Obstacle crossing method and system for tracked robot

Abstract

The present invention relates to the field of robot obstacle crossing technology, and discloses a method and system for tracked robots to cross obstacles, including: extracting multi-dimensional features and accurately identifying obstacle types from data collected by sensors based on an intelligent analysis algorithm; optimizing and regulating track drive speed difference instructions based on the identification results to ensure the dynamic balance of the robot in complex terrain; utilizing an adaptive update mechanism to adjust key variables in the control algorithm to avoid track slippage on different slopes; and improving the robot's passability in narrow or dense obstacle spaces through precise regulation of path planning priority allocation logic. The solution of the disclosed embodiment can solve the problem of insufficient accuracy in intelligently analyzing data collected by robot sensors to identify obstacle types, as well as the problem of unstable robot balance when crossing complex terrain and easy track slippage during climbing.

Description

Technical Field

[0001] The present invention relates to the technical field of robot obstacle crossing, and in particular to an obstacle crossing method and system for a tracked robot. Background Art

[0002] The obstacle crossing capability of a tracked robot is mainly improved through intelligent technology and algorithm optimization, which can enhance the robot's ability to autonomously cross obstacles in complex environments. However, in practical applications, there are still some challenges and problems to be solved:

[0003] First, how to improve the accuracy of the sensor's recognition of different types of obstacles, as this is crucial for the robot's path selection and safe avoidance capabilities. Second, in complex terrain conditions, such as hillsides or gravel, adjusting the speed differences of the track drive is crucial for maintaining the robot's balance. Furthermore, the track is prone to slipping when climbing different slopes, necessitating improvements to the control algorithm's adaptive mechanism to dynamically adjust relevant variables to maintain stable driving performance.

[0004] Efficient navigation of confined spaces and densely packed obstacles is also a challenge for the navigation system when planning paths, requiring more precise allocation of logical priorities to improve maneuverability. Furthermore, effective power management based on energy consumption models during long missions is a major challenge, ensuring optimal energy consumption during the robot's obstacle-crossing mission. Addressing these issues will help further improve the practicality and reliability of this type of tracked robot. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] A method and system for overcoming obstacles of a tracked robot, comprising:

[0007] Based on intelligent analysis algorithms, the system extracts multi-dimensional features from sensor data and accurately identifies obstacle types. Based on the identification results, it optimizes and controls the track drive speed difference command to ensure the robot's dynamic balance in complex terrain.

[0008] An adaptive update mechanism is used to adjust key variables in the control algorithm to prevent track slippage on different slopes. Precise control of path planning priority allocation logic improves the robot's maneuverability in narrow or densely obstructed spaces.

[0009] The robot sensor data is pre-filtered and denoised, and a convolutional neural network (CNN) is used for initial obstacle classification: after obtaining the original image and depth sensor input, a bilateral filter is used to reduce random noise I = bilateral(I), where I represents the original image. Primary object classification is performed through CNN. If P(class)>θ, that is, the probability is higher than the threshold, where P is the probability that the obstacle is a specific class and θ = 7 is the set threshold. Finally, the preliminary classification information is used as one of the precursors for further intelligent analysis.

[0010] Preferably, a nonlinear proportional controller is introduced in the process of optimizing the adjustment of the crawler drive speed difference to ensure dynamic balance:

[0011] After collecting the instantaneous posture change angle Δα data, the maximum allowable angle α_max is determined;

[0012] When it is detected that Δα>δ (δ is the minimum valid angle increment), the compensation strategy is started;

[0013] Adjust the speed commands Vleft and Vright of the left and right tracks to satisfy the equation: ΔV = K*(sin(Δα / 2)), where K is the control gain coefficient. The specific value is adjusted in real time according to the application scenario to ensure that the overall movement is stable while achieving the optimal obstacle crossing capability.

[0014] Preferably, the adaptive update mechanism incorporates a slope estimation step to address issues arising from varying terrain conditions:

[0015] The tilt angle β is monitored in real time using the outputs of the gyroscope and accelerometer;

[0016] If abs(β)>βlim, that is, the absolute slope exceeds the limit, where βlim ranges from 15° to 30°, then enter the special walking mode;

[0017] The actual slope angle ψ is estimated according to the algorithm tan(ψ) = |gyro_y / g|, where g is the acceleration due to gravity and gyro_y is the angular velocity on the Y axis. This is used to fine-tune the track tension and reduce the possibility of slippage.

[0018] Preferably, a path cost evaluation function is introduced to improve the traversal performance in dense areas:

[0019] In the initialization stage, the initial cost of all grid points is set to Cost_0;

[0020] When encountering an obstacle boundary, increase the cost of adjacent grid nodes. Cost = Cost_pre + w * Dist(obs, node), where the Dist function returns the Euclidean distance between two points, and w is the weight coefficient, with a value greater than 1 to ensure sensitivity to the risk of nearby obstacles. The specific value is determined according to the environmental safety requirements. Once the path is generated, the best path is selected based on the cumulative cost value. This method can significantly improve the safety and efficiency of navigation.

[0021] Preferably, combine the reinforcement learning (RL) method to train the optimal behavior decision model to improve the response speed in complex terrain:

[0022] Set the reward function reward = c * (1 / sqrt(dist_to_closest_obstacle)) – t * speed_change_cost, where the parameter c is a positive coefficient, and its value range is optimized according to the specific application scenario; t is a non-zero negative number representing the cost of direction change energy, and its value needs to balance efficiency and energy consumption;

[0023] A certain reward will be given whenever an obstacle is crossed or bypassed;

[0024] Use the QLearning update formula Q(s,a) < Q(s,a) + α * [reward + γmaxQ(s,a) - Q(s,a)] to iteratively approximate the optimal action selection. Here, s represents the state, a is the action taken, α is the learning rate, and γ is the discount factor, which can enhance the machine learning ability and promote the automatic improvement process.

[0025] Preferably, implement the dynamic environment perception function to more effectively respond to unexpected situations:

[0026] Continuously monitor the position Pos_other of other moving entities in the surrounding environment;

[0027] Assume that the relative velocity v_rel between two objects is defined as dP / dt, and measure v_relative using the Doppler effect.

[0028] Preferably, predict future terrain changes by integrating the moment of inertia algorithm to prevent possible skidding in advance:

[0029] Periodically collect the rotational speed Ω from the gyroscope;

[0030] Calculate the total external torque τ from J * dΩ / dt = ∑τ.

[0031] Preferably, add a battery power management module for energy consumption modeling to extend the battery life and improve the energy efficiency ratio:

[0032] Statistically calculate the average power demand Power_avg according to the historical task pattern;

[0033] The estimated energy E_est required to complete a specific path is E_est = Power * time_spent;

[0034] For energy conservation considerations, if E_used > E_remaining / (1 + risk_margin), the path is re-planned. Here, E_used is the energy consumed, E_remaining is the remaining available energy of the battery, and risk_margin is the emergency reserve percentage, set at 10% - 30%. Then the path is re-planned. This strategy helps to maximize the operation range and maintain a low working burden.

[0035] Preferably, a hybrid drive architecture is adopted, combined with an electromagnetic brake and a hydraulic buffer structure to enhance stability in sharp turn and high-load scenarios:

[0036] Configure the main motor to cooperate with the auxiliary magnetic powder clutch to form a complete drive system;

[0037] Design the working switching logic under emergency braking conditions as f(t) = max(Pe, Pm), where Pe is the maximum torque value generated by the motor and Pm corresponds to the resistance torque generated by the magnetic braking part;

[0038] If F_applied < F_required, the reaction force acting on the ground is less than the required friction coefficient multiplied by the vehicle gravity, that is, it cannot stop safely. Enable the liquid pressure provided by the reservoir to apply a large-range braking pressure Fb until it returns to the controllable range.

[0039] Preferably, a multi-scale mapping mechanism is added to strengthen the understanding of the fine structure of the local terrain and improve the construction of the global map:

[0040] Create multi-level sub-region maps, including the fine layer Tile_High, the general layer Tile_Middle, and the schematic layer Tile_Low;

[0041] For each new detection result D, apply the fusion rule R: if scale(D) < Smin goto L_high; else if Smax < scale(D) goto L_low; otherwise classify it into the middle level, where the specific thresholds of Smin and Smax are determined comprehensively according to the actual sensor accuracy, computing hardware capabilities, and application scenario requirements. This mapping relationship helps to refine the specific details around the path and simplify the remote observation objects.

[0042] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0043] 1. Based on intelligent analysis algorithms, multi-dimensional feature extraction and precise obstacle identification are performed on sensor data. Based on the identification results, track drive speed differential commands are optimized and controlled to ensure the robot's dynamic balance in complex terrain. Key variables in the control algorithm are adjusted using an adaptive update mechanism to prevent track slippage on varying slopes. Precise control of path planning priority allocation logic improves the robot's maneuverability in narrow or densely packed obstacle spaces. The solutions of the disclosed embodiments address the issue of intelligent analysis of data collected by robot sensors to address the lack of accuracy in obstacle identification.

[0044] 2. This invention solves the problem of intelligently analyzing and extracting features from data collected by robot sensors. Accurately identifying obstacle types in complex environments is a significant challenge, especially when obstacles vary in shape and background information is complex. To address these issues, the present invention establishes a method framework based on advanced intelligent algorithms. This framework can exploit the implicit characteristics of sensor data from multiple perspectives, and combines learning results from a large number of known cases for efficient calculation and real-time processing. This allows for accurate obstacle classification and three-dimensional position estimation, ensuring a reliable basis for subsequent operational decisions. This fundamentally improves the consistency and accuracy of the system's judgment of various obstacle types. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Schematic diagram of the specific process of the present invention;

[0046] Figure 2 Schematic diagram of the original image and sensor algorithm flow of the present invention;

[0047] Figure 3 A schematic diagram of the posture change process flow is collected for the present invention;

[0048] Figure 4 This is a schematic diagram of the gyroscope and acceleration flow of the present invention;

[0049] Figure 5 This is a schematic diagram of the initialization phase flow of the present invention;

[0050] Figure 6 It is a schematic diagram of the function flow of the present invention;

[0051] Figure 7 This is a schematic diagram of the surrounding environment monitoring process of the present invention;

[0052] Figure 8 This is a schematic diagram of the process of collecting rotation speed by the gyroscope of the present invention;

[0053] Figure 9 This is a schematic diagram of the power requirement flow of the task of the present invention;

[0054] Figure 10 This is a schematic diagram of the process of forming the magnetic powder clutch of the present invention;

[0055] Figure 11 This is a schematic diagram of the multi-level sub-area map process of the present invention; DETAILED DESCRIPTION

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0057] See also Figure 1-11 The present invention provides a method and system for tracked robots to cross obstacles. The method first extracts multi-dimensional features from the data collected by sensors and accurately identifies the types of obstacles based on an intelligent analysis algorithm. The robot is equipped with a variety of sensors, such as lidar, cameras, infrared sensors, and ultrasonic rangefinders. These sensors can continuously obtain information about the surrounding environment in different environments and transmit the large amount of collected data to the central processing unit. Specifically, the intelligent analysis algorithm implements two important steps by processing these sensor data: one is data cleaning, that is, filtering out anomalies and noise; the other is feature engineering, extracting multiple dimensional characteristics including shape, color, and texture. For example, in a specific instance, the robot encountered an area of ​​weeds. It used the camera to capture images and combined them with a plant recognition model trained by machine learning to quickly determine that this was grass rather than an obstacle with different characteristics. This accurate classification capability is the basis for ensuring subsequent obstacle-crossing behavior.

[0058] After identifying the obstacle type based on the aforementioned characteristics, the approach then shifts to optimizing the track drive speed differential to ensure the robot's dynamic balance in complex terrain. Tracked structures offer superior obstacle-crossing performance because they can adjust their trajectory and inclination angles by adjusting the speed differential between the two sides to adapt to varying terrain. However, improper configuration can increase the risk of rollover or even rollover, necessitating precise modulation. In a practical operation, the engineering team designed a feedback control loop that senses the current posture in real time and uses an algorithm to calculate the ideal acceleration and deceleration commands, ensuring a smooth and stable motion without excessive fluctuations that could affect overall safety and stability. For example, when faced with an uneven road with varying surface friction, the system adjusts the speed parameters of the left and right tracks, taking into account the gravitational acceleration component and the moment of inertia factor. This ensures that the vehicle maintains its center of gravity at the optimal point throughout its intended trajectory, avoiding unnecessary deviations and oscillations, and ensuring a smooth transition to flat terrain to resume normal operation without disruption or damage.

[0059] It is particularly important to utilize an adaptive update mechanism to adjust key parameters within the control algorithm to effectively overcome track slippage during hill climbing. When faced with the challenges of varying slope angles, traditional fixed mechanical compensation methods struggle to adapt to the ever-changing demands of actual application scenarios, increasing the likelihood of slippage and threatening mission reliability. The solution involves introducing an intelligent learning component that iterates online based on historical case studies and on-site perception, dynamically adjusting threshold ranges that are beneficial for increasing ground pressure to achieve enhanced adhesion. Specifically, if a construction vehicle attempts to ascend a steep mountain road at a 45-degree angle, this mechanism will activate. Based on a pre-set weighting function, an appropriate proportional factor will be selected to amplify the pressure applied to the two chains, ensuring adequate grip and preventing idling and loss of traction, leading to an inability to successfully reach the summit and ultimately complete construction-related activities.

[0060] Then, in order to further enhance smooth passage in small-scale or narrow spaces, the path planning algorithm must follow specific rules to provide optimal decision recommendations. Typically, map data is discretized into many small nodes to represent possible paths for traversal. Each route has its own corresponding priority weight index, which is automatically evaluated, scored, and ranked by the computer in order to select the best one that is most conducive to rapid evacuation from dangerous situations without colliding with surrounding obstacles or getting stuck, wasting precious rescue opportunities. For example, suppose an emergency search and rescue team member carries equipment deep into an underground pipeline to perform a detection and search mission. Since the passage itself is relatively narrow, with sharp turns and frequent complex and changeable layouts, it is necessary to rely on navigation software to formulate a detailed plan in advance, taking into account all potential adverse factors. After repeated simulations and calculations, a coherent and compact ideal trajectory that takes into account both safety and time efficiency is selected to provide users with a reference basis, thereby greatly improving work efficiency, reducing the number of human errors, and improving overall performance.

[0061] Finally, considering that efficient energy utilization is crucial for long distances or when traversing obstacles for extended periods, a dynamic energy consumption prediction model has been developed for this purpose. Based on real-time monitoring, configuration adjustments are made to reduce inefficient power consumption and improve overall economic efficiency. Embedded high-performance metering devices continuously record the power consumption of various components. Using big data mining, a comprehensive behavioral pattern prediction curve is then analyzed and summarized. This predicts the required power reserve and provides timely adjustment recommendations, such as reducing unnecessary function activation, lowering operating frequency, and shortening charging cycles. This not only extends battery life and reduces maintenance and replacement frequency, but also promotes the practical implementation of green development concepts, contributing to environmental protection and sustainable social development goals. Imagine a fleet of unmanned transport vehicles busy shuttling materials back and forth between mines. Their daily mileage is long, the workload is demanding, and battery life is critical. With the aforementioned technology, it is possible to rationally plan routes and allocate resources, maintain efficient operation, and minimize energy consumption and environmental impact, reducing operating costs, saving costs for the owner and generating greater value for shareholders. This solves the problem of limited maneuverability in confined spaces or with closely spaced obstacles, and optimizes energy management.

[0062] Within this framework, all the steps involved are interconnected and complementary to each other, jointly constructing a complete tracked robot obstacle-crossing walking solution, effectively solving a series of difficult technical bottlenecks, achieving remarkable application results, and demonstrating strong vitality and broad prospects.

[0063] Next, the claims and guidelines of the present invention are described:

[0064] The first step is to expand and refine the content of the first claim, clarifying the specific parameters and limitations of the technology in the specific application environment. This means starting from a narrower and more specific perspective to set a benchmark framework for subsequent operations. For example, in the context of a tracked robot's obstacle-crossing method, the material of the track can be considered, stipulating that it should have a certain degree of heat resistance and flexibility to ensure that it will not be damaged when traversing complex terrain. Specifically, the elongation at break of this material should be greater than or equal to 30%. This range can both adapt to environmental changes and ensure that material costs are controlled within a reasonable range.

[0065] The second step is to develop action guidelines based on the defined technical details. These guidelines typically involve the application of abstract logic such as algorithms or physical rules. Taking the aforementioned example, in one embodiment, the tracked robot relies on visual sensors to identify obstacles ahead and uses this data as a trigger to determine whether to raise the chassis. The threshold for raising the chassis is set to occur when the sensor determines that the height difference exceeds the maximum allowable ground clearance of 8 cm in the robot's normal walking posture. This is a relatively scientific value calculated based on the overall design of the machine and its motion requirements.

[0066] The third step is to add a layer of verification or adjustment factors to the behavioral model formed in the previous step. For this tracked equipment, this may include introducing a feedback system to monitor whether the lifting effect during actual travel is within the expected safety level. If the test results show that the current solution may cause structural instability, the response strategy is re-planned by changing the contact surface size or other related factors according to the formula P = F / m (P represents pressure; F represents force; m refers to the mass of the force-bearing part per unit area).

[0067] The fourth step is to integrate all of these elements into a coherent, integrated process. The resulting solution must not only meet the original purpose of the invention but also operate efficiently in real-world applications. In one example, after performing a series of obstacle avoidance maneuvers, the robotic arm must return to its initial position to prepare for the next task. This process requires smooth and precise movement, while minimizing energy consumption.

[0068] Next, the present invention will be described in detail. The method for overcoming obstacles of a crawler robot according to the present invention includes the following steps:

[0069] The first step is to process the robot's sensor data, pre-filtering and denoising the acquired data. To ensure the original image more accurately reflects real-world information, a bilateral filter is used to reduce random noise contamination caused by various environmental or device factors. Here, I = bilateral(I) is a formula, where the parameter I represents the original image input, and the function returns the denoised image. This operation helps improve the accuracy of subsequent obstacle recognition. The purpose of setting this formula is to preserve edge features while reducing unnecessary interference.

[0070] Then, after acquiring and processing the image, the depth sensor is used to obtain additional spatial dimension information input. These preparatory tasks are all aimed at more accurately identifying obstacles that may appear in the environment. Preliminary classification judgment is performed through a convolutional neural network that has been trained and is suitable for the current application scenario. When an object is determined to belong to a specific type, P(class) represents the probability output value under this category; and θ is set to 0.7 as the critical point to screen out more reliable results. P(class)>θ means that if the P(class) value corresponding to a detected obstacle category is higher than the given threshold of 0.7, it means that the model has a high level of confidence in this judgment.

[0071] Finally, a preliminary classification conclusion is drawn. The information obtained from the above processing serves as the basic material and can pave the way for further more complex analysis and processing.

[0072] This process ensures that tracked robots can better adapt to varying terrain conditions, detect potential obstacles promptly and accurately, and ensure safe and efficient movement. This enables tracked robots to complete their assigned tasks more stably and reliably, expanding their application scenarios and scope.

[0073] Next, the present invention will be described. In the tracked robot's obstacle-crossing method, the robot's posture change data must first be acquired and processed in real time. Specifically, as the robot traverses complex terrain, the sensor system continuously monitors and collects data on the instantaneous posture change angle Δα, ensuring that subsequent calculations and adjustments are based on accurate information. During this phase, the maximum allowable tilt angle α_max is determined, which represents the maximum threshold for the robot to maintain dynamic stability without tipping over.

[0074] To achieve stable obstacle-crossing behavior, the algorithm sets a minimum valid angle increment, δ, and implements a check: if Δα > δ is detected, the compensation strategy is immediately initiated. δ sets a small but significant angle value as sensitivity, ensuring that any posture changes that could affect stability are promptly addressed. This detection and judgment is essential, for example, when encountering uneven terrain or slopes.

[0075] Subsequently, the speed difference between the left and right crawler drives is optimized and adjusted based on the collected Δα data. This process uses a nonlinear proportional control method. After adjustment, the speed commands Vleft and Vright of the left and right crawler drives satisfy the formula: ΔV = K*(sin(Δα / 2)). Where:

[0076] -ΔV is the speed difference between the tracks;

[0077] -K is a non-zero constant proportional control gain. K ranges from 0.01 to 1.5 m / s / rad. The optimal value depends on different application scenarios. For example, a smaller range can be selected when fast movement is required, while a larger range can be selected when higher precision is required.

[0078] The -sin function is used to limit instability caused by overcorrection. By converting half the attitude offset angle to radians and then applying a sine function, the model is simplified and the control effect is smoothed. This is because larger angles correspond to greater correction forces, which gradually decrease until the vehicle approaches the horizontal position.

[0079] Finally, while ensuring the overall stability of the entire motion system, optimal obstacle-crossing capabilities are achieved. Specifically, this includes maintaining forward direction while appropriately correcting the trajectory to avoid obstacles and smoothly traversing uneven terrain. In one embodiment, consider a tracked exploration vehicle operating in rugged mountainous terrain. Faced with complex environments, the aforementioned method allows the device to flexibly turn even in a high roll state, and safely and smoothly traverse obstacles formed by rocks of varying sizes without becoming unstable or stalling.

[0080] This solution combines precise perception, real-time control, and intelligent decision-making to enhance the autonomous operating efficiency of tracked robots and their adaptability to changing working conditions. By precisely managing speed differentials, these devices maintain excellent balance and maneuverability regardless of terrain.

[0081] Next, the present invention further refines the tracked robot obstacle crossing method mentioned in the claims. First, in this process, the tilt angle β is monitored in real time using a gyroscope and an accelerometer. Here, the gyroscope can sense changes in angular velocity to determine changes in orientation, and the accelerometer is used to sense acceleration conditions including gravity to assist in determining the angle. This step is intended to provide basic data support for subsequent adjustments, ensuring that the slope information is collected as accurately as possible so that the machine can respond in a timely manner.

[0082] If the absolute value of the tilt angle (abs(β)) exceeds a set limit (βlim), the device enters a special travel mode to protect the equipment and more smoothly cross obstacles. This threshold (βlim) is a preset safety limit. It can be set between 15-30°, depending on factors such as track material and terrain type, with a preferred setting of approximately 20° to balance sensitivity and stability.

[0083] Then, the actual slope angle ψ is calculated based on the mathematical model tan(ψ) = |gyro_y / g|, where g represents 9.8 m / s. 2The gravitational acceleration is represented by gyro_y, which represents the angular velocity measured along the Y axis. Its absolute value is taken here to obtain the positive ratio in the tangent function, which in turn derives the inclination angle. This formula allows for a quick estimation of the angle of the ground relative to the horizontal without knowing the current height or length. This method of slope estimation is chosen because it allows for rapid and reasonable determination based solely on the built-in sensor, minimizing misjudgments caused by external interference.

[0084] Finally, based on the slope angle information obtained above, the system applies appropriate tension to each track and attempts to reduce the risk of slippage. For example, if a steeply rising dirt slope is detected, in one embodiment, the system increases the pressure of the upper track against the underlying dirt based on the previously calculated larger angle, while slightly relaxing the lower track to allow a certain degree of bending to adapt to the slope curve. This improves grip efficiency, prevents slippage, and ensures the robot can successfully climb over it.

[0085] Next, we'll describe one embodiment of the present invention. When a tracked robot executes an obstacle-crossing method, it first sets an initial cost for each grid point on the path. This occurs before the robot begins planning. When initializing the entire map area, all possible traversal grid points are assigned the same initial cost value, Cost_0. For example, for a standard 10×10 grid map, assuming both robot and obstacle information have been preloaded into the system, all feasible locations that haven't yet been affected by actual terrain (i.e., nodes that are traversable but haven't yet encountered specific conditions) will default to the same initial cost value.

[0086] When an obstacle boundary is detected, the cost weight is adjusted based on the relative spatial geometry between the current grid point and the nearest static or dynamic obstructing object to account for the risk of approaching the obstacle. The Dist function calculates the distance from the node to the obstacle boundary. This range depends on the actual scale of the scene; for example, a greater number of pixels per unit area results in a smaller minimum non-zero distance. A multiplication factor, w, is introduced to linearly amplify the risk cost. When the distance is shorter and closer to the obstacle, a greater weight is added (w>1); vice versa. Ideally, w takes a fixed, appropriately large positive value to ensure sensitivity to near-critical situations without over-magnifying distant hazards. The formula is: Cost = Cost_pre + w*Dist(obs,node), where w is determined based on the robot's application scenario and requirements. Higher weights are used for situations where environmental safety requirements are high and adjustable paths are redundant.

[0087] After the path generation process is completed, a forward route with the lowest overall cost is selected based on the accumulated cost calculations in the previous steps. The accumulated total cost of each node reflects the comprehensive evaluation result under a series of local optimal decisions. After the entire journey simulation is completed or the goal is reached ahead of schedule, the overall performance of multiple rounds of different direction plans is compared to confirm the optimal solution, thereby improving the safety and reliability of the navigation process and saving time. Taking a 15*15 map as an example, if there are two sets of paths A and B, after traversing these paths, the total accumulated cost values ​​of the two paths are compared to select the one with the smaller value as the preferred recommended walking route. This mechanism allows tracked automatic equipment to effectively avoid complex terrain structures and achieve mission objectives more smoothly, significantly improving its operational performance.

[0088] The above process aims to optimize the behavior of autonomous mobile devices in complex terrain conditions. Through rational strategy design, it significantly improves mission success rates while enhancing adaptability and stability in the face of unknown challenges. It also emphasizes how to balance immediate sensory feedback with pre-planned layout to achieve the optimal path matching solution.

[0089] Next, we'll describe how the tracked robot, when performing an obstacle-crossing task, first sets up and trains an optimal behavioral decision-making model using reinforcement learning methods to improve its response speed and decision-making accuracy in complex terrain. The core of this step is to construct a reasonable environmental model and, based on this, establish an appropriate reward mechanism to guide the behavioral learning process.

[0090] The robot generates its current state s by sensing the distance to the nearest obstacle in its surroundings. The parameter `dist_to_closest_obstacle` defines the distance from the robot to the closest obstacle. To ensure that the robot can handle complex terrain conditions efficiently, the reward function reward is set equal to a positive coefficient c multiplied by the inverse square root of the distance to the obstacle minus the speed change cost representing the energy cost of changing direction. The positive coefficient c here represents the rate of increase in the score obtained the closer to the obstacle; and the negative factor t indicates the additional energy cost required for each direction adjustment. This formula encourages the machine to move quickly while ensuring safety, while avoiding unnecessary or too frequent turning operations that consume too many resources. The value of c is usually greater than zero, the range depends on the application scenario requirements, and can be optimized according to actual conditions; t is a non-zero negative number used as a lever to adjust the flexibility of the strategy. Ideally, a balance should be maintained between efficiency and energy consumption.

[0091] Each time the robot completes bypassing or crossing an obstacle at a specific location, it will be given a positive evaluation score. The purpose of this design is to strengthen beneficial behaviors and encourage the exploration of better solution paths. Subsequently, the update rule is used in the Q-learning algorithm: the previous Q value is added to the new increment α, which is weighted by the difference between the immediate feedback and the discounted future expected maximum reward. In this process, s and a represent the set of choices made by the system at different times and states; α, as an important factor for adjusting the weight of new information, generally takes a small positive real value such as around 0.1; γ, as an indicator of the importance of future predictions, should also choose a probability value less than 1 and close to 1 but not 1. For example, the default value of 0.95 can be used.

[0092] The action update method described above enables the algorithm to gradually converge on a set of best practice guidelines, ultimately achieving autonomous adaptive improvement. For example, in a real-world implementation, if a tracked unmanned vehicle is navigating a wilderness area filled with gravel, trenches, and other unforeseen obstacles, this reinforcement learning-based approach will continuously try new navigation routes and accumulate experience from each action feedback. Specifically, the vehicle may initially choose a safer approach (away from potential threats). Through continuous trial and error experience and memory consolidation, it gradually learns to identify areas suitable for rapid movement and areas that require caution or special skills—for example, climbing a hill without taking a detour, or making a sharp turn at the right time to better enter the next straight track. This not only reduces the total time required to complete the task but also improves the intelligence and flexibility of the entire system.

[0093] Next, the obstacle-crossing method of the tracked robot according to the present invention is described, which specifically involves a series of continuous dynamic environment perception steps to deal with unexpected situations:

[0094] The first step is to continuously monitor the positions of other moving entities in the environment. The Pos_other parameter represents the spatial coordinates of all moving objects around the robot. For example, a lidar sensor mounted on top of a crawler robot continuously sends signals. These signals reflect off obstacles and return to the sensor, allowing the robot to record and track the positions of other moving obstacles.

[0095] The relative speed v_rel is then defined based on the rate of change of the relative displacement between the two objects. Here, v_rel = dP / dt means that after each time period Δt, the distance between the other object and the tracked robot will change by a certain amount ΔP; dP represents the position difference, that is, the degree of change in the distance between the two, and dt is a very short time interval. The Doppler effect is used to measure the frequency change caused by this relative motion, and then indirectly obtain the specific speed value v_rel of the two objects approaching or moving away from each other. For example, when a tracked vehicle is approaching a car that is also moving in a different direction, the data obtained by the sensor can be analyzed by a built-in algorithm to accurately determine the current relative speed and direction of the two, providing data support for possible emergency treatment measures.

[0096] The formula dTTC = min((|d / P_rel|^2)*safe_margin) < ε, mentioned in the next step, is used to assess collision risk. Here, TTC refers to the Time-To-Collision (TTC). min selects the minimum of all candidate values ​​as part of the judgment criteria; |d / P_rel| refers to the minimum foreseeable probability of contact (in seconds) given the ratio of the current safe distance d to the relative speed; and safe_margin is the manually defined safety margin required to avoid collision. In most cases, a setting of 1 to 3 seconds is appropriate, depending on the specific workplace conditions and safety level requirements. If the calculated result is less than the predetermined error value ε, it indicates a high probability of an imminent collision or that the vehicle is approaching too close to another moving object, and the danger level warning level is reached.

[0097] The last step requires the immediate transmission of an alarm signal to the control system so that the path planning strategy can be adjusted in time. When it is determined that the potential collision risk has reached a critical state, for example, a pedestrian suddenly appears to cross the road quickly not far ahead, causing the calculated collision time to be far less than the preset minimum safety boundary value (such as dTTC = 1s and lower than the system default value), the controller should immediately initiate the avoidance action plan after receiving the prompt information - either reducing the travel speed until it stops, or quickly repositioning the navigation path to bypass the obstacle area, ensuring that the obstacle crossing task is successfully completed without compromising the life and property of other traffic participants. The purpose of designing this entire dynamic environment recognition mechanism is to ensure that the crawler machine can respond more efficiently and intelligently to the challenges brought by various complex environments during the entire operation.

[0098] First, while the gyroscope is collecting rotational velocity, this step requires periodic acquisition of the mechanical system's rotational state parameter, Ω. The gyroscope can sense changes in the tracked robot's posture in real time and updates this data at a specific sampling interval (e.g., 1ms to 50ms) to obtain the most accurate possible attitude change rate. This continuous recording of Ω ensures that the algorithm can track sudden environmental changes. In one embodiment, rotational velocity data is collected at the highest rate for subsequent calculations when the tracked robot passes over a rock pile.

[0099] Secondly, calculations related to the moment of inertia are performed based on known physical quantities. J represents the ability of an object to resist rotational changes or a quantitative representation of the inertia of the system. This value is usually provided by the manufacturer; dΩ / dt reflects the instantaneous angular acceleration vector. According to J*dΩ / dt=∑τ, the sum of all external rotational force factors τ that cause the overall rotation effect of the system is analyzed, that is, the torsional effect of the entire robot due to external forces is solved. Specifically, if the robot drives down from the top of the slope into the flat area, this equation can evaluate the impact of the terrain difference on the motion trajectory. Since the value of J varies greatly depending on the specific design and configuration, under the optimal design scheme, a mass distribution form that is larger but does not affect the balance between structural stability and flexibility will be selected to achieve optimal controllability.

[0100] Then, if a risk of slip is detected, the pre-set safety mechanism is immediately activated. abs(J*(final_velocityinitial_velcity) / dt) represents the absolute difference in the average kinetic energy increase or decrease over a period of time, reflecting the amplitude of the object's velocity fluctuations. ∑tau is the set of all externally applied torques required to overcome relative displacement within the device. slip_torque is an empirical limit indicator obtained from prior experimental testing, used to characterize the maximum threshold allowed by the friction conditions between the tire and the supporting surface. Once the former exceeds the latter, it indicates that signs of wheel separation are imminent or have already occurred. At this point, the control program should be activated to reduce driving force to prevent accidents and prevent the uncontrollable situation from further deteriorating. For example, if encountering a surface with insufficient friction, such as slippery grass, power output should be promptly limited to maintain stable forward movement.

[0101] Next, the present invention describes the implementation of energy consumption modeling based on task requirements and the addition of a battery power management system module:

[0102] The first step is to analyze and utilize data from past missions to calculate the average power requirement, Power_avg. The average power represents the average energy consumption per unit time when the robot has performed similar tasks in the past. This value is derived from historical data accumulated over long periods of operation.

[0103] The subsequent step involves estimating the total energy cost of the entire task, E_est = Power * time_spent, based on the average power obtained and the time spent on the predefined route. Here, time_spent represents the estimated time required to complete the given route (in seconds or minutes), and Power uses the average power from historical statistics as an approximation or the expected maximum power (in watts). This formula provides a preliminary understanding of the energy requirements of the upcoming task, ensuring that underestimated energy consumption does not lead to unsuccessful task completion. The optimal setting depends on the specific circumstances of each operating environment.

[0104] Another step is to monitor real-time energy usage. When the actual energy used (E_used) exceeds the current remaining battery capacity (E_remaining) divided by (1 + risk_margin), a new route must be planned immediately to ensure sufficient power for subsequent operations and to reserve power for emergency situations. The risk_margin is the percentage of energy reserved for emergency situations, and is generally recommended to be set between 0.1 and 0.3, depending on the specific situation. This not only helps to address potential errors but also ensures system stability and operational efficiency.

[0105] Next, the first step of describing the present invention is to configure the drive system:

[0106] In this step, the crawler robot combines the main motor with an auxiliary magnetic powder clutch to form a hybrid drive architecture, and introduces electromagnetic brakes and hydraulic buffer structures to enhance stability under sharp turns and high load conditions. This means that the robot can have stable performance in different driving scenarios. Specifically, in a crawler robot application, for example, when operating in a complex terrain environment such as a rocky area, the hybrid drive system ensures that the machine can quickly adjust its motion state after encountering obstacles such as large rocks and respond quickly, achieving better operating accuracy.

[0107] The second step is to define the working switching logic under emergency braking conditions, using the formula f(t) = max(Pe, Pm). Here, Pe refers to the maximum torque value that the motor can output. The general range depends on the motor power level. For medium-sized robots, it may be between 0 and several Nm; and Pm corresponds to the torque value generated by magnetic braking. The purpose of this formula is to determine which force can more quickly and effectively cause the vehicle to decelerate to a stop. In one embodiment, assuming that Pe = 30 Nm and Pm = 25 Nm, in certain special circumstances, such as facing an emergency and needing to stop immediately, the larger of the two, that is, 30 Nm, is used to judge the braking effect. This ensures that in an emergency, all current braking means can be maximized to achieve the deceleration target as quickly as possible.

[0108] The third step is to check whether a safe stop is possible. Calculate the relationship between the ground reaction force F_applied and the required frictional force F_required (the coefficient of friction multiplied by the vehicle's gravity). If there is a situation where F_applied < F_required, it means that it is impossible to ensure a reliable stationary state relying on the existing contact surface characteristics, and it is necessary to further enhance the braking performance to prevent accidental slipping. For example, when the robot is traveling on a muddy road section and the effective downward normal pressure exerted by its weight on a specific road surface decreases due to load or angle problems, resulting in insufficient friction to support the stopping action, the next measure will be triggered to intervene and supplement the braking force to restore the operation within the control range. If it is detected at this time that the acting reaction force is less than the required amount, that is, in the case of being unable to stop safely,

[0109] then take the fourth step, which is to increase the braking force to restore the normal operating performance range. This operation is carried out by the reservoir supplying high-pressure fluid to push the braking mechanism to generate an additional strong force Fb for suppression until F_applied is increased back to not less than or even exceeding F_required. For example, when the machine travels on soft soil and the tires slip and it is difficult to stop normally, this emergency treatment plan is activated to provide uniform and continuously increasing braking air pressure to each wheel to ensure the overall structure is stable until sufficient conditions to support a smooth landing and stop are re-established. The above steps fully demonstrate the important strategies and process details for ensuring safety and stability during the obstacle crossing of the tracked robot.

[0110] Next, describe the establishment of the multi-scale mapping (MSM) mechanism of the present invention to strengthen the understanding of the fine structure of the local terrain, thereby optimizing the global map construction. This mechanism splits the overall environmental information into maps with different levels of detail: from the most detailed fine layer (Tile_High) to the relatively rough general layer (Tile_Middle) and then to the general overview layer (Tile_Low), ensuring that both the fine features in the close range and the overall trends in a large area can be taken into account at any given moment.

[0111] For each new round of detection data D, during the processing, the fusion rule R is used to determine which level of the map these data should belong to. There are thresholds Smin and Smax, where scale(D) represents the spatial scale or resolution degree of the area involved in the detection result; if scale(D) < Smin, it means that the currently obtained information has a very high resolution and is suitable for the fine layer Tile_High to accurately capture the local small-scale terrain features, that is, the trigger condition if scale(D) < Smin goto L_high. If the scale exceeds Smax, that is, when scale(D) > Smax, it means that relatively large-scale but less accurate geographical features are obtained, so it should be classified as the rough layer Tile_Low of the low-detail level, and the trigger condition if Smax < scale(D) goto L_low. All other data that do not meet the above two extreme cases are directly classified into the intermediate general layer, and in this way, the smooth transition and connection between the sub-region maps of each level are completed.

[0112] For example, in one embodiment, when the tracked robot approaches a complex section composed of boulders and small obstacles, the device can quickly sense the dense objects nearby and allocate them to the fine layer for further accurate description, while for the larger and flatter surface contours in the distance, it is only necessary to simply record them in the general layer or even the rough layer to be sufficient for navigating the traveling direction and route planning. This not only saves computing resources but also improves efficiency.

[0113] In addition, in terms of parameter setting, no specific values of Smin and Smax are defined, which depends on the actual situation, such as the required map accuracy, sensor type, and computing hardware capabilities, etc., under the combined effect of multiple factors. The definition of the optimal value will also vary with the application scenario; generally speaking, the best state to be achieved is to ensure sufficient accuracy without causing an excessive processing load, and ensure that the MSM can work efficiently and reasonably. In this way, this obstacle-crossing method can make timely and accurate judgments when encountering various different obstacles and terrain conditions.

[0114] An obstacle-crossing method for a tracked robot of the present invention includes multiple key steps and technical solutions to solve important problems such as how the tracked robot can efficiently and safely cross obstacles in complex terrain, maintain its own dynamic balance, improve passability, and optimize energy consumption.

[0115] First, the invention solves the problem of intelligent analysis and feature extraction of data collected by robot sensors. When faced with complex environments, accurately identifying obstacle types is a huge challenge, especially when the obstacles are of different shapes and the background information is complex. To address these problems, the present invention has established a method framework based on advanced intelligent algorithms, which can explore the implicit characteristics of the data sent back by sensors from multiple aspects, and combine the learning results of a large number of known cases for efficient calculation and real-time processing, to achieve accurate obstacle category determination and three-dimensional position estimation, to ensure that subsequent operational decisions have a basis for reference; thereby fundamentally improving the consistency and accuracy of the system's judgment of various types of obstacles, effectively making up for the shortcomings caused by the uncertainty errors in previous identification methods, and greatly improving the accuracy of obstacle identification.

[0116] Secondly, the present invention establishes a technical innovation in maintaining the stability of the robot by further optimizing the speed difference control rules of the drive device based on the results of the above-mentioned precise identification. Traditional adjustment methods often rely on preset fixed response curves, making it difficult to flexibly adjust speed parameter values ​​when encountering unknown height differences or unexpectedly inclined surfaces. This can cause the entire platform to become unstable or even tip over. To overcome these difficulties, a dynamic learning feedback closed-loop control system architecture is adopted. By continuously collecting information on the robot's motion posture angle change signals and the fluctuation of the ground reaction torque, and then leveraging a knowledge base derived from deep reinforcement training, the relative difference in the speed of the two tracks and its gradual change are finely adjusted. This ensures that the overall center of gravity remains within a reasonable range and does not move beyond the limits of the extreme boundary conditions. This achieves the goal of maintaining stable forward movement, whether driving straight on flat roads or turning and circling on hilly terrain, successfully achieving dynamic balance and stability when traversing complex terrain.

[0117] Next, the technical details for preventing slip under climbing conditions of varying slopes are highlighted. This method incorporates an adaptive update mechanism. By quantitatively modeling an evaluation system based on multiple factors, including actual load weight sensing and tire grip friction monitoring under the current operating environment, and drawing on previously accumulated experience and experimental database data, a reasonable initial parameter set is established before each new attempt. The numerical combinations of these dimensions are then fine-tuned and optimized as the operating conditions evolve. For example, upon detecting signs of increased resistance on an uphill slope, engine torque is promptly and appropriately increased to compensate for the traction required to overcome gravitational potential energy. If a potential slip signal (such as an excessive speed difference alarm) is detected, preventive intervention measures are immediately triggered, such as reducing acceleration or activating electronic anti-skid functions, until all associated risks are eliminated. This ensures consistent adhesion under all specified conditions, minimizing track slippage during climbing on varying slopes and ensuring reliable and safe climbing.

[0118] Finally, targeted path planning strategies were developed to improve the robot's ability to navigate narrow spaces or areas with densely packed obstacles. The core idea is to establish a set of high-dimensional discrete point array structure expressions that comprehensively consider the effects of multiple constraints (such as the minimum allowable turning radius, the maximum lateral offset tolerance, and the distribution pattern of obstacle spacing density) to characterize the optimal feasible route to be sought. A heuristic search optimization algorithm is then applied to find the global optimal solution trajectory. Specifically, the order of obstacle avoidance measures is arranged according to the priority of task urgency: first, large solid debris with sharp edges that are easily scratched is avoided, and then a path is planned to pass through the relatively open small gaps between them until the target point is migrated. In this way, not only is the probability of success in narrow and curved tunnels increased, but the time loss rate caused by frequent U-turns is also effectively reduced.

[0119] In addition, to address the issue of maintaining good energy utilization during long-term operations, this invention also involves an energy consumption estimation model that can automatically sense the correlation between the current work progress percentage and the remaining power reserve, and provide appropriate response suggestions based on the gain and loss scores of predetermined standard thresholds. Based on this model, the allocation ratios of various key performance parameters of the power supply system can be regularly recalibrated, such as the battery charge and discharge current intensity control limit value, the sleep mode trigger frequency setting criteria, etc. In this way, the power loss can be reduced to the lowest possible level without affecting the overall mission efficiency, meeting the needs of continuous and stable energy supply in long-distance transportation or rescue and search activities, thereby improving the endurance and reliability of the crawler robot in performing complex tasks.

[0120] In summary, the patent of this invention covers a series of comprehensive and complete measures to comprehensively enhance the ability of tracked robot groups to cope with challenges in diverse outdoor environments.

[0121] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and alterations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for overcoming obstacles of a crawler robot, characterized in that: include: Based on intelligent analysis algorithms, the system extracts multi-dimensional features from sensor data and accurately identifies obstacle types. Based on the identification results, it optimizes and controls the track drive speed difference command to ensure the robot's dynamic balance in complex terrain. An adaptive update mechanism is used to adjust key variables in the control algorithm to prevent track slippage on different slopes. Precise control of path planning priority allocation logic improves the robot's maneuverability in narrow or densely obstructed spaces. The robot sensor data is pre-filtered and denoised, and a convolutional neural network (CNN) is used for initial obstacle classification. After obtaining the original image and depth sensor input, a bilateral filter is used to reduce random noise I = bilateral(I), where I represents the original image. The CNN performs preliminary object classification. If P(class) > θ, the probability is above a threshold, where P is the probability of the obstacle being a specific class and θ = 7 is the set threshold. Finally, the preliminary classification information is used as one of the precursors to further intelligent analysis. A nonlinear proportional controller is introduced to optimize the adjustment of the track drive speed difference to ensure dynamic balance: After collecting the instantaneous posture change angle Δα data, the maximum allowable angle α_max is determined; When it is detected that Δα>δ, δ is the minimum valid angle increment, and the compensation strategy is started; Adjust the speed commands Vleft and Vright of the left and right tracks to satisfy the equation: ΔV = K*(sin(Δα / 2)), where K is the control gain coefficient. The specific value is adjusted in real time according to the application scenario to ensure that the overall movement is stable while achieving the optimal obstacle crossing capability. The adaptive update mechanism incorporates a slope estimation step to address issues caused by varying terrain conditions: The tilt angle β is monitored in real time using the outputs of the gyroscope and accelerometer; If abs(β)>βlim, that is, the absolute slope exceeds the limit, where βlim ranges from 15° to 30°, then enter the special walking mode; The actual slope angle ψ is estimated according to the algorithm tan(ψ) = |gyro_y / g|, where g is the acceleration of gravity and gyro_y is the angular velocity on the Y axis.

2. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: The path cost evaluation function is introduced to improve the traversal performance in dense areas: In the initialization stage, the initial cost of all grid points is set to Cost_0; When encountering the boundary of an obstacle, the cost of the adjacent grid node is increased by Cost = Cost_pre + w*Dist(obs,node). The Dist function returns the Euclidean distance between two points. w is the weight coefficient, which is greater than 1 to ensure risk sensitivity to close obstacles. The specific value is determined according to the environmental safety requirements.

3. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: Combine reinforcement learning methods to train the optimal behavior decision model to improve the response speed in complex terrain: Set the reward function reward = c*(1 / sqrt(dist_to_closest_obstacle))–t*speed_change_cost, where parameter c is a positive coefficient whose value range is determined by optimizing the specific application scenario; t is a non-zero negative number representing the energy cost of the change of direction, and its value needs to balance efficiency and energy consumption; A certain reward is given whenever an obstacle is crossed or bypassed; Use the QLearning update formula Q(s,a) < Q(s,a) + α * [reward + γmaxQ(s,a) - Q(s,a)] to iteratively approximate the optimal action selection, where s represents the state, a is the action taken, α is the learning rate, and γ is the discount factor.

4. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: Implement a dynamic environment perception function to more effectively handle unexpected situations: Continuously monitor the position Pos_other of other moving entities in the surrounding environment; Assume that the relative velocity v_rel between two objects is defined as dP / dt, and measure v_relative using the Doppler effect.

5. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: Predict future terrain changes through the integrated moment of inertia algorithm to prevent possible skidding in advance: Periodically collect the rotational speed Ω from the gyroscope; Calculate the total external torque τ from J * dΩ / dt = ∑τ, where J is the moment of inertia of the robot about the drive axis.

6. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: Added a battery power management module for energy consumption modeling to extend the battery life and improve the energy efficiency ratio: Statistically calculate the average power demand Power_avg based on the historical task pattern; Estimate the energy E_est required to complete a specific path as E_est = Power * time_spent; For energy conservation considerations, if E_used > E_remaining / (1 + risk_margin), then re-plan the path, where E_used is the energy consumed, E_remaining is the remaining available energy of the battery, and risk_margin is the emergency reserve percentage, set to 10% - 30%.

7. The method for overcoming obstacles of a crawler robot according to claim 1, characterized in that: Adopt a hybrid drive architecture combined with electromagnetic brakes and hydraulic buffer structures to enhance stability in sharp turns and high-load situations: Configure the main motor with an auxiliary magnetic particle clutch to form a complete drive system; Design the working switching logic under emergency braking conditions as f(t) = max(Pe, Pm), where Pe is the maximum torque value generated by the motor, Pm corresponds to the resistance torque generated by the magnetic braking part, and f(t) is the total braking torque actually issued at time t; If F_applied < F_required, that is, the reaction force acting on the ground is less than the required friction coefficient multiplied by the vehicle gravity, meaning it cannot stop safely. F_applied is the actual braking force between the current tire / track and the ground, and F_required is the minimum required braking force calculated based on the friction coefficient μ and the vehicle weight mg; Apply a large-range braking pressure Fb using the liquid pressure provided by the reservoir. The specific value of Fb is determined according to the actual braking demand until it returns to the controllable range.

8. A system for a tracked robot obstacle-crossing method according to any one of claims 1 to 7, characterized in that: Add a multi-scale mapping mechanism to strengthen the understanding of the fine structure of the local terrain and improve the construction of the global map: Create multi-level sub-region maps, including the fine layer Tile_High, the general layer Tile_Middle, and the schematic layer Tile_Low; For each new detection result D, apply the fusion rule R: if scale(D) < Smin, go to L_high; elseif Smax < scale(D), go to L_low; otherwise, classify it into the intermediate level, where the specific thresholds of Smin and Smax are determined comprehensively according to the actual sensor accuracy, computing hardware capabilities, and application scenario requirements.

Images