priority background thread to solve the background problem, updates the reachability of each task point, re-plans the execution order of the remaining task points, and decides whether to immediately interrupt the current task in order to maximize the completion of the high-priority task.

[0014] As an optional implementation, in step 3, multi-source sensing data is acquired through a multi-sensor fusion sensing system, including surrounding static obstacles, dynamic objects, high-temperature areas, and gas concentration distribution, and a dynamic risk layer is generated regarding high-temperature risk, gas risk, and moving object risk, including:

[0015] Step 3-1: Use the LiDAR to scan the point cloud and match it with the preloaded semantic map. Combine the angular velocity and acceleration of the IMU to output the pose (x, y, yaw) of the robot dog in the global map.

[0016] Step 3-2: After the LiDAR point cloud is segmented on the ground, continuous points that are 0.1 meters above the ground and cannot be crossed by gait are projected onto the 2.5D grid map and marked as occupied grids.

[0017] Step 3-3: The infrared thermal imager detects and outputs high-temperature pixels, maps them to 3D coordinates, records the surface temperature, and generates the high-temperature risk R_hot according to the exponential decay model.

[0018] R_hot = A_hot * exp(-d / σ_hot);

[0019] Where A_hot represents the basic risk value determined by the surface temperature, d represents the distance from the robot dog to the high temperature point, and σ_hot is the attenuation distance constant;

[0020] Steps 3-4: For a grid that detects harmful gas, the risk value is set as R_gas = C / C_threshold; C is the harmful gas concentration value of the current grid output by the gas sensor, and C_threshold is the safety limit. If the value is exceeded, R_gas is truncated to 1.0 and diffuses to neighboring grids.

[0021] Steps 3-5: For each detected dynamic object, generate an elliptical risk field around its current position, with the major axis pointing in its direction of motion, to simulate its probability of occupancy in the near future. The risk amplitude is proportional to its speed: for every 1 m / s increase in speed, the risk amplitude increases by 0.2; the semi-major axis of the ellipse = speed × 1.5 seconds, and the semi-minor axis = 0.5 meters.

[0022] As an optional implementation, in step 4, within each 0.2s planning cycle, for each grid, the risk values ​​corresponding to the basic risk layer and the dynamic risk layer are taken as the maximum value according to the grid to obtain the current risk field R(x,y) of each grid.

[0023] In the z-direction, only the robot dog's standing plane and the foot contact surface when crossing obstacles are considered. The z-dimensional risk value is defined using a passability label, where a step height < 0.2 meters indicates passability with a risk value of 0, otherwise a risk value of 1.

[0024] If a grid is occupied by a static obstacle, the risk is immediately set to 1.

[0025] As an optional implementation, in step 5, the first task point in the inspection task list is used as the current task target point. Starting from the robot dog's current position, a global planner is used to perform a global search in the grid area of ​​non-static obstacles to plan a global reference path from the current position to the target task point, including the following steps:

[0026] Generate and maintain a task queue. Initially, the task queue is sorted according to the initial priority of the inspection task list.

[0027] Using the task point at the head of the team as the target point of the current task, and taking the robot dog's current position as the starting point, the global planner is used to perform a global search in the grid area of ​​non-static obstacles. First, the area to which the task point belongs is found, and then the grid search is performed in the area using the A* search algorithm.

[0028] As an optional implementation, in step 5, a grid search is performed within the region using the A* search algorithm, including:

[0029] Based on the A* algorithm, the cost f(n) for each grid n is designed as follows:

[0030] f(n)= g(n) + h(n) + α·R(n);

[0031] Where g(n) represents the actual path length from the starting point to grid n; h(n) represents the Euclidean distance from grid n to the target point, which is used as a heuristic function; R(n) represents the risk value of grid n in the current risk field; α is the risk penalty coefficient, which is determined by the safety level of the high-risk area of ​​the cement plant, representing the additional path cost corresponding to each 0.1 risk value;

[0032] The initial polyline path is obtained by searching for the grid coordinates. After smoothing the path and eliminating the corners, a global reference path is obtained.

[0033] As an optional implementation, in step 6, guided by the global reference path, in each control cycle, a set of speed combinations (v, ω) that can be reached in the next cycle are generated based on the robot dog's current linear velocity v and angular velocity ω. For each candidate speed combination (v, ω), a fixed time is used to extrapolate forward using a uniform motion model to generate multiple trajectories. Local path planning is then performed at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking motion.

[0034] The cost function design for each trajectory includes a fusion cost (Cost) of orientation cost (heading), obstacle distance cost (1 / dist_obs), velocity cost (v_m-v), and risk accumulation cost (risk_sum), specifically expressed as follows:

[0035] Cost=w1·heading+w2·(1 / dist_obs) +w3·(v_m-v)+ w4·risk_sum;

[0036] In the formula, the heading cost represents the angle between the trajectory endpoint orientation and the global path tangent direction;

[0037] The obstacle distance cost (1 / dist_obs) represents the distance on the trajectory to the nearest static obstacle. If the distance is less than the safe radius, the cost is set to infinity.

[0038] The velocity cost (v_m-v) represents the difference between the target velocity v_m and the current linear velocity v corresponding to the trajectory;

[0039] The cumulative risk cost, risk_sum, represents the sum of risk values ​​accumulated over sampling points every 0.1 meters along the trajectory.

[0040] w1, w2, w3, and w4 represent the weighting coefficients of the corresponding cost components.

[0041] As an optional implementation, in step 8, starting a low-priority background thread to perform background solving, updating the reachability of each task point, re-planning the execution order of the remaining task points, and deciding whether to immediately interrupt the current task to maximize the completion rate of high-priority tasks specifically includes the following steps:

[0042] First, based on the current risk field, determine whether each remaining task point is reachable. Reachability is defined as: there exists a path from the robot dog's current position to the task point, and the maximum risk value on the path is ≤ 0.8; if no path exists, mark the task point as skipped.

[0043] Then, the priority is adjusted according to the event type, with the priority order as follows: environmental risk events > device status events > path blocking events > task-specific abnormal events;

[0044] With the objective of minimizing the sum of total time and total risk integral, a genetic algorithm is used to solve for the optimal task sequence, and the updated task sequence is output:

[0045] If the first task in the updated task sequence is the same as the current task, then continue executing the original path;

[0046] If the speeds differ, determine if the robot dog is currently moving. If the speed is greater than 0.1 m / s, perform a smooth braking maneuver to bring the robot dog to a stop, then use the task point of the first task in the updated task sequence as the target and trigger global path replanning. If the robot dog has already stopped, proceed directly with global path replanning.

[0047] According to a second aspect of the present invention, a computer system is provided, comprising:

[0048] One or more processors;

[0049] The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the process of executing the aforementioned method for autonomous obstacle avoidance and task rescheduling of a robotic dog targeting high-risk areas of a cement plant.

[0050] In a third aspect of the present invention, a computer-readable medium for storing software is provided, the software comprising instructions executable by one or more computers, the instructions, when executed by the one or more computers, performing the aforementioned process of an autonomous obstacle avoidance and task rescheduling method for a robotic dog targeting high-risk areas in a cement plant.

[0051] The autonomous obstacle avoidance and task rescheduling method for robotic dogs in high-risk areas of cement plants, as described in the above embodiments of the present invention, firstly utilizes multi-modal sensor fusion to perceive the environment in real time, including static / dynamic obstacles, high temperatures, and harmful gases. It then rapidly generates a dynamic risk field based on an offline-calibrated 2.5D grid and a lookup table method, quantifying the safety risk at each spatial location. A hierarchical path planning method is employed, using global A* search for low-risk reference paths, combined with Local Dynamic Window (DWA) for high-frequency obstacle avoidance and smooth control. Finally, asynchronous task rescheduling is implemented in an independent thread using a genetic algorithm to solve the problem, updating the task queue executed by the robotic dog with the goal of minimizing the sum of total time consumption and total risk integral. This allows for rapid response to internal and external emergencies, dynamic adjustment of inspection strategies, and improved timeliness in responding to equipment anomalies and safety risks. The present invention achieves deep collaboration between the navigation layer and the task layer, ensuring the operational safety of the robotic dog in high-risk environments such as cement plants, and significantly enhancing its robustness and responsiveness in high-risk dynamic environments.

[0052] Compared with the prior art, the significant advantages of the present invention are:

[0053] (1) By modeling and sensing the dynamic risk field in real time, the 2.5D grid and offline lookup table method are used to replace the complex online physical simulation. The risk values ​​of high temperature, gas and dynamic obstacles are quickly superimposed through the pre-calibrated distance-risk mapping, so that the robot dog can quickly and proactively adapt to the complex and ever-changing environment of the cement plant. When encountering dynamic obstacles or risks, it can intelligently detour. Combined with task rescheduling, the achievement of the core inspection target is guaranteed, and the task interruption rate is greatly reduced.

[0054] (2) In the path planning of robot dog inspection, a hierarchical planning method is adopted. The global A* search guides the low-risk macro path at the cost of path length and risk penalty. Combined with local DWA sampling in the velocity space and embedding risk accumulation cost, high-frequency obstacle avoidance is achieved, avoiding the heavy deployment and computational efficiency problems of complex model predictive control (MPC). It can still operate stably under harsh conditions such as dust and high temperature in cement plants, and has low computing power requirements. It is especially conducive to lightweight deployment and operation on the robot dog end.

[0055] (3) Task rescheduling runs on an independent thread, allowing for rapid timeouts and returns to suboptimal solutions, ensuring continuous execution of the main control loop. Through event-driven rescheduling, the task graph is dynamically updated, unreachable points are marked and their priorities are adjusted, and the remaining tasks are reordered while ensuring safety. Compared with fixed-sequence inspection, this can reduce invalid waiting caused by path blockages and improve the completion rate of high-value inspection points.

[0056] Especially for emergencies such as sudden increases in gas concentration or open flames, preset task templates can be invoked, eliminating the need to solve optimization problems and achieving a response time of less than 10ms, which greatly improves the speed of handling sudden dangers.

[0057] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below may be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other. Furthermore, all combinations of the claimed subject matter are considered part of the inventive subject matter of this disclosure.

[0058] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0059] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings.

[0060] Figure 1 This is a flowchart illustrating a method for autonomous obstacle avoidance and task rescheduling for robotic dogs in high-risk areas of cement plants, according to an embodiment of the present invention.

[0061] Figure 2 This is an exemplary workflow diagram of autonomous obstacle avoidance and task rescheduling of a robot dog according to an embodiment of the present invention. Detailed Implementation

[0062] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0063] Various aspects of the invention are described in this disclosure with reference to the accompanying drawings, which illustrate numerous illustrative embodiments. The embodiments of this disclosure are not necessarily intended to encompass all aspects of the invention. It should be understood that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of many ways, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0064] Combination Figure 1 As shown, a method for autonomous obstacle avoidance and task rescheduling of a robotic dog in a high-risk area of ​​a cement plant according to an embodiment of the present invention includes the following steps:

[0065] Step 1: Load the plant area semantic map, process safety knowledge base, and inspection task list; and construct the plant area's basic risk layer based on the static objects in the semantic map according to the process safety knowledge base;

[0066] Step 2: Construct a 2.5D grid map centered on the robot dog, with a front length of H meters and left and right lengths of D meters. Each grid cell stores a risk value between 0 and 1 to represent passability. The initial value is 0, indicating that the robot dog can pass, and 1 indicates that the robot dog cannot pass.

[0067] Step 3: Utilize the multi-sensor fusion perception system on the robot dog to obtain the robot dog's real-time location and the distribution of static obstacles, dynamic objects, high-temperature areas, and gas concentrations in the surrounding area, and generate a dynamic risk layer regarding high-temperature risk, gas risk, and moving object risk.

[0068] Step 4: According to the preset planning cycle, superimpose the basic risk layer and the dynamic risk layer to obtain the current risk field R(x, y); where, if the height of the step exceeds the threshold, the risk is set to 1;

[0069] Step 5: Using the task head point of the task queue as the current task target point, and starting from the robot dog's current position, use the global planner to perform a global search in the grid area of ​​non-static obstacles to plan a global reference path from the current position to the target task point.

[0070] Step 6: Guided by the global reference path, perform local path planning through DWA in each control cycle, and execute the local path planning at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking movement.

[0071] Step 7: Continuously monitor the robot dog's operating status and external input, and when path blocking events, environmental risk events, device status events, and task-specific abnormal events are detected, write the event information to shared memory and set a rescheduling request flag; and

[0072] Step 8: In response to the detection of the rescheduling request flag, the robot dog stores the current state in shared memory and continues to execute the current local path planning. At the same time, it starts a low-priority background thread to solve the background problem, updates the reachability of each task point, re-plans the execution order of the remaining task points, and decides whether to immediately interrupt the current task in order to maximize the completion of the high-priority task.

[0073] In an embodiment of the present invention, during the system startup and initialization phase, before the robot dog enters the plant area, maintenance personnel or users can load the semantic map and process safety knowledge base of the cement plant offline. This allows the high-precision two-dimensional / three-dimensional semantic map of the cement plant to be imported into the robot dog's industrial control computer. Fixed equipment (such as rotary kilns, grate coolers, and fans) and static structures such as walls, stairs, and railings are marked, and a safe distance threshold is preset for each object. For example, the surface temperature of a rotary kiln can reach 400℃, requiring a safe distance ≥ 2.0 meters; the coal mill area has a dust explosion risk, requiring a safe distance ≥ 3.0 meters; and the high-altitude area at the top of the preheater needs to prevent falls, with the path set inside the railing.

[0074] In embodiments of the present invention, the robot dog is equipped with a lidar, an RGB-D depth camera, an infrared thermal imager, a multi-functional gas sensor (detecting CO, SO2, and dust concentration), and an inertial measurement unit (IMU), forming a multi-sensor fusion sensing system. During initial startup, each sensor outputs a self-test status code. Thus, the robot dog's real-time location and the surrounding static obstacles, dynamic objects, high-temperature areas, and gas concentration distribution are obtained through the individual sensors.

[0075] It should be understood that the cement plant's process safety knowledge base records the fixed equipment within the plant area and their safety distance thresholds. Therefore, for each fixed equipment in the semantic map, the risk impact radius and risk decay curve can be pre-determined. For static structures, the risk impact locations are pre-determined; these static structures include walls, stairs, and railings. Combining the risk impact data of fixed equipment and static structures, a basic risk layer for the plant area is constructed.

[0076] In an embodiment of the present invention, a 2.5D grid map is constructed with the robot dog as the center, extending H meters in front and D meters to the left and right. Each grid cell stores a risk value ranging from 0 to 1 to characterize passability, initially set to 0 (passable) and 1 (prohibited). Simultaneously, the spatial resolution of the risk field (0.2-meter grid in this example) and the update frequency (10Hz as an example) are set to initialize the risk field parameters.

[0077] For each grid cell, based on each fixed device in the semantic map, the risk impact radius and attenuation curve are obtained from a table according to its type. The risk value is then determined and filled into the grid cell. For example, taking a rotary kiln as an example, its impact radius is 2.5 meters, and the risk value decreases linearly with distance, taking values ​​of: max(0, 1-d / 2.5), where d is the nearest distance to the kiln surface. For the edge of the guardrail, the risk is 0 on the inner side and 1 (impassable) on the outer side (suspended side). This basic risk layer is loaded when the robot dog starts up and does not change over time, serving as a static base map.

[0078] In this invention, the risk value is normalized and defined as being between 0 and 1:

[0079] 0: Safe area, unobstructed, free of high temperatures and harmful gases;

[0080] 0.5: Passable, for example, near the high-temperature safety boundary;

[0081] 1.0: No passage allowed, including grid areas where obstacles are present, the distance to high-temperature radiation is less than the minimum safe distance, or the concentration of toxic gases exceeds the threshold.

[0082] Furthermore, in step 3, multi-source sensing data is acquired through a multi-sensor fusion sensing system, including surrounding static obstacles, dynamic objects, high-temperature areas, and gas concentration distribution, and a dynamic risk layer is generated regarding high-temperature risk, gas risk, and moving object risk, including:

[0083] Step 3-1: Use the LiDAR to scan the point cloud and match it with the preloaded semantic map. For example, use the Normal Distribution Transform (NDT) algorithm for matching. Combine the angular velocity and acceleration output by the IMU inertial sensor to output the robot dog's pose (x, y, yaw) in the global map.

[0084] Step 3-2: After the LiDAR point cloud is segmented on the ground, continuous points that are 0.1 meters above the ground and cannot be crossed by gait are projected onto the 2.5D grid map and marked as occupied grids.

[0085] Step 3-3: The infrared thermal imager detects and outputs high-temperature pixels, maps them to 3D coordinates, records the surface temperature, and generates the high-temperature risk R_hot according to the exponential decay model.

[0086] R_hot = A_hot * exp(-d / σ_hot);

[0087] Where A_hot represents the base risk value determined by the surface temperature, the higher the temperature, the larger A_hot, up to a maximum of 1.0; d represents the distance from the robot dog to the high-temperature point, and σ_hot is the attenuation distance constant, which is taken as 0.8 for the cement plant environment in this example; for example, for a pipe surface at 300℃, A_hot=1.0, Risk_hot≈0.53 at 0.5 meters, and ≈0.15 at 1.5 meters;

[0088] Steps 3-4: For a grid that detects harmful gas, the risk value is set as R_gas = C / C_threshold; C is the current concentration of harmful gas in the grid output by the gas sensor, and C_threshold is the safety limit. If the concentration is exceeded, R_gas is truncated to 1.0 and diffuses to neighboring grids, with a diffusion gradient of 30% reduction in concentration every 0.5 seconds.

[0089] Steps 3-5: For each detected dynamic object, generate an elliptical risk field around its current position, with the major axis pointing in its direction of motion, to simulate its probability of occupancy in the near future. The risk amplitude is proportional to its speed: for every 1 m / s increase in speed, the risk amplitude increases by 0.2; the semi-major axis of the ellipse = speed × 1.5 seconds, and the semi-minor axis = 0.5 meters.

[0090] In an optional embodiment, a point cloud clustering algorithm (such as Euclidean clustering) can be further employed to separate moving objects (personnel, vehicles, temporary stockpiling forklifts) and calculate their velocity vectors (by dividing the position difference between two consecutive frames by time). The classification criteria are illustrated below:

[0091] If the object's height is greater than 1.5 meters and its width is less than 2 meters, it is identified as a person.

[0092] If the object is longer than 3 meters and exhibits wheel-like movement characteristics, it is identified as a vehicle.

[0093] If an object is stationary but not in the semantic map, it is judged as a temporary material pile.

[0094] In an optional embodiment, the infrared thermal imager acquires the temperature distribution within the field of view at a frequency of 5 frames per second, and maps pixels with temperatures exceeding a set threshold (e.g., 80°C) into 3D space to form a high-temperature surface point cloud.

[0095] Therefore, based on the real-time acquisition of static obstacles, dynamic objects, high-temperature areas, and gas concentration distribution around the robot dog, in step 4, for each 0.2s planning cycle, for each grid, the risk values ​​corresponding to the basic risk layer and the dynamic risk layer are taken as the maximum value according to the grid to obtain the current risk field R(x, y) of each grid; this is used as a risk heat map as a safety constraint for subsequent path planning and task scheduling.

[0096] In the z-direction (i.e., the height direction), only the robot dog's standing plane and the foot contact surface when crossing obstacles are considered. The z-dimensional risk value is defined using a passability label, where a step height < 0.2 meters indicates passability with a risk value of 0, otherwise a risk value of 1.

[0097] If a grid is occupied by a static obstacle, the risk is immediately set to 1.

[0098] In an embodiment of the present invention, during the initialization phase, the robot dog downloads a list of inspection tasks from the host computer (central control room), and each task includes:

[0099] Task point coordinates (x, y, z);

[0100] Task type (e.g., infrared temperature measurement, gas sampling, equipment status image recognition, etc.);

[0101] Priority (levels 1 to 5, with 1 being the highest, such as emergency shutdown for inspection);

[0102] Estimated stay time (seconds);

[0103] Dependencies, for example, the fan bearing temperature must be measured before the fan vibration can be checked.

[0104] As an optional implementation, in step 5, the robot dog uses the head of the inspection task list as the current task target point, and its current position as the starting point. It then uses a global planner to perform a global search within the grid area of ​​non-static obstacles to plan a global reference path from its current position to the target task point, including the following steps:

[0105] Generate and maintain a task queue. Initially, the task queue is sorted according to the initial priority of the inspection task list.

[0106] Using the task point at the head of the team as the target point of the current task, and taking the robot dog's current position as the starting point, the global planner is used to perform a global search in the grid area of ​​non-static obstacles. First, the area to which the task point belongs is found, and then the grid search is performed in the area using the A* search algorithm.

[0107] A grid search is performed within the region using the A* search algorithm, including:

[0108] Based on the A* algorithm, the cost f(n) for each grid n is designed as follows:

[0109] f(n)= g(n) + h(n) + α·R(n);

[0110] Where g(n) represents the actual path length from the starting point to grid n; h(n) represents the Euclidean distance from grid n to the target point, which is used as a heuristic function; R(n) represents the risk value of grid n in the current risk field; α is the risk penalty coefficient, which is determined by the safety level of the high-risk area of ​​the cement plant, representing the additional path cost corresponding to each 0.1 risk value;

[0111] The initial polyline path is obtained by searching for the grid coordinates. After smoothing the path and eliminating the corners, a global reference path is obtained.

[0112] The aforementioned path smoothing uses cubic spline interpolation to smooth the inflection points of the searched polyline path, ensuring that the robot dog does not need to make sharp turns when moving.

[0113] Therefore, based on the known static map and the current risk field, a low-risk reference path from the current location to the target task point is quickly searched through global planning. This path guides the local obstacle avoidance module. Path planning considers not only the shortest distance but also avoids areas near high-temperature equipment and toxic gas zones. For example, although the straight-line distance from the kiln head to the coal mill is short, it requires passing through a high-temperature radiation zone. By increasing the risk penalty coefficient α, the A* global search automatically selects a safe detour area.

[0114] Based on the overall low-risk path determined by global planning, DWA local path rolling planning can be further combined within the high-frequency control cycle to generate smooth and safe speed commands according to the real-time risk field, avoid dynamic obstacles and follow the global reference path.

[0115] In step 6, guided by the global reference path, in each control cycle, a set of speed combinations (v, ω) that can be reached in the next cycle is generated based on the robot dog's current linear velocity v and angular velocity ω. For each candidate speed combination (v, ω), a fixed time is used to extrapolate the motion forward using a uniform motion model to generate multiple trajectories. Local path planning is then performed at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking motion.

[0116] The cost function design for each trajectory includes a fusion cost (Cost) of orientation cost (heading), obstacle distance cost (1 / dist_obs), velocity cost (v_m-v), and risk accumulation cost (risk_sum), specifically expressed as follows:

[0117] Cost=w1·heading+w2·(1 / dist_obs) +w3·(v_m-v)+ w4·risk_sum;

[0118] In the formula, the heading cost represents the angle between the trajectory endpoint orientation and the tangent direction of the global path. The smaller the angle, the better the robot dog can maintain its forward movement along the global path.

[0119] The obstacle distance cost (1 / dist_obs) represents the distance on the trajectory to the nearest static obstacle. If the distance is less than the safe radius (e.g., 0.4 meters), the cost is set to infinity.

[0120] The speed cost (v_m-v) represents the difference between the target speed v_m and the current linear speed v corresponding to the trajectory; a higher linear speed is encouraged to improve inspection efficiency. The target speed v_m is 1.2 m / s for example.

[0121] The risk accumulation cost, risk_sum, represents the cumulative sum of risk values ​​at every 0.1-meter sampling point on the trajectory. This cost enables the robot dog to actively choose a path with lower risk, even if it is slightly longer. Thus, by adding the risk accumulation cost, the robot dog will actively avoid the edge of the high-temperature area instead of walking along the boundary.

[0122] w1, w2, w3, and w4 represent the weighting coefficients of the corresponding cost components.

[0123] As a concrete example, the high-frequency local programming method in step 6 includes the following process:

[0124] In each control cycle (0.05 seconds or 20 Hz), velocity space sampling is performed based on the robot dog's current linear velocity v (0 to 1.2 m / s) and angular velocity ω (-1.5 to 1.5 rad / s). Combined with its acceleration and deceleration capabilities (maximum linear acceleration 0.8 m / s², maximum angular acceleration 2 rad / s²), a set of velocity combinations (v, ω) that can be reached in the next cycle is generated. Taking 30 to 50 candidate velocity pairs as an example.

[0125] Then, for each candidate velocity (v, ω), assuming the robot dog moves at that velocity at a constant speed, a fixed time interval (usually 1.0 second) is extrapolated forward to generate an arc-shaped trajectory. The coordinates of each point on the trajectory are calculated using a kinematic model:

[0126] x(t) = x0 + v·t·cos(θ0 + ω·t);

[0127] y(t) = y0 + v·t·sin(θ0 + ω·t);

[0128] Where θ0 is the current orientation angle;

[0129] Then, four cost components are calculated for each candidate trajectory and weighted summed to obtain the fusion cost (Cost); where, according to the cement plant experiment parameter tuning, w1=0.3, w2=0.5 (if dist_obs<0.4, then automatically infinite), w3=0.1, w4=0.4;

[0130] Then, the candidate speed (v, ω) with the lowest cost is selected, converted into gait instructions for the robot dog (completed by the underlying controller), and sent to the servo motor to drive the robot dog to walk.

[0131] As an optional implementation, if the obstacle distance cost of all candidate trajectories is infinite within 10 consecutive cycles (i.e., no safe path can be found within 0.5 seconds), it is determined that the local planning has failed and the current path to the target point is completely blocked, such as a landslide or temporary scaffolding completely blocking the passage. This triggers a path blocking event, which is reported to the task scheduler. At the same time, the robot dog performs emergency deceleration to a stop and waits in place.

[0132] In an optional embodiment, the robot dog runs an event detection thread in the background. The event detection thread continuously monitors the robot dog's operating status and external inputs at a preset frequency (1Hz), including: path blocking events, environmental risk events, device status events, and task-specific abnormal events.

[0133] Furthermore, upon identifying path blocking events, environmental risk events, device status events, and task-specific abnormal events, the event information is written to shared memory, and a rescheduling request flag is set.

[0134] If multiple events occur simultaneously, they are ordered by priority as follows: environmental risk events > device status events > path blocking events > task-specific abnormal events.

[0135] In an optional embodiment, an environmental risk event is defined as:

[0136] Gas sensor readings rise more than twice the safety threshold within 1 second (e.g., CO concentration suddenly increases from 10 ppm to 50 ppm).

[0137] The infrared thermal imager detected an open flame (a moving or flickering hot spot with a temperature >500℃).

[0138] The dust concentration sensor exceeded the lower explosive limit by 10% (i.e., 10% LEL).

[0139] Environmental risk events have the highest priority and will immediately interrupt the current task upon being triggered.

[0140] In an optional embodiment, the device status event includes:

[0141] Signals from the DCS system in the central control room are received via the factory's industrial Ethernet (using the OPC UA protocol). For example: if the bearing temperature of fan #3 is abnormal, immediate on-site confirmation is required; or the raw material mill is shut down urgently, and nearby inspections are suspended.

[0142] The device status event includes the location coordinates of the event and the emergency level.

[0143] In an optional embodiment, the task's own abnormal events include:

[0144] If the current task check result is abnormal (e.g., temperature measurement shows that the bearing is overheating), additional detection actions are required (e.g., re-measure the temperature, or perform vibration detection).

[0145] The robot dog's battery level is below 20%, and it needs to be put into a return-to-home charging task; or

[0146] The prerequisite task for the current task point has not been completed (for example, vibration must be measured before temperature, but the vibration task is skipped for some reason).

[0147] When any event is detected, the event detection thread writes the event information (type, location, priority, timestamp) to shared memory and sets a rescheduling request flag. If multiple events occur simultaneously, they are ordered by priority.

[0148] In an embodiment of the invention, after an event is detected, the execution order of the remaining task points needs to be replanned, and a decision needs to be made on whether to immediately interrupt the current task to maximize the completion rate of high-priority tasks. Specifically, upon detecting a rescheduling request flag, the task is not immediately blocked, but the current state (machine dog pose, remaining task list, risk field) is sent to an independent background task rescheduling thread and set to low priority. The background thread starts the solver and continues to perform local obstacle avoidance at the original target point (or, if the event is urgent, performs safety braking first). The solver is set to a maximum running time of 2 seconds.

[0149] In a specific embodiment, firstly, based on the current risk field, it is determined whether each remaining task point is reachable. Reachability is defined as: there exists a path from the robot dog's current position to the task point, and the maximum risk value on the path is ≤ 0.8; if no path exists, the task point is marked as skipped.

[0150] Then, the priority is adjusted according to the event type, with the priority order as follows: environmental risk events > device status events > path blocking events > task-specific abnormal events;

[0151] With the objective of minimizing the sum of total time and total risk integral, a genetic algorithm is used to solve for the optimal task sequence, and the updated task sequence is output:

[0152] If the first task in the updated task sequence is the same as the current task, then continue executing the original path;

[0153] If they are different, determine whether the robot dog is moving. If the speed is >0.1 m / s, perform a smooth braking first to stop the robot dog and then use the task point of the first task in the updated task sequence as the target to trigger global path replanning. If the robot dog has already stopped, perform global path replanning directly.

[0154] As a concrete example, the process of finding the optimal task sequence using a genetic algorithm includes:

[0155] Encode each task sequence into a chromosome (skip unreachable tasks);

[0156] Design the fitness function: Fitness = - (total travel distance + average path risk × 100), where the travel distance is estimated based on the pairwise distance matrix pre-calculated by A*;

[0157] The genetic operation is performed with a population size of 50 and an iteration count of 200. The goal is to minimize the sum of the total time consumption and the total risk integral, and output the result task ranking sequence. If the genetic time exceeds the maximum running time of 2 seconds, the current best individual is returned.

[0158] In an optional embodiment, in response to a rescheduling request, the event level can be determined first. If it is a non-urgent event, a background thread is started to solve the problem using a genetic algorithm; if it is an urgent environmental risk, a preset action sequence is directly invoked. For example, for an emergency event involving environmental risk and emergency equipment shutdown, instead of invoking the genetic algorithm, a simple sequence is directly generated: [emergency task point, unfinished high-priority tasks in the original queue, ...], making the rescheduling solution time <10 milliseconds. Therefore, for the environmental risk of an emergency gas leak, inspections must be immediately interrupted and personnel rushed to the leak point to avoid potential safety accidents caused by delays.

[0159] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant, characterized in that, Includes the following steps: Step 1: Load the plant semantic map, process safety knowledge base, and inspection task list; And based on the process safety knowledge base, a basic risk layer for the plant area is constructed from the static objects in the semantic map; Step 2: Construct a 2.5D grid map centered on the robot dog, with H meters in front and D meters to the left and right. Each grid cell stores a risk value of 0 to 1 to represent passability. The initial value is 0, indicating that it is passable, and 1 indicates that it is prohibited from passing. Step 3: Utilize the multi-sensor fusion perception system on the robot dog to obtain the robot dog's real-time location and the distribution of static obstacles, dynamic objects, high-temperature areas, and gas concentrations in the surrounding area, and generate a dynamic risk layer regarding high-temperature risk, gas risk, and moving object risk. Step 4: According to the preset planning cycle, superimpose the basic risk layer and the dynamic risk layer to obtain the current risk field R(x,y); where, if the height of the step exceeds the threshold, the risk is set to 1; Step 5: Using the task head point of the task queue as the current task target point, and starting from the robot dog's current position, use the global planner to perform a global search in the grid area of ​​non-static obstacles to plan a global reference path from the current position to the target task point. Step 6: Guided by the global reference path, perform local path planning through DWA in each control cycle, and execute the local path planning at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking movement. Step 7: Continuously monitor the robot dog's operating status and external input, and when path blocking events, environmental risk events, device status events, and task-specific abnormal events are detected, write the event information to shared memory and set a rescheduling request flag; and Step 8: In response to the detection of the rescheduling request flag, the robot dog stores the current state in shared memory and continues to execute the current local path planning. At the same time, it starts a low-priority background thread to solve the background problem, updates the reachability of each task point, re-plans the execution order of the remaining task points, and decides whether to immediately interrupt the current task in order to maximize the completion of the high-priority task.

2. The method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 1, the process safety knowledge base records the fixed equipment within the plant area and their safety distance thresholds; For each fixed device in the semantic map, the risk impact radius and risk decay curve are pre-determined; For static structures, the locations of risk impacts are determined in advance; the static structures include walls, stairs, and railings. By combining risk impact data of fixed equipment and static structures, a basic risk layer for the plant area is constructed.

3. The method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 2, for each grid cell, the risk impact radius and attenuation curve are obtained by looking up a table based on the type of each fixed device in the semantic map, the risk value is determined and filled into the grid cell.

4. The method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 3, multi-source sensing data is acquired through a multi-sensor fusion sensing system, including surrounding static obstacles, dynamic objects, high-temperature areas, and gas concentration distribution, and a dynamic risk layer is generated regarding high-temperature risk, gas risk, and moving object risk, including: Step 3-1: Use the LiDAR to scan the point cloud and match it with the preloaded semantic map. Combine the angular velocity and acceleration of the IMU to output the pose (x, y, yaw) of the robot dog in the global map. Step 3-2: After the LiDAR point cloud is segmented on the ground, continuous points that are 0.1 meters above the ground and cannot be crossed by gait are projected onto the 2.5D grid map and marked as occupied grids. Step 3-3: The infrared thermal imager detects and outputs high-temperature pixels, maps them to 3D coordinates, records the surface temperature, and generates the high-temperature risk R_hot according to the exponential decay model. R_hot = A_hot * exp(-d / σ_hot); Where A_hot represents the basic risk value determined by the surface temperature, d represents the distance from the robot dog to the high temperature point, and σ_hot is the attenuation distance constant; Steps 3-4: For a grid that detects harmful gas, the risk value is set as R_gas = C / C_threshold; C is the harmful gas concentration value of the current grid output by the gas sensor, and C_threshold is the safety limit. If the value is exceeded, R_gas is truncated to 1.0 and diffuses to neighboring grids. Steps 3-5: For each detected dynamic object, generate an elliptical risk field around its current position, with the major axis pointing in its direction of motion, to simulate its probability of occupancy in the near future. The risk amplitude is proportional to its speed: for every 1 m / s increase in speed, the risk amplitude increases by 0.2; the semi-major axis of the ellipse = speed × 1.5 seconds, and the semi-minor axis = 0.5 meters.

5. The method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 4, within each 0.2s planning cycle, for each grid cell, the risk values ​​corresponding to the basic risk layer and the dynamic risk layer are taken as the maximum value according to the grid cell to obtain the current risk field R(x, y) of each grid cell; In the z-direction, only the robot dog's standing plane and the foot contact surface when crossing obstacles are considered. The z-dimensional risk value is defined using a passability label, where a step height < 0.2 meters indicates passability with a risk value of 0, otherwise a risk value of 1. If a grid is occupied by a static obstacle, the risk is immediately set to 1.

6. The method for autonomous obstacle avoidance and task rescheduling of a robot dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 5, based on the first task point in the inspection task list as the current task target point, and starting from the robot dog's current position, the global planner performs a global search in the grid area of ​​non-static obstacles to plan a global reference path from the current position to the target task point, including the following steps: Generate and maintain a task queue. Initially, the task queue is sorted according to the initial priority of the inspection task list. Using the task point at the head of the team as the target point of the current task, and taking the robot dog's current position as the starting point, the global planner is used to perform a global search in the grid area of ​​non-static obstacles. First, the area to which the task point belongs is found, and then the grid search is performed in the area using the A* search algorithm.

7. The method for autonomous obstacle avoidance and task rescheduling of a robot dog in high-risk areas of a cement plant according to claim 6, characterized in that, In step 5, a raster search is performed within the region using the A* search algorithm, including: Based on the A* algorithm, the cost f(n) for each grid n is designed as follows: f(n)= g(n) + h(n) + α·R(n); Where g(n) represents the actual path length from the starting point to grid n; h(n) represents the Euclidean distance from grid n to the target point, which is used as a heuristic function; R(n) represents the risk value of grid n in the current risk field; α is the risk penalty coefficient, which is determined by the safety level of the high-risk area of ​​the cement plant, representing the additional path cost corresponding to each 0.1 risk value; The initial polyline path is obtained by searching for the grid coordinates. After smoothing the path and eliminating the corners, a global reference path is obtained.

8. The method for autonomous obstacle avoidance and task rescheduling of a robot dog in high-risk areas of a cement plant according to claim 6, characterized in that, In step 6, guided by the global reference path, in each control cycle, a set of speed combinations (v, ω) that can be reached in the next cycle is generated based on the robot dog's current linear velocity v and angular velocity ω. For each candidate speed combination (v, ω), a fixed time is used to extrapolate the motion forward using a uniform motion model to generate multiple trajectories. Local path planning is then performed at the speed corresponding to the trajectory with the minimum cost to drive the robot dog's walking motion. The cost function design for each trajectory includes a fusion cost (Cost) of orientation cost (heading), obstacle distance cost (1 / dist_obs), velocity cost (v_m-v), and risk accumulation cost (risk_sum), specifically expressed as follows: Cost=w1·heading+w2·(1 / dist_obs) +w3·(v_m-v)+ w4·risk_sum; In the formula, the heading cost represents the angle between the trajectory endpoint orientation and the global path tangent direction; The obstacle distance cost (1 / dist_obs) represents the distance on the trajectory to the nearest static obstacle. If the distance is less than the safe radius, the cost is set to infinity. The velocity cost (v_m-v) represents the difference between the target velocity and the current linear velocity v corresponding to the trajectory; The cumulative risk cost, risk_sum, represents the sum of risk values ​​accumulated over sampling points every 0.1 meters along the trajectory. w1, w2, w3, and w4 represent the weighting coefficients of the corresponding cost components.

9. The method for autonomous obstacle avoidance and task rescheduling of a robot dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 7, the event detection thread continuously monitors the robot dog's operating status and external inputs at a preset frequency, including: path blocking events, environmental risk events, device status events, and task-specific abnormal events. Furthermore, upon identifying path blocking events, environmental risk events, device status events, and task-specific abnormal events, the event information is written to shared memory, and a rescheduling request flag is set. If multiple events occur simultaneously, they are ordered by priority as follows: environmental risk events > device status events > path blocking events > task-specific abnormal events.

10. The method for autonomous obstacle avoidance and task rescheduling of a robotic dog in high-risk areas of a cement plant according to claim 1, characterized in that, In step 8, starting a low-priority background thread to solve the problem in the background, updating the reachability of each task point, re-planning the execution order of the remaining task points, and deciding whether to immediately interrupt the current task to maximize the completion rate of the high-priority task specifically includes the following steps: First, based on the current risk field, determine whether each remaining task point is reachable. Reachability is defined as: there exists a path from the robot dog's current position to the task point, and the maximum risk value on the path is ≤ 0.8; if no path exists, mark the task point as skipped. Then, the priority is adjusted according to the event type, with the priority order as follows: environmental risk events > device status events > path blocking events > task-specific abnormal events; With the objective of minimizing the sum of total time and total risk integral, a genetic algorithm is used to solve for the optimal task sequence, and the updated task sequence is output: If the first task in the updated task sequence is the same as the current task, then continue executing the original path; If they are different, determine whether the robot dog is moving. If the speed is >0.1 m / s, perform a smooth braking first to stop the robot dog and then use the task point of the first task in the updated task sequence as the target to trigger global path replanning. If the robot dog has already stopped, perform global path replanning directly.

Images