A robot manipulator collision avoidance planning method, system, medium and device

Abstract

The application relates to the technical field of human-computer interaction, and provides a robot mechanical arm collision avoidance planning method, system, medium and equipment, which comprises the following steps: observing human body posture features, processing the posture features through a feature coding layer and an encoder to obtain the posture features, processing the posture features through a decoder and a feature decoding layer, and then performing inverse discrete cosine transformation to obtain human body posture features at a future time; observing human body trajectory features, processing the trajectory features through the feature coding layer and the encoder to obtain the trajectory features, performing feature fusion on the trajectory features and the posture features through a multi-head attention layer, processing the fused trajectory features through the decoder and the feature decoding layer, and then performing inverse discrete cosine transformation to obtain human body trajectory features at the future time; calculating human body models from the human body posture features at the future time, and point cloudizing the human body models to obtain predicted dynamic human body point clouds; and acquiring object point clouds, taking the dynamic human body point clouds and the object point clouds as obstacles to perform collision avoidance trajectory planning.

Description

Technical Field

[0001] This invention belongs to the field of human-computer interaction technology, and in particular relates to a collision avoidance planning method, system, medium and equipment for a robotic arm. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Human pose prediction has wide applications in human-computer interaction, motion capture, virtual reality, and other fields. The goal of human pose prediction is to predict the movement trend of a human body over a given historical motion sequence. With the development of deep learning technology, neural network-based human pose prediction models have made significant progress. However, challenges remain in terms of real-time performance and accuracy. Furthermore, using geometric envelope models to model the human body struggles to meet the required level of detail.

[0004] Avoiding collisions between robotic arms and human users is a crucial objective in human-computer interaction. There are numerous collision avoidance planning methods, such as sampling-based methods, optimization-based methods, and learning-based methods.

[0005] Sample-based model predictive control (MPPI) has gradually become the mainstream method for collision avoidance path planning due to its advantages such as being able to handle nonlinear systems and high-dimensional state spaces, and being able to adjust online in real time in dynamic environments. However, traditional sample-based model predictive control algorithms are not optimized for collision avoidance with humans, and in dynamic environments, it is quite difficult for robotic arms to maintain real-time performance when facing dynamic obstacles. Summary of the Invention

[0006] To address the technical problems mentioned above, this invention provides a robot arm collision avoidance planning method, system, medium, and device. Based on predicted human trajectory features and human posture features, the human posture prediction network is jointly optimized to improve the effect of human posture prediction. Furthermore, the predicted human posture sequence is calculated into a human body model, which is then cloudified and added to the robot arm collision avoidance process. This ensures that the cost calculation of the robot arm's sampling trajectory is related not only to the current human and object point clouds but also to the predicted future human point clouds, making the robot arm's collision avoidance trajectory more forward-looking.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] The first aspect of the present invention provides a collision avoidance planning method for a robotic arm, comprising:

[0009] Acquire the observed human posture characteristics and observed human trajectory characteristics at several historical moments;

[0010] For observing human posture features, the feature encoding layer and encoder are used to obtain posture features. After processing the posture features through the decoder and feature decoding layer, an inverse discrete cosine transform is performed to obtain the human posture features at future time. For observing human trajectory features, the feature encoding layer and encoder are used to obtain trajectory features. The trajectory features and posture features are fused through a multi-head attention layer. The fused trajectory features are then processed through the decoder and feature decoding layer, and an inverse discrete cosine transform is performed to obtain the human trajectory features at future time. The feature encoding layer, encoder, decoder, and feature decoding layer are optimized by combining the human trajectory features at future time and the human posture features.

[0011] The human body posture features at future moments are calculated into a human body model, and the human body model is converted into a point cloud to obtain the predicted dynamic human body point cloud.

[0012] The object point cloud is acquired and, together with the dynamic human body point cloud, is used as an obstacle to plan a collision avoidance trajectory.

[0013] Furthermore, the feature fusion step of the multi-head attention layer includes:

[0014] Based on trajectory features and posture features Calculate query Q Key K Sum V ,in, , and These are the weights of three linear layers;

[0015] Calculate association features based on query, key, and value. V, where, Used for normalization processing;

[0016] The associated features and the attitude features are multiplied by a matrix and then concatenated to obtain the fused trajectory features.

[0017] Furthermore, the calculation of the human body model adopts a linear model used for human body modeling.

[0018] Furthermore, the collision avoidance trajectory planning employs sampling-based model predictive control.

[0019] Furthermore, the feature encoding layer comprises a discrete cosine transform, a fully connected layer, and a transpose layer connected in sequence.

[0020] Furthermore, the encoder employs a multilayer perceptron comprising multiple linear layers and normalization layers.

[0021] Furthermore, the feature decoding layer has the same structure as the feature encoding layer; the decoder has the same structure as the encoder.

[0022] A second aspect of the present invention provides a collision avoidance planning system for a robotic arm, comprising:

[0023] The data acquisition module is configured to acquire the observed human posture features and observed human trajectory features at several historical moments.

[0024] The feature extraction module is configured as follows: For observed human posture features, the module processes them through a feature encoding layer and an encoder to obtain posture features. After processing the posture features through a decoder and a feature decoding layer, an inverse discrete cosine transform is performed to obtain the human posture features at future time. For observed human trajectory features, the module processes them through a feature encoding layer and an encoder to obtain trajectory features. The trajectory features and posture features are fused through a multi-head attention layer. The fused trajectory features are then processed through a decoder and a feature decoding layer, and an inverse discrete cosine transform is performed to obtain the human trajectory features at future time. The module further optimizes the feature encoding layer, encoder, decoder, and feature decoding layer by combining the human trajectory features at future time and the human posture features.

[0025] The point cloudification module is configured to: calculate the human body posture features at future moments into a human body model, and then convert the human body model into a point cloud to obtain the predicted dynamic human body point cloud.

[0026] The trajectory planning module is configured to: acquire object point clouds, and together with the dynamic human body point cloud, use them as obstacles to perform collision avoidance trajectory planning.

[0027] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the collision avoidance planning method for a robotic arm as described above.

[0028] A fourth aspect of the present invention provides a computer device including a computer-readable storage medium, a processor, and a computer program stored on the computer-readable storage medium and executable on the processor, wherein the processor executes the program to implement the steps in the collision avoidance planning method for a robotic arm as described above.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This invention calculates the predicted human posture sequence into a human body model using the SMPL model, and then converts the human body model into a point cloud. This is then incorporated into the collision avoidance of the robotic arm based on MPPI, so that the cost calculation of the robotic arm's sampling trajectory is not only related to the current human body and object point cloud, but also to the predicted future human body point cloud. This allows the robotic arm to select more appropriate control inputs and also has a certain degree of foresight in the collision avoidance process.

[0031] This invention considers that human posture is greatly affected by the position (trajectory) of the hip, predicts human trajectory features, and fuses them with posture features through a multi-head attention mechanism. Based on the predicted human trajectory features and human posture features, the human posture prediction network is optimized together, which can better improve the effect of human posture prediction. Attached Figure Description

[0032] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0033] Figure 1 This is a flowchart of a collision avoidance planning method for a robotic arm according to Embodiment 1 of the present invention;

[0034] Figure 2 This is a schematic diagram of the human posture prediction network according to Embodiment 1 of the present invention;

[0035] Figure 3 This is a schematic diagram of the multi-head attention layer structure of Embodiment 1 of the present invention;

[0036] Figure 4 This is a schematic diagram of using SMPL to cloudify human body points according to Embodiment 1 of the present invention;

[0037] Figure 5 This is a schematic diagram of the first control input with the lowest cost in the application of MPPI to the robotic arm in Embodiment 1 of the present invention;

[0038] Figure 6 This is a schematic diagram of the sampling-based model predictive control planning process according to Embodiment 1 of the present invention;

[0039] Figure 7 This is a schematic diagram of the structure of a computer device according to Embodiment 4 of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0041] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] Example 1

[0043] This embodiment provides a collision avoidance planning method for a robotic arm.

[0044] This embodiment provides a collision avoidance planning method for a robotic arm. Based on MPPI collision avoidance planning, a human pose point cloud prediction network is added. The predicted future human pose sequence is calculated into a human model using the SMPL (Linear Model for Human Modeling) model, and the human model is converted into a point cloud. The point cloud in the environment and the predicted point cloud are used together for MPPI algorithm collision avoidance planning.

[0045] This embodiment provides a collision avoidance planning method for a robotic arm, such as... Figure 1 As shown, it includes the following steps:

[0046] Step 1: Predict the future posture and trajectory of the human body using a human posture prediction network.

[0047] In step 1 of this embodiment, a multilayer perceptron human pose point cloud prediction subnetwork and a human trajectory prediction subnetwork that combine a multi-head attention layer are provided.

[0048] like Figure 2 As shown, the two prediction sub-networks have the same structure, both including a feature encoding layer, an encoder, a decoder, and a feature decoding layer, and the two sub-networks are connected through a multi-head attention layer.

[0049] Step 101: Obtain the user's observed pose using a depth camera, i.e., the past... Observations of human posture characteristics at historical moments and observation of human trajectory characteristics ;

[0050] Step 102: Observe the characteristics of human posture. Perform Discrete Cosine Transform (DCT) to encode the action time series, resulting in... .

[0051] Step 103, A single-layer linear layer (fully connected layer) processes the features in the action space, and a transpose layer is used to transpose the features to obtain the output. ;in, and These are the learnable parameters for a single linear layer (fully connected layer);

[0052] Step 104: Add features Input the attitude encoder, which contains m perceptron modules.

[0053] Features The input is a multilayer perceptron module consisting of multiple linear layers (fully connected layers) and normalization layers, which encodes the pose features in time.

[0054] A residual network structure is used between adjacent perceptron layers. The residual network only learns the residual values ​​of the features before and after the transition, using the formula... ,in, Indicates the normalization layer. and Let be the learnable parameters of the i-th fully connected layer. The output of the i-th fully connected layer and the output features of the (i-1)-th perceptron module are shown. After passing through the i-th layer perceptron module, the original features are merged across layers. The output features of the i-th layer perceptron module are obtained. .

[0055] The output of the attitude encoder is the attitude feature. m represents the number of multilayer perceptrons in the attitude encoder. This is the output feature of the (m-1)th layer perceptron module.

[0056] Step 105: Assign pose features The input is an attitude decoder with the same structure as the attitude encoder, which decodes the attitude features in the time dimension; the decoded attitude features are then transposed to... ;Will A single linear layer is input to extract spatial features.

[0057] Step 106: Use inverse discrete cosine transform to transform the output back to the original attitude representation, obtaining the future 1 Human posture characteristics at any moment .

[0058] Step 107: Observe the characteristics of human body trajectory The input discrete cosine transform is used to encode the action time series, and the encoded result is... .

[0059] Step 108, A single linear layer processes the features in the action space, and a transpose layer transposes the features to obtain the output. .

[0060] Step 109: Add features Input a multilayer perceptron module with the same structure as described above to encode the trajectory features in time.

[0061] Step 110: Obtain the pose features Trajectory Features Perform feature fusion and input it into a multi-head attention layer.

[0062] like Figure 3 As shown, the inputs to the multi-head attention layer are query Q, key K, and value V, respectively. , , ;in, , and These are the weights of the three linear layers; expressed by the formula. V, Used for normalization to obtain the correlation features between pose features and trajectory features. ; associate features With posture features Performing matrix multiplication followed by matrix concatenation yields new trajectory features. .

[0063] Step 111: Merge the trajectory features The trajectory decoder, with the same structure as the trajectory encoder, is used to decode the trajectory features in the time dimension; the decoded trajectory features are then transposed to... ;Will A single linear input layer extracts spatial features; the inverse discrete cosine transform is used to transform the output back to the original trajectory representation, yielding the future 1 Human trajectory characteristics .

[0064] Since human posture is greatly influenced by the position (trajectory) of the hip, predicting human trajectory features and fusing them with posture features through a multi-head attention mechanism, and then optimizing the human posture prediction network based on the predicted human trajectory features and human posture features, can better improve the performance of human posture prediction.

[0065] Step 2: Analyze the predicted human posture features Calculate the human body model and convert the human body model into a point cloud to obtain the predicted dynamic human body point cloud.

[0066] In step 2 of this embodiment, the following is used: Figure 4 The SMPL shown converts the predicted human pose features into a human model and then cloudifies the human model into points.

[0067] Step 3: Construct a sampling-based model predictive control (MPPI) to achieve collision avoidance path planning in an environment with obstacle point clouds.

[0068] In step 3 of this embodiment, a collision avoidance planning method for a robotic arm based on MPPI is provided, such as... Figure 6 As shown, it includes:

[0069] (1) Select the state of the model predictive control as , where q, , These are the robot arm's angle, angular velocity, and angular acceleration, respectively.

[0070] (2) The motion model of the robotic arm is the optimal control of a discrete stochastic dynamic system. ,in, The state at time t In the input Switching to under control state of time The probability of.

[0071] (3) Sample from the current input distribution Batch, length is control sequence For each batch, the sampled control sequence is input into the approximate dynamic model of the robotic arm to obtain the trajectory and the corresponding state sequence. and cost sequence The batch size of the state sequence and its corresponding cost sequence is N, and the length is H.

[0072] In this embodiment, the cost of each batch is calculated in parallel. ,in, It is length Other terminal costs, It is a discount factor used to favor immediate costs; the sampling batch size is N. The cost obtained by inputting a batch control sequence. For the h-th control in the control sequence of length H at time t, similarly... Let H be the h-th state in the current batch of state sequence of length H at time t. The parameters for the terminal cost are reasonable parameters that are specified by the user.

[0073] (4) such as Figure 5 As shown, from In the batch control sequence, the first input of the control sequence with the lowest cost is selected as the input of the current state, i.e., the optimal input.

[0074] It should be noted that what is obtained is a control sequence, which is a control sequence with a field of view length of H. The first control sequence is selected as the input.

[0075] (5) Update the distribution for the next sampling using the control sequence of the current sample, updating the variance and mean of the distribution for the next sampling. Mean update equation: Covariance update equation: , and It is the update step size and weight. The calculation equation is as follows: ;in, Let be the mean of the sampling distribution at the h-th position of the entire sequence at time t. Let be the covariance of the sampling distribution at the h-th position of the entire sequence at time t; weights In the calculation formula, This represents the control at the h-th position of the i-th batch sequence. This represents the state of the h-th position in the i-th batch sequence. The parameter represents the terminal cost.

[0076] (6) Sample the control sequence from the distribution again, calculate the control input, and update the input distribution. Repeat this process until the target point is reached.

[0077] Step 4: The predicted dynamic human point cloud and the real-time object point cloud of the environment are used as obstacles and added together to the collision avoidance planning based on sampling model prediction control.

[0078] Obstacles affect the cost of collisions, which in turn affects the total cost. The cost affects the weights in the update equation, which in turn affects the sampling distribution.

[0079] In step 4 of this embodiment, the human pose prediction point cloud is integrated into the sampling-based model prediction control to provide a collision avoidance planning method based on MPPI and combined with the human pose prediction point cloud.

[0080] During the motion planning process of the robotic arm based on sampled model predictive control, the point cloud of real-time objects and human bodies is treated as obstacles.

[0081] The robotic arm samples the trajectory and calculates the collision avoidance cost of obstacles to obtain the control input with the lowest cost, which then controls the next movement of the robotic arm.

[0082] like Figure 1 As shown, the dynamic human point cloud predicted by the prediction network is combined with the real-time point cloud and incorporated into the motion planning of the robotic arm. Each step of the robotic arm's sampling trajectory not only avoids the real-time human point cloud, but also takes into account the predicted dynamic human point cloud, making the robotic arm's collision avoidance trajectory more forward-looking.

[0083] Collision avoidance planning based on sampling model predictive control employs basic mesh collision detection.

[0084] Step 5: By using sampling-based model predictive control to plan a more forward-looking collision avoidance trajectory, the robotic arm completes the movement.

[0085] This embodiment provides a robot arm collision avoidance planning method. The predicted human posture sequence is calculated into a human body model through the SMPL model, and the human body model is converted into a point cloud and added to the MPPI-based robot arm collision avoidance. This makes the cost calculation of the robot arm's sampling trajectory not only related to the current human body and object point cloud, but also related to the predicted future human body point cloud, allowing the robot arm to select more appropriate control inputs, and also has a certain degree of foresight in the collision avoidance process.

[0086] Example 2

[0087] This embodiment provides a collision avoidance planning system for a robotic arm, which specifically includes:

[0088] The data acquisition module is configured to acquire the observed human posture features and observed human trajectory features at several historical moments.

[0089] The feature extraction module is configured as follows: For observed human posture features, the module processes them through a feature encoding layer and an encoder to obtain posture features. After processing the posture features through a decoder and a feature decoding layer, an inverse discrete cosine transform is performed to obtain the human posture features at future time. For observed human trajectory features, the module processes them through a feature encoding layer and an encoder to obtain trajectory features. The trajectory features and posture features are fused through a multi-head attention layer. The fused trajectory features are then processed through a decoder and a feature decoding layer, and an inverse discrete cosine transform is performed to obtain the human trajectory features at future time. The module further optimizes the feature encoding layer, encoder, decoder, and feature decoding layer by combining the human trajectory features at future time and the human posture features.

[0090] The point cloudification module is configured to: calculate the human body posture features at future moments into a human body model, and then convert the human body model into a point cloud to obtain the predicted dynamic human body point cloud.

[0091] The trajectory planning module is configured to: acquire object point clouds, and together with the dynamic human body point cloud, use them as obstacles to perform collision avoidance trajectory planning.

[0092] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0093] Example 3

[0094] This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the collision avoidance planning method for a robotic arm as described in Embodiment 1 above.

[0095] Example 4

[0096] This embodiment provides a computer device, such as... Figure 7 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps in the robot arm collision avoidance planning method described in Embodiment 1 above.

[0097] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A robot manipulator collision avoidance planning method, characterized by, The method comprises the following steps: obtaining observation human body posture features and observation human body trajectory features at a plurality of historical time points; for the observation human body posture features, obtaining posture features through feature encoding layers and an encoder, and obtaining future time human body posture features through inverse discrete cosine transformation after the posture features are processed through a decoder and feature decoding layers; for the observation human body trajectory features, obtaining trajectory features through the feature encoding layers and the encoder, and obtaining future time human body trajectory features through inverse discrete cosine transformation after the trajectory features and the posture features are fused through a multi-head attention layer and the fused trajectory features are processed through the decoder and the feature decoding layers; and optimizing the feature encoding layers, the encoder, the decoder and the feature decoding layers in combination with the future time human body trajectory features and the human body posture features; the step of fusing features through the multi-head attention layer comprises: Based on trajectory features and pose features , compute a query Q , a key K and a value V , wherein based on the query, the key and the value, compute an association feature V, wherein is used to perform normalization processing; calculating the future time human body posture features into a human body model, and point cloudizing the human body model to obtain predicted dynamic human body point clouds; obtaining object point clouds as obstacles together with the dynamic human body point clouds to plan an avoidance trajectory; wherein: (1) The state of the selected model predictive control is wherein q, , are the angle, angular velocity, and angular acceleration of the robot arm, respectively. (2) Optimal control of a mechanical arm modeled as a discrete stochastic dynamical system where is the state at time t is the state at time t+1 is the control at time t is the probability of transitioning to the state at time t+1 under the control (3) sampling from the current input distribution batch, length of control sequence For each batch, input the sampled control sequence into the robot approximate dynamics model to obtain the trajectory and the corresponding state sequence and cost sequence ; wherein the batch of the state sequence and the corresponding cost sequence is N and the length is H; Parallel computing the cost of each batch where, is the length of the batch the terminal cost except for the batch cost, is the discount factor, which is used to bias the current cost; the sampling batch is N, is the cost obtained by inputting a batch control sequence, is the hth control of the control sequence of the current batch with a length of H at time t, and is the hth state of the state sequence of the current batch with a length of H at time t, is the parameter of the terminal cost, which is a reasonable parameter artificially specified; (4) from selecting as the input of the current state the first input of the batch control sequence with the lowest cost, i.e. the optimal input; a control sequence is obtained, which is a control sequence with a field of view length H, and the first control sequence is selected as an input; (5) using the current sampling control sequence to update the distribution of the next sampling, updating the variance and mean of the distribution, for the next sampling; the mean updating equation: , the covariance updating equation: , and is the updating step, the weight the calculation equation of ; wherein, is the mean of the sampling distribution of the hth position of the entire sequence at time t, is the covariance of the sampling distribution of the hth position of the entire sequence at time t; the weight in the calculation formula of represents the control of the hth position of the i th batch sequence, represents the state of the hth position of the i th batch sequence, is the parameter of the terminal cost; (6) the control input is calculated again by sampling the control sequence from the distribution, and the input distribution is updated, and the cycle is repeated until the target point is reached. , and are weights of the three linear layers.

2. The robot arm collision avoidance planning method of claim 1, wherein: the fused trajectory features are obtained by matrix multiplication of the associated features and the posture features and then matrix splicing.

3. The robot arm collision avoidance planning method of claim 1, wherein, The human body model is calculated using a linear model for human body modeling.

4. The robot arm collision avoidance planning method of claim 1, wherein, The avoidance trajectory planning adopts a model predictive control based on sampling.

5. The robot arm collision avoidance planning method of claim 1, wherein, The feature encoding layer comprises a discrete cosine transformation, a fully connected layer and a transpose layer connected in sequence.

6. The robot arm collision avoidance planning method of claim 1, wherein, The encoder adopts a multi-layer perceptron comprising multiple linear layers and normalization layers.

7. The robot arm collision avoidance planning method of claim 1, wherein, The feature decoding layer has the same structure as the feature encoding layer, and the decoder has the same structure as the encoder.

8. A robot manipulator collision avoidance planning system, characterized by, The method comprises the following steps: a data acquisition module configured to obtain observation human body posture features and observation human body trajectory features at a plurality of historical time points; a feature extraction module configured to, for the observation human body posture features, obtain posture features through feature encoding layers and an encoder, and obtain future time human body posture features through inverse discrete cosine transformation after the posture features are processed through a decoder and feature decoding layers; for the observation human body trajectory features, obtaining trajectory features through the feature encoding layers and the encoder, and obtaining future time human body trajectory features through inverse discrete cosine transformation after the trajectory features and the posture features are fused through a multi-head attention layer and the fused trajectory features are processed through the decoder and the feature decoding layers; and optimizing the feature encoding layers, the encoder, the decoder and the feature decoding layers in combination with the future time human body trajectory features and the human body posture features; the step of fusing features through the multi-head attention layer comprises: Based on trajectory features and pose features , compute query Q , key K and value V wherein based on the query, the key and the value, compute association features V, wherein is used to perform normalization a point cloudization module configured to calculate the future time human body posture features into a human body model, and point cloudize the human body model to obtain predicted dynamic human body point clouds; A trajectory planning module is configured to: acquire an object point cloud, which is taken as an obstacle together with the dynamic human body point cloud, and perform collision avoidance trajectory planning; Wherein, including: (1) The state of the selected model predictive control is wherein q, , are the angle, angular velocity, and angular acceleration of the robot arm, respectively. (2) The motion model of the robotic arm is the optimal control of a discrete stochastic dynamic system. ,in, The state at time t In the input Switching to under control state of time The probability of; (3) sampling from the current input distribution batch, length of control sequence for each batch, input the sampled control sequence into the robot approximate dynamics model to obtain the trajectory and the corresponding state sequence and cost sequence ; wherein the batch of the state sequence and the corresponding cost sequence is N, and the length is H; Parallelly compute the cost of each batch where, is the length of the batch is the terminal cost except for the batch, is the discount factor to bias the immediate cost; the sampling batch is N, is the cost of inputting a batch control sequence, is the hth control of the control sequence of the current batch with length H at time t, and similarly is the hth state of the state sequence of the current batch with length H at time t, is the parameter of the terminal cost, which is a reasonable parameter artificially specified; (4) from The first input of the batch control sequence with the lowest cost is selected as the input of the current state, i.e. the optimal input. A control sequence is obtained, which is a control sequence with a field of view length H, and the first control sequence is selected as an input; (5) using the current sampling control sequence to update the distribution of the next sampling, updating the variance and mean of the distribution, for the next sampling; the mean updating equation: , the covariance updating equation: , and is the updating step, the weight the calculation equation: ; wherein, is the mean of the sampling distribution of the hth position of the entire sequence at time t, is the covariance of the sampling distribution of the hth position of the entire sequence at time t; the weight in the calculation formula, represents the control of the hth position of the i th batch sequence, represents the state of the hth position of the i th batch sequence, is the parameter of the terminal cost; (6) The control sequence is sampled from the distribution again, the control input is calculated, and the input distribution is updated, and the cycle is repeated until the target point is reached. , and are weights of the three linear layers.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps in the collision avoidance planning method of the robot manipulator according to any one of claims 1-7.

10. A computer device, comprising a computer readable storage medium, a processor, and a computer program stored on the computer readable storage medium and executable on the processor, wherein, The processor executes the computer program to implement the steps in the collision avoidance planning method of the robot manipulator according to any one of claims 1-7.

Images