Track picking robot stability judgment and control method based on random forest regression

Abstract

The invention discloses a crawler picking robot stability judgment and control method based on random forest regression. The method comprises the following steps: firstly, acquiring angles, angular velocities and angular accelerations of all joints of the mechanical arm, an inclination angle of a crawler chassis and constraining force and moment of force at the joint of the mechanical arm and the chassis; secondly, a heterogeneous feature selection mechanism is constructed through a multi-output random forest regression model, and X / Y / Z axial force and X / Y axial torque are predicted; secondly, establishing an overturning stability evaluation model for the crawler chassis, and calculating a stable region; and finally, comparing the actual support area with the range of the stable boundary to judge the system stability. And the mechanical arm movement path with the best stability is selected, and if the overturning danger exists in the movement process, the overall stability is maintained by adjusting the joint speed and the acceleration which generate the maximum overturning moment. According to the method, real-time response to the dynamic load is achieved through the multi-target machine learning model, the anti-overturning capacity of the crawler chassis is improved, and the method is suitable for the field of orchard picking robots and has remarkable application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of agricultural robots, and specifically relates to a stability judgment and control method for a crawler harvesting robot based on random forest regression, which is particularly suitable for optimizing the dynamic stability of a robotic arm and a mobile platform when working together in complex terrain. Background Art

[0002] Tracked harvesting robots are widely used in complex agricultural scenarios such as orchards and greenhouses. Their robotic arms need to frequently adjust their posture during operation to complete the picking task. However, the dynamic load changes of the robotic arms and the undulations of the terrain can easily increase the risk of overturning of the mobile platform. Traditional stability judgment methods are mostly based on static mechanical models, such as the center of gravity projection method or the torque balance method, but such methods are difficult to adapt to the dynamic load changes caused by the movement of the robotic arm in real time, and do not consider the coupling effect between the multi-degree-of-freedom robotic arm and the mobile platform. In addition, the existing technology lacks the ability to predict the constraint force and torque in real time, resulting in a delayed control response and an inability to effectively prevent sudden instability.

[0003] In recent years, machine learning technology has been introduced into the field of robot control, but its application in stability prediction still has the following problems:

[0004] Limitations of single-target prediction: Existing models mostly predict a single physical quantity (such as overturning moment), and it is difficult to simultaneously output multi-dimensional constraint forces and moments; Insufficient data preprocessing: Sensor data under dynamic conditions is prone to missing or noisy, and traditional interpolation methods lead to reduced model accuracy; Poor real-time performance: Complex models have a large amount of calculation and are difficult to meet the real-time control needs of field operations.

[0005] Therefore, there is an urgent need for an efficient and robust dynamic stability judgment and control method to improve the operational safety and adaptability of crawler picking robots. Summary of the Invention

[0006] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a stability judgment and control method for a crawler harvesting robot based on random forest regression. The multi-output random forest model is used to predict the constraint force and torque at the connection between the robot arm and the mobile platform in real time, and the dynamic stability criterion is combined to optimize the robot arm motion parameters, thereby improving the overall stability of the system.

[0007] The present invention solves the existing technical problems by adopting the following technical solutions. A crawler picking robot stability judgment and control method based on random forest regression includes the following steps: Step 1, collecting dynamic parameters of the crawler picking robot during the movement of the manipulator arm, including the angle, angular velocity, angular acceleration of each joint of the manipulator arm, the inclination angle of the crawler chassis, and the constraint force and torque at the connection between the manipulator arm and the mobile platform; Step 2, constructing a multi-output random forest regression model, whose input is the joint angle, angular velocity, angular acceleration and chassis inclination angle, and the output is the predicted constraint force (Fx', Fy', Fz') and predicted torque (Mx', My'); Data processing, implementing segmented Z-score normalization based on the operation scenario, the formula is: z = (x-μ) / σ. α is the chassis roll angle. When |α|>5°, the slope mode standardized parameters μ_slope and σ_slope are used. When |α|≤5°, the flat mode parameters μ_flat and σ_flat are used. Step 4: Determine the actual support area of ​​the crawler chassis based on the signal from the axle pressure sensor. Compare the actual support area with the stability boundary in real time to determine the robot's stability. Compare the distance between the stability boundary and the actual support boundary in each support direction to determine the direction in which the manipulator system will tip over. Step 5: When the manipulator end position is known, select the manipulator motion path with the best stability using the stability criterion as a constraint. If there is a risk of tipping during manipulator motion, dynamically adjust the manipulator's speed to optimize overall system stability.

[0008] Furthermore, the robotic arm is a 6-DOF serial robotic arm, and the first joint is equipped with a force and torque sensor to detect the force and torque applied by the robotic arm to the chassis.

[0009] Furthermore, the specific implementation method of step 1 is: through the encoder, IMU and force / torque sensor, the angle, angular velocity and angular acceleration of each joint of the robotic arm are collected, and the tilt angle of the chassis, as well as the constraint force (X / Y / Z axis) and torque (X / Y axis) at the connection between the first joint of the robotic arm and the mobile platform are measured.

[0010] Furthermore, the multi-output random forest regression model in step 2 is specifically implemented as follows: During the data preprocessing phase, missing values ​​are filled using median interpolation, and features are normalized using the Z-score normalization method; z = (x - μ) / σ, where x is the original data point, μ is the feature mean, and σ is the standard deviation. When |α| > 5°, the slope mode normalization parameters μ_slope and σ_slope are used; when |α| ≤ 5°, the flat mode parameters μ_flat and σ_flat are used.

[0011] Furthermore, a heterogeneous feature selection mechanism was constructed, and a feature subset with maximum information coefficient (MIC) > 0.6 was used for the force prediction channel, and a feature subset with Pearson correlation coefficient greater than 0.5 and VIF < 5 was used for the torque prediction channel.

[0012] Furthermore, through bootstrap sampling and parallel computing, grid search (GridSearchCV) is used to optimize model parameters and select the best model configuration. The best hyperparameter combination is:

[0013] max_depth=None; max_features='sqrt'; min_samples_split=2;

[0014] n_estimators=200.

[0015] Furthermore, 3-fold cross-validation was used to evaluate the model performance to ensure the generalization ability of the model on different datasets.

[0016] Furthermore, the stability evaluation model applicable to the crawler chassis in step 3 is:

[0017] M·e+(F×H)·e+(G×h)·e+Fz·(h+s·cosθ)·e=0

[0018] Where M is the torque of the manipulator acting on the crawler chassis, F is the horizontal force of the manipulator acting on the crawler chassis, G is the gravity of the chassis, and F is the horizontal force of the manipulator acting on the crawler chassis. z is the vertical force exerted by the manipulator on the crawler chassis. H is the moment arm from F to the overturning axis. h is the distance from the chassis' center of gravity to the overturning axis in the horizontal plane, which determines the boundary of the stable region. s is the horizontal distance from the chassis' center of gravity to the contact point between the manipulator's first joint and the chassis. e is the unit vector in the direction of the overturning axis. θ is the angle between the overturning axis and the X-axis. h is used to determine the extent of the stable region.

[0019] Furthermore, the actual support area of ​​the tracked chassis in step 4 is: a thin film pressure sensor is arranged at the wheel axle of the track wheel, the ground wheel is determined according to the signal of the pressure sensor, the connecting line between adjacent ground wheels determines the boundary of the actual support area, and the shortest distance from the center of gravity of the chassis to the actual support boundary in the horizontal plane is d.

[0020] Furthermore, the judgment logic of the stability state in step 4 is:

[0021] If the shortest distance from the chassis center of gravity to the actual support boundary is greater than the distance to the stability boundary, it is considered stable;

[0022] If the shortest distance from the chassis center of gravity to the actual support boundary is equal to the distance from the chassis center of gravity to the stability boundary, it is judged to be critically stable;

[0023] If the shortest distance from the center of gravity of the chassis to the actual support boundary is less than the distance to the stability boundary, it is judged to be unstable.

[0024] Furthermore, in step 4, the distances between the stable boundary and the actual support boundary in the direction of each support edge are compared to determine the direction in which the robotic arm system overturns.

[0025] Furthermore, the specific method for dynamically adjusting the robot arm's speed in step 5 is as follows: when the robot arm's end position is known, the robot arm's motion path with optimal stability is selected using a stability criterion as a constraint. During robot arm motion, if a risk of tipping over is detected, the joint generating the maximum tipping moment can be identified by calculating the tipping moment generated by the motion of each joint. Subsequently, the speed and acceleration of that joint can be appropriately reduced to effectively reduce the risk of tipping over. By dynamically adjusting the robot arm's speed, the stability of the entire system is optimized.

[0026] The advantages and positive effects of the present invention are:

[0027] The present invention uses a multi-output random forest regression model to predict the constraint force and torque at the connection between the robotic arm and the mobile platform in real time, and optimizes the robotic arm motion parameters in combination with dynamic stability criteria, solving the problem that traditional static models cannot adapt to dynamic load changes; provides an efficient and robust dynamic stability judgment and control method, significantly improving the safety and efficiency of operations under complex terrain; realizes the robotic arm's active change of its own position and posture to reduce the risk of system overturning, effectively improving the anti-overturning ability of the crawler picking robot, and improving the environmental adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is an overall flow chart of the stability judgment and control method of the crawler picking robot based on random forest regression of the present invention.

[0029] Figure 2 Flowchart for constructing the random forest regression prediction model of the present invention.

[0030] Figure 3 This is a comparison chart of the predicted values ​​and true values ​​of the random forest model of the present invention.

[0031] Figure 4 This is a flow chart of the stability control of the crawler picking robot of the present invention.

[0032] Figure 5 This is a structural diagram of a crawler picking robot.

[0033] Figure 6Schematic diagram of the actual support area of ​​the chassis on the ground in a stable state.

[0034] Figure 7 Schematic diagram of the actual support area of ​​the chassis on the ground in an unstable state.

[0035] Figure 8 Schematic diagram of the actual support area of ​​the chassis on the ground under inclined road conditions. DETAILED DESCRIPTION

[0036] The following describes an embodiment of the present invention in detail with reference to the accompanying drawings. This embodiment takes a crawler-type picking robot applied to an orchard environment as an example. Its robotic arm has a 6-DOF serial structure, and the first joint is equipped with a force and torque sensor.

[0037] According to the above crawler picking robot system, the present invention proposes a crawler picking robot stability judgment and control method based on random forest regression, such as Figures 1 to 7 As shown, the implementation process of the present invention includes the following steps:

[0038] Step 1. Dynamic parameter acquisition and preprocessing: Collect the motion parameters of the robotic arm and measure the angle, angular velocity, and angular acceleration of the robotic arm through encoders, IMUs, and six-dimensional force / torque sensors. At the same time, measure the chassis tilt angle and the force and torque applied by the robotic arm to the chassis.

[0039] Construct a multi-output random forest regression prediction model, such as Figure 2 The raw data is processed, and the input features (angular velocity, angular acceleration, and angle) are normalized using the Z-score according to the following formula: z = (x - μ) / σ, where x is the raw data point, μ is the feature mean, and σ is the standard deviation. When |α| > 5°, the slope mode normalization parameters μ_slope = -0.02 and σ_slope = 10.41 are used. α is the chassis roll angle. When |α| ≤ 5°, the flat mode parameters μ_flat = 0.03 and σ_flat = 2.89 are used.

[0040] A heterogeneous feature selection mechanism was constructed. For the force prediction channel, a feature subset with a maximum information coefficient (MIC) greater than 0.6 was used to ensure that the selected features had a strong dependency on force prediction. For the torque prediction channel, a feature subset with a Pearson correlation coefficient greater than 0.5 and a VIF less than 5 was used to ensure that the selected features had strong linear correlation and low multicollinearity. A MultiOutputRegressor was used to wrap the RandomForestRegressor to construct a multi-target prediction architecture. Bootstrap sampling and parallel computation were enabled, and hyperparameter combinations were optimized through grid search (GridSearchCV). Hyperparameter optimization ensured the stability of the random forest model, avoiding overfitting by limiting the maximum depth and minimum number of split samples, and ensuring that the model could effectively handle multiple output targets.

[0041] The optimal hyperparameter combination is:

[0042] max_depth=None; max_features='sqrt'; min_samples_split=2;

[0043] n_estimators=200.

[0044] The model was evaluated using 3-fold cross validation, and the scoring indicators were negative mean square error (NMSE) and determination coefficient (R 2 ), and the mean square error (MSE) is calculated for each target variable on the test set.

[0045] In step 3, the stable area of ​​the crawler picking robot under different road conditions is calculated:

[0046] M·e+(F×H)·e+(G×h)·e+Fz·(h+s·cosθ)·e=0

[0047] Where M is the torque of the manipulator acting on the crawler chassis, F is the horizontal force of the manipulator acting on the crawler chassis, G is the gravity of the chassis, and F is the horizontal force of the manipulator acting on the crawler chassis. z is the vertical force applied by the manipulator arm to the crawler chassis. H is the moment arm from F to the rollover axis. h is the distance from the chassis' center of gravity to the rollover axis in the horizontal plane, which determines the boundaries of the stable region. s is the horizontal distance from the chassis' center of gravity to the contact point between the manipulator's first joint and the chassis. e is the unit vector in the direction of the rollover axis. θ is the angle between the rollover axis and the X-axis. The value of h represents the distance between the chassis' center of gravity and the rollover axis, and h is used to determine the extent of the stable region.

[0048] In the vertical direction, the distance from the contact point between the first joint of the robot arm and the chassis to the overturning axis is 0.85 meters, the mass of the chassis is 115 kilograms, the distance between the center of mass of the chassis and the contact point in the Y-axis direction is 0.133 meters, and the acceleration of gravity is g.

[0049] The actual support area of ​​the crawler chassis depends on the contact area between the crawler and the ground. Under different road conditions, the actual support area is complex and changeable. The ground wheel is determined according to the signal of the pressure sensor, and the connection line between adjacent ground wheels determines the boundary of the actual support area, such as Figure 5-7 shown.

[0050] Under different road conditions, the contact surface between the track and the ground changes, and the angle θ between the overturning axis and the X-axis is obtained through the pressure sensor signal.

[0051] When 0≤θ<90°,

[0052] M=M x cosθ+M y sinθ,F=F x sinθ-F y cosθ

[0053] When 90°≤θ<180°,

[0054] M=-M x cosθ-M y sinθ,F=F x sinθ-F y cosθ

[0055]

[0056] Where L is F z The moment arm to the overturning axis, α is the roll angle of the chassis, and β is the pitch angle of the chassis. The distance h from the center of mass of the chassis to the stability boundary is obtained.

[0057] The thin film pressure sensor is placed at the wheel axle of the track wheel. The ground wheel is determined based on the signal of the pressure sensor. The line between adjacent ground wheels determines the boundary of the actual support area. The shortest distance d from the center of gravity of the chassis to the actual support boundary in the horizontal plane is calculated.

[0058] The stability determination logic in step 4 is as follows: if the distance from the chassis center of gravity to the actual support boundary is greater than the distance to the stability boundary, the system is considered stable; if the distance from the chassis center of gravity to the actual support boundary is equal to the distance to the stability boundary, the system is considered critically stable; if the distance from the chassis center of gravity to the actual support boundary is less than the distance to the stability boundary, the system is considered unstable. The distance between the stability boundary and the actual support boundary in each support edge direction is compared to determine the direction in which the robotic arm system would tip over.

[0059] The specific method for dynamically adjusting the robot arm speed in step 5 is as follows: when the end position of the robot arm is known, multiple motion trajectories can be obtained through inverse kinematics. The stability of the robot arm in each path is calculated using the overturning stability criterion, and the robot arm motion path with the best stability is selected.

[0060] During the movement of the robotic arm, it may encounter interference from external factors such as vibration or unexpected forces, which can affect its stability. When the risk of overturning is detected, the overturning moment generated by the movement of each joint can be calculated to identify the joint generating the greatest overturning moment. Subsequently, the speed and acceleration of this joint can be appropriately reduced to effectively reduce the risk of overturning. By dynamically adjusting the robotic arm's motion parameters, the stability of the entire system can be optimized to effectively cope with these interferences.

Claims

1. A method for judging and controlling the stability of a crawler harvesting robot arm based on a random forest regression model, characterized in that: The following steps are involved: Step 1: Use the embedded IMU array to collect the three-axis angular velocity ω and angular acceleration α of each joint of the robotic arm in real time at a sampling rate of 200Hz ， The dynamic force (Fx, Fy, Fz) and torque (Mx, My) at the connection between the first joint and the crawler chassis are obtained through a strain gauge six-dimensional force sensor; Step 2: Build a multi-output random forest regression model whose input is joint angle, angular velocity, angular acceleration and chassis tilt angle, and output is the predicted constraint force (Fx', Fy', Fz') and predicted torque (Mx', My'); Step 3: Based on the moment stability criterion, combined with the predicted restraint force and moment, and the changes in the actual support area of ​​the crawler chassis under different road conditions, a stability evaluation model suitable for the crawler chassis is established; Step 4: Determine the actual support area of ​​the crawler chassis based on the signal of the pressure sensor at the axle, compare the actual support area of ​​the crawler chassis with the stability boundary in real time, and judge the stability state of the robot; The distance between the stable boundary and the actual support boundary in each support edge direction is compared to determine the direction in which the robotic arm system overturns. Step 5: When the end position of the robot arm is known, the robot arm motion path with the best stability is selected using the stability criterion as a constraint. If there is a risk of tipping during the robot arm's motion, the robot arm's speed is dynamically adjusted to optimize the stability of the entire system.

2. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: Data processing in step 2: Standardization: Implement segmented Z-score standardization based on the job scenario α is the roll angle of the chassis; When |α|>5°, the slope mode is used to standardize the parameters μ_slope and σ_slope; When |α|≤5°, the flat-earth model parameters μ_flat and σ_flat are used.

3. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The training of the multi-output random forest regression model includes: constructing a heterogeneous feature selection mechanism, using a feature subset with a maximum information coefficient (MIC) greater than 0.6 for the force prediction channel, and using a feature subset with a Pearson correlation coefficient greater than 0.5 and a VIF less than 5 for the torque prediction channel.

4. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The optimal hyperparameter combination is: max_depth=None; max_features='sqrt'; min_samples_split=2; n_estimators=200.

5. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The output target variables of step 2 include: X / Y / Z axial constraint force (Fx, Fy, Fz) and X / Y axial torque (Mx, My) at the connection between the first joint of the robotic arm and the mobile platform, a total of 5 physical quantities.

6. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: In step 3, the stable boundary distance of the crawler picking robot under different road conditions is calculated: M·e+(F×H)·e+(G×h)·e+Fz·(h+s·cosθ)·e=0, where h is the desired stable boundary distance.

7. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The actual support area of ​​the crawler chassis in step 4 is as follows: a thin film pressure sensor is arranged at the wheel axle of the track wheel, the ground wheel is determined according to the signal of the pressure sensor, the line between adjacent ground wheels determines the boundary of the actual support area, and the shortest distance d from the center of gravity of the chassis to the actual support boundary in the horizontal plane is calculated.

8. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The judgment logic of the stability state in step 4 is: If d>h, it is determined to be a stable state; if d=h, it is determined to be a critical stable state; if d<h, it is determined to be an unstable state.

9. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The specific method for determining the overturning direction of the robotic arm system in step 4 is: comparing the distance between the stable boundary and the actual support boundary in the direction of each support edge, The direction of minimum value of S is the overturning direction.

10. The method for determining and controlling the stability of a crawler harvesting robot arm based on a random forest regression model according to claim 1, wherein: The specific method for dynamically adjusting the robot arm's speed in step 5 is to select the robot arm's motion trajectory with the best stability, using the stability criterion as a constraint. When a tipping risk is detected, the tipping moment generated by each joint's motion is calculated to identify the joint generating the maximum tipping moment. Subsequently, the speed and acceleration of that joint are reduced to mitigate the tipping risk.

Images