Obstacle avoidance control method of mobile mechanical arm

Abstract

The invention relates to an obstacle avoidance control method for a mobile mechanical arm. The obstacle avoidance control method comprises the steps of establishing a dynamic priority distribution mechanism, adjusting a weight matrix of a chassis and the mechanical arm in real time based on environmental threats and the state of the mechanical arm, and balancing obstacle avoidance and joint optimization. A repulsive potential field model is improved, a dynamic obstacle track is predicted in combination with a visual Transform, and short-term repulsive force is enhanced; designing space-time constraint model predictive control, and generating a smooth trajectory through quaternion attitude tracking and dynamic constraint; the potential field intensity, the control weight and the priority parameters are globally optimized by adopting a double-delay depth deterministic strategy gradient algorithm. According to the method, the obstacle avoidance success rate can be improved, the robot motion trail is smoothed, and the singular configuration risk is avoided.

Description

Technical Field

[0001] The present invention belongs to the field of robots, and particularly relates to an obstacle avoidance control method for a mobile manipulator. Background Art

[0002] As a core device in scenarios such as industrial automation and warehousing logistics, a mobile manipulator needs to achieve efficient obstacle avoidance and precise operation in a dynamic and complex environment. However, the existing technologies face multiple challenges: firstly, the prediction accuracy of the dynamic obstacle trajectory is insufficient. The traditional potential field method only relies on static environment modeling and is difficult to cope with the sudden interference of high-speed moving obstacles; secondly, the priority allocation of the manipulator obstacle avoidance and the chassis movement is fixed, resulting in the manipulator being easily trapped in a singular configuration or the chassis having a lag in obstacle avoidance response; thirdly, although model predictive control can plan a smooth trajectory, the fixed parameters are difficult to balance the multi-objective conflicts of attitude tracking, energy consumption optimization, and real-time obstacle avoidance; fourthly, traditional reinforcement learning algorithms are prone to control instability due to policy oscillation in a dynamic environment.

[0003] The current research mainly optimizes local problems through a single technology. For example, improving the potential field model to enhance the repulsive force, or introducing an adaptive weight to adjust the manipulator joint priority, but has not achieved global parameter collaborative optimization and efficient coupling with a dynamic environment. In addition, the spatio-temporal feature extraction of dynamic obstacles relies on fixed rules and lacks an accurate modeling of the motion coupling relationship of multiple obstacles. Therefore, there is an urgent need for a method that integrates dynamic priority allocation, high-precision obstacle prediction, spatio-temporal constraint optimization, and adaptive parameter collaboration to improve the obstacle avoidance robustness, motion smoothness, and system energy efficiency of the mobile manipulator in complex scenarios. Summary of the Invention

[0004] The present invention provides a dynamic obstacle avoidance control method for a mobile manipulator, which comprises the following steps: Based on the environmental threat and the manipulator motion state, a dynamic weight matrix for adjusting the chassis movement and the manipulator joint optimization is real-time adjusted through a dynamic priority allocation mechanism; Using a vision Transformer to extract the spatio-temporal features of dynamic obstacles, predict their motion trajectories, and generate an obstacle avoidance guiding force in combination with an improved repulsive potential field model; Adopting spatio-temporal constraint model predictive control, embedding the dynamic weight matrix into the objective function and constraint conditions to generate a smooth motion trajectory; Optimizing the dynamic priority parameters, the repulsive potential field strength, and the weight parameters of the spatio-temporal constraint model predictive control through a double-delayed deep deterministic policy gradient algorithm; Generating joint angular velocity and chassis movement speed commands.

[0005] Specifically, the dynamic priority allocation mechanism is specifically: constructing a dynamic weight matrix , where ζ is a forgetting factor, σ base , σarm are the chassis movement priority and the manipulator joint movement priority, respectively, R base is the chassis motion reachability matrix, R i (q) is the manipulator joint reachability index.

[0006] Specifically, its characteristic is that the priority coefficient adjustment rule is: when the distance to the obstacle is less than the safety threshold, according to σ base ←σ base +0.7×(1 - d / d0) for update; when the joint is close to the singular configuration, according to σ arm ←σ arm +0.5×(1 - ||Ri(q)||) for update, where d is the current distance to the obstacle and d0 is the safety threshold.

[0007] Specifically, the improved repulsive potential field model is: the repulsive potential field function , where η is the potential field strength coefficient, λ is the velocity decay factor, v obs is the predicted relative velocity vector of the obstacle, d is the current distance to the obstacle, and d0 is the safety threshold.

[0008] Specifically, the spatio-temporal feature extraction of the Vision Transformer specifically includes: Segment the environmental point cloud into 16×16×16 grid sub-regions, extract the geometric center coordinates, point cloud density, and normal vector direction features of each sub-region, calculate the motion correlation degree between blocks through the multi-head attention mechanism, and output the predicted velocity vector of the dynamic obstacle.

[0009] Specifically, the objective function of the spatio-temporal constraint model predictive control is , where Q des , Q act are the desired and actual attitude quaternions respectively, τ is the joint torque, W τ is the torque weight matrix τ max is the maximum joint torque, θ i is the joint angle, θ safe is the joint safety threshold, k is the steepness coefficient, u is the control input vector, t0, t f are the start and end times of the prediction horizon respectively, and α, β, γ are the weights of the kinematic term, dynamic term, and ergonomic term respectively.

[0010] Specifically, the constraint conditions of the spatio-temporal constraint model predictive control are .

[0011] Specifically, the double-delayed deep deterministic policy gradient algorithm specifically includes: The state space is defined as the joint state, dynamic weight, reachability index and obstacle speed. The action space is the parameter set. The reward function is designed as Among them, Δt collision is the collision time interval, E consumption is the energy consumption calculated by integrating the motor current, θ tool is the angle between the tool coordinate system and the target direction, σ τ is the standard deviation of joint torque, and A, B, C, and D are the weights of each item.

[0012] Specifically, the double-delayed deep deterministic policy gradient algorithm adopts a dual critic network to suppress over-estimation, the target network achieves smooth parameter transition through soft update φ′←0.995φ′+0.005φ, and delays the policy network update to enhance stability.

[0013] The present invention also provides a mobile robot arm system, which is used to execute any of the control methods described above.

[0014] Beneficial technical effects: The combination of dynamic priority allocation and improved repulsive potential field model enables the system to respond quickly when dynamic obstacles approach, and the obstacle avoidance success rate is improved, which is especially suitable for sudden scenarios with high-speed moving obstacles; the robot arm posture is monitored in real time through the joint accessibility index, and the joint optimization priority is dynamically improved to avoid loss of control due to singular configurations; the spatiotemporal constraint model predictive control generates continuous smooth trajectories through quaternion posture tracking and dynamic weight constraints, and the trajectory jitter amplitude is reduced, which is suitable for high-precision grasping and assembly tasks; the improved dual-delay deep deterministic policy gradient algorithm globally optimizes joint torque and chassis speed, and combined with the energy consumption term in the reward function, the overall energy consumption of the system is reduced, the continuous operation time of the mobile robot arm is extended, and the global optimal balance between obstacle avoidance sensitivity, posture accuracy and energy consumption is achieved; the visual Transformer works in conjunction with the dynamic potential field model to reduce the trajectory prediction error of dynamic obstacles and shorten the system's full-cycle response delay; through dynamic weight constraints and contact force threshold limits, the system can still operate stably in complex scenarios such as narrow spaces and multi-obstacle coupled motion. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Attached Figure 1 is a flow chart of a dynamic obstacle avoidance control method for a mobile mechanical arm according to the present invention; Attached Figure 2 Schematic diagram of dynamic obstacle avoidance of a mobile robotic arm according to the present invention. DETAILED DESCRIPTION

[0016] like Figure 1 As shown, the present invention discloses a dynamic obstacle avoidance control method for a mobile mechanical arm, which comprises the following steps: Allocate dynamic priorities: According to the environmental threats and the motion states of the robotic arm, adjust the priority weights of the chassis and the robotic arm in real time to achieve a balance between obstacle avoidance and motion planning.

[0017] Obtain environmental perception data and joint states through a depth camera and sensors, and calculate the reachability index Ri(q) and the dynamic weight matrix W(t). The dynamic weight matrix W(t) serves as the weight matrix for the subsequent control objective function, and the value of the reachability index Ri(q) is used as the input for the parameter optimization module.

[0018] Among them, the dynamic weight matrix W(t) is:

[0019] Among them, ζ is the forgetting factor, ζ ∈ [0.05, 0.2], which is used to control the attenuation rate of historical data. This parameter can be set and adjusted according to experience or optimized in real time through the parameter optimization module. σ base , σ arm are the chassis movement priority and the robotic arm joint movement priority respectively, which can be calculated by fuzzy logic or set and adjusted according to experience. σ base , σ arm ∈ (0, 1), R base is the chassis motion reachability matrix, which represents the speed limit matrix of the movable directions of the chassis at the current position. For example, it can be diag(v x,max , v y,max , ω max ), R i (q) is the robotic arm joint reachability index, R i (q) = det(JJ T ) + h, h = 10 -6 , which is used to prevent singularity. J is the Jacobian matrix.

[0020] When the detected distance d to the obstacle ahead is less than the safety threshold d0, the chassis movement priority is set to σ base ← σ base + 0.7×(1 - d / d0). The closer the obstacle is, the more the chassis obstacle avoidance action takes precedence over the robotic arm movement. When the joint approaches the singular configuration, that is, when R i (q) tends to 0, the robotic arm joint movement priority is set to σ arm ← σ arm + 0.5×(1 - ||Ri(q)||), so as to avoid the robotic arm getting out of control due to the singular configuration. By allocating dynamic priorities, the success rate of obstacle avoidance can be improved, the motion flexibility of the mobile robotic arm can be monitored in real time, and the risk of singular configurations can be reduced.

[0021] Dynamic Obstacle Prediction and Repulsive Potential Field Planning: By fusing visual perception and motion prediction, a dynamic repulsive force field is generated to guide obstacle avoidance, thereby improving the prediction accuracy of dynamic obstacle trajectories.

[0022] Among them, the improved potential field function U rep is:

[0023] Among them, η is the potential field strength coefficient, the initial value of η can be 2.5, and it can be adjusted through a parameter optimization module or empirical setting; λ is the velocity decay factor, and λ can be 0.1; v obs is the predicted vector of the relative velocity of the obstacle, which is predicted and output by the Vision Transformer. d is the distance from the current position to the obstacle, and d0 is the safety threshold. When the obstacle is stationary, that is, v obs = 0, it is transformed into a traditional potential field; when the obstacle approaches at high speed, the short-term repulsive force is enhanced through the exponential term.

[0024] Among them, the specific prediction of the Vision Transformer is as follows: (1) Input the environmental point cloud data. The point cloud can be divided into sub-regions of 16×16×16 according to the spatial grid, and the geometric center coordinates, point cloud density, average color, and normal vector direction features of each sub-region are extracted.

[0025] (2) The encoding layer calculates the motion correlation degree of the obstacle through the multi-head attention mechanism:

[0026] Among them, Q, K, and V represent the query, key, and value matrices respectively. The query matrix Q focuses on the motion correlation requirements of the current block. If a moving obstacle is detected in the block, Q will strengthen its temporal dimension features. The key matrix K describes the spatio-temporal features of other blocks and calculates the correlation weights between other blocks and the current block. The value matrix V is the actual transmitted feature value, such as the velocity change amount, d k is the feature dimension, and the softmax function is used to convert the attention scores into a probability distribution, , and a high probability value indicates that the block has a significant impact on the current motion prediction. For example, an obstacle approaching rapidly, and irrelevant noise is filtered through a low probability value, such as a distant static object.

[0027] (3) Output the predicted motion vector of the obstacle.

[0028] Through the calculation of the obstacle motion correlation degree, the motion trend of the obstacle can be identified, such as lateral movement or accelerating approach, distinguish static obstacles from dynamic obstacles, and determine the motion coupling relationship between multiple obstacles.

[0029] Spatio-temporal Constraint Model Predictive Control: Under the premise of meeting the set constraints, continuous and smooth motion trajectory planning is achieved through model predictive control. The objective function of model predictive control is:

[0030] where Q des , Q act are the desired and actual attitude quaternions respectively, τ is the joint torque, W τ is the torque weight matrix, W τ = diag(1 / τ max 2 ), τ max is the maximum joint torque, θ i is the joint angle, θ safe is the joint safety threshold, k is the steepness coefficient, u is the control input vector, which includes joint torque and chassis speed, t0, t f are the start and end times of the prediction horizon respectively, α, β, γ are the weights of the kinematic term, dynamic term, and ergonomic term respectively, which are used to adjust the balance between attitude tracking accuracy and obstacle avoidance sensitivity, suppress sudden changes in joint torque, and avoid joints exceeding the safe angle range, and can be set to 0.6, 0.3, 0.1 respectively.

[0031] The constraint conditions of model predictive control are:

[0032] The constraint conditions include quaternion differential constraint, maximum torque constraint, contact force constraint, and dynamic priority constraint. By embedding the dynamic weight matrix W(t) into the constraint conditions, the motion bandwidth of high-priority tasks is guaranteed, and the Euler angle singularity problem is avoided through the quaternion differential constraint.

[0033] Parameter Co-optimization: Dynamically adjust the parameters of each module through the improved double-delay deep deterministic policy gradient algorithm to achieve optimal global performance. Specifically: (1) State-Action Space Definition The state s includes joint state, dynamic weight, reachability index, and obstacle speed prediction,

[0034] The action a = [ζ, η, α, β, γ], that is, simultaneously optimize the dynamic priority, potential field strength, and model predictive weight coefficient.

[0035] (2) Reward Function Design

[0036] where Δt collision is the collision time interval, set to ∞ or a fixed value such as 10s when there is no collision, E consumptionThe energy consumption for calculating the integral of the motor current, θ tool The angle between the tool coordinate system and the target direction, σ τ The standard deviation of the joint torque, which is used to reflect the load balance. The weights A, B, C, and D of each item can be 0.4, 0.2, 0.3, and 0.1 respectively.

[0037] (3) Network update strategy, which is specifically as follows: 3.1 Double Critic to suppress overestimation: , where Q θi′ is the target Critic network, which provides a stable Q-value estimate to prevent policy oscillation caused by overfitting of a single network. π(s′) is the target Actor network, which generates sub-optimal action exploration. Delayed update of the policy network improves stability. j is the exploration noise, which enhances environmental exploration at the beginning of training to avoid falling into local optima.

[0038] 3.2 Delayed policy update: Perform 1 Actor update after every 2 Critic updates; 3.3 Soft update the target network as φ′←0.995φ′+0.005φ.

[0039] Through the improved double-delayed deep deterministic policy gradient algorithm, the output parameter compensation amount is output. By adjusting the model predictive control weight parameters α, β, γ, the dynamic priority parameter ζ, and the potential field strength η, the final command is affected to achieve dynamic response adjustment. For example, when an emergency obstacle avoidance is detected, the algorithm increases η, resulting in an enhanced potential field repulsive force, and the v base generated by the model predictive control increases accordingly.

[0040] Control command synthesis: Generate joint angular velocity and chassis movement speed commands.

[0041] As Figure 2 shown, a series of obstacles (triangles in the figure) are set on the planned motion trajectory (dashed line in the figure) of the mobile manipulator, and the actual motion trajectory (solid line in the figure) of the mobile manipulator can be observed. It can be seen that the control method of the present invention has a good control effect. The present invention directly embeds the dynamic priority weight into the objective function of the model predictive control to achieve a decision-making - planning closed loop; through obstacle speed prediction, the potential field and motion constraints are corrected in real time; through the improved double-delayed deep deterministic policy gradient algorithm, global optimization is achieved, which runs through all adjustable parameters and breaks through the local optimum limit.

[0042] The present invention also proposes a clamping control method for a mobile manipulator, which includes the following steps: Contact force perception and dynamic coordination, which is specifically as follows: Real-time collect the contact force F contact =[f x ,fy , f z , τ x , τ y , τ z T , the chassis linear velocity v base and the manipulator joint torque τ arm ; Dynamically coordinate and calculate the motion coupling factor: ω c = k1 * (v base × τ arm ) + k2 * ||ΔF||, where k1 is the motion coupling coefficient, the initial value can be 0.8, k2 is the force field sensitivity coefficient, the initial value can be 1.2, and ΔF is the contact force gradient field, specifically , where R b w is the rotation matrix from the chassis to the world coordinate system, is the tensor product; Output the motion coupling factor ω c and the contact force gradient magnitude ||ΔF||, where the motion coupling factor ω c is used to limit the chassis speed, and the contact force gradient magnitude ||ΔF|| is used for sliding mode controller gain adjustment.

[0043] Sliding mode control: Real-time collect the desired trajectory error E, , that is, the joint angle and velocity error, and the motion coupling factor ω c and the contact force gradient magnitude ||ΔF||.

[0044] The time-varying sliding mode surface is designed as: , where γ is the dynamic integral gain. When the error is large, γ can increase the integral gain to accelerate the system to converge to the sliding mode surface. When approaching the steady state, the gain is reduced to avoid overshoot, balancing the response speed and stability. By dynamically adjusting the integral action, it can effectively handle uncertainties such as load mutations and external disturbances, reducing the risk of oscillation or instability that may be caused by traditional fixed-gain integration. The gain matrix K(t) satisfies: , where R(x) is the environmental curvature tensor, a tensor describing the geometric characteristics of the robot's motion environment, determined by the curvature of the chassis motion trajectory and the manipulator configuration. In an unstructured environment (such as an uneven ground), the curvature tensor reflects the degree of bending of the motion path, , where κ i is the principal curvature of the trajectory. By affecting the adjustment of the sliding mode gain matrix K(t), when the path curvature is larger, the control gain is reduced to avoid overshoot. The energy constraint matrix Q(x) is determined by the conservation of kinetic energy, a symmetric positive definite matrix derived from the robot's kinetic energy conservation equation, used to constrain the change range of the sliding mode gain K(t) to prevent instability caused by sudden energy changes​ The control law is as follows:

[0045] where the pseudo-inverse J of the Jacobian matrix c + is the pseudo-inverse of the Jacobian matrix J of the end effector of the robotic arm c and is calculated as =J c T (J c J c T ). The adaptive gain ρ = ρ0 + α||ΔF||, where α is the optimization parameter of reinforcement learning. The dynamic coupling constraint matrix Λi = diag(ω -1 *λ c ), M is the inertia matrix, C is the Coriolis force estimation matrix, and G is the gravity compensation term; δ is the thickness of the sliding mode boundary layer. The smaller the value, the higher the accuracy of approaching the ideal sliding mode surface, but it may cause chattering. The larger the value, the smoother the control signal, but it will introduce a steady-state error. Therefore, δ can be set as δ = 0.1||ΔF|| + 0.05*ω i to balance the accuracy and smoothness, and it can be dynamically adjusted through parameter optimization later; σ is the sliding mode surface. c . The parameter optimization of the double-delay deep deterministic policy gradient is as follows:

[0046] 1. State-action space: Collect q, joint angular velocity, chassis attitude angle θ , chassis speed v base , and F base , and set the state as contact , and the action a = [Δk1, Δk2, Δρ, Δδ], that is, the parameter increment t . 2. Reward function design: where σ

[0047] is the force volatility, σ F = std(F F ), that is, the standard deviation of the contact force. θ contact align is the alignment angle between the tool and the target, specifically the angle between the end effector coordinate system and the target direction. The energy consumption , Vibration amp amp is the vibration amplitude, which is the maximum vibration amplitude of the end of the robotic arm in the frequency domain analysis, usually taking the peak value in the 0 - 100 Hz frequency band. The weights A, B, C, and D of each item can be 0.4, 0.2, 0.3, and 0.1 respectively.

[0048] 3. Network update strategy: (1)Double Critic to Suppress Overestimation: Q = min(Q θ1′ , Q θ2′ ), where Q θ1′ , Q θ2′ are the target Critic networks, neural networks for calculating the target Q-values. The minimum value is taken through the double Critic network to suppress the overestimation of the Q-values.

[0049] (2)Delayed Actor Update: φ′ ← 0.995φ′ + 0.005φ; (3)Perform 1 Actor update after every 2 Critic updates.

[0050] Output joint torque τ and chassis speed command v base ′ = v base ·sat(ωc), speed limit through the saturation function sat. sat can be: .

[0051] Real-time constraint on the chassis speed through ω c , and at the same time adjust the sliding mode control gain by ||ΔF|| to form a bidirectional coupling. In the sliding mode control, combine the double-delay deep deterministic policy to dynamically adjust δ, suppress chattering while maintaining robustness. The double-delay deep deterministic policy realizes multi-objective optimization through the composite index of the reward function to ensure the overall optimality of the system.

[0052] This scheme deeply integrates dynamic coordination, sliding mode control and reinforcement learning, realizes a closed-loop from force perception to motion control and then to parameter optimization, and significantly improves the stability and accuracy of the mobile manipulator in the gripping operation.

[0053] In several embodiments provided by the present invention, it should be understood that the disclosed method can be implemented in other ways. For example, the above-described invention embodiments are merely illustrative. For example, the division of method modules and steps is only a logical function division, and there may be other division methods in actual implementation. The modules and steps described as separate components or separate steps may or may not be physically separated. The components shown as modules may or may not be physical modules, and may be located in one place or distributed to multiple network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0054] In addition, in each embodiment of the present invention, the functional modules and steps can be integrated in one processing module or method, or each module and step can exist physically alone, or two or more modules and steps can be integrated in one module or method. The above integrated modules can be implemented in the form of hardware or in the form of hardware plus software functional modules.

[0055] It is obvious to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the basic characteristics of the present invention.

[0056] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic obstacle avoidance control method for a mobile manipulator, characterized in that, It includes the following steps: Based on the environmental threats and the motion state of the robotic arm, the dynamic weight matrices of the chassis movement and the robotic arm joints are adjusted in real time through a dynamic priority allocation mechanism; Utilize a Vision Transformer to extract the spatio-temporal features of dynamic obstacles, predict their motion trajectories, and generate an obstacle avoidance guiding force in combination with an improved repulsive potential field model; Adopt spatio-temporal constraint model predictive control, embed the dynamic weight matrices into the objective function and constraint conditions to generate a smooth motion trajectory; Optimize the dynamic priority parameters, the repulsive potential field strength, and the weight parameters of the spatio-temporal constraint model predictive control through the Twin Delayed Deep Deterministic Policy Gradient algorithm; Generate joint angular velocity and chassis movement speed commands.

2. The method according to claim 1, wherein The dynamic priority allocation mechanism is specifically as follows: construct a dynamic weight matrix , where ζ is the forgetting factor, and σ base , σ arm are the chassis movement priority and the manipulator joint movement priority respectively, R base is the chassis motion reachability matrix, and R i (q) is the manipulator joint reachability index.

3. The method according to claim 2, characterized in that, The priority coefficient adjustment rule is: when the distance to the obstacle is less than the safety threshold, update it according to σ base ←σ base +0.7×(1 - d / d0); when the joint is approaching the singular configuration, update it according to σ arm ←σ arm +0.5×(1 - ||Ri(q)||), where d is the current distance to the obstacle and d0 is the safety threshold.

4. The method according to claim 1, wherein The improved repulsive potential field model is: the repulsive potential field function , where η is the potential field strength coefficient, λ is the velocity decay factor, v obs is the predicted vector of the relative velocity of the obstacle, d is the current distance from the obstacle, and d0 is the safety threshold.

5. The method according to claim 4, characterized in that, The specific spatio-temporal feature extraction of the Vision Transformer includes: Divide the environmental point cloud into 16×16×16 grid sub-regions, extract the geometric center coordinates, point cloud density, and normal vector direction features of each sub-region, calculate the motion correlation degree between blocks through the multi-head attention mechanism, and output the predicted velocity vector of the dynamic obstacle.

6. The method according to claim 1, characterized in that, The objective function of the spatio-temporal constraint model predictive control is , where Q des , Q act are the desired and actual attitude quaternions respectively, τ is the joint torque, W τ is the torque weight matrix, θ i is the joint angle, θ safe is the joint safety threshold, k is the steepness coefficient, u is the control input vector, t0, t f are the start and end times of the prediction horizon respectively, and α, β, γ are the weights of the kinematic term, dynamic term, and ergonomic term respectively.

7. The method according to claim 6, wherein The constraint conditions of the spatio-temporal constraint model predictive control are .

8. The method according to claim 1, characterized in that, The Twin Delayed Deep Deterministic Policy Gradient algorithm specifically includes: defining the state space to include joint states, dynamic weights, reachability indicators, and obstacle speeds, the action space as a parameter set, and the reward function is designed as Among them, Δt collision is the collision time interval, E consumption is the energy consumption calculated by integrating the motor current, θ tool is the angle between the tool coordinate system and the target direction, σ τ is the standard deviation of the joint torque, and A, B, C, and D are the weights of each sub-item.

9. The method according to claim 8, characterized in that, The Twin Delayed Deep Deterministic Policy Gradient algorithm uses a Twin Critic network to suppress overestimation. The target network realizes smooth parameter transition through soft update φ′←0.995φ′+0.005φ, and delays the update of the policy network to enhance stability.

10. A mobile robotic arm system, characterized in that, For implementing the control method according to any one of claims 1-9.

Images