Robot motion generation method and robot control system

Abstract

The invention provides a robot action generation method and a robot control system, and the method comprises the steps: obtaining multi-modal observation data at a current moment; the modal weight and the initial noise trajectory of each modal are acquired; and according to the multi-modal observation data, the modal weight of each modal and the noise estimation model corresponding to each modal, performing iterative exploration on the initial noise trajectory according to a time inverse sequence, and generating an action trajectory sequence of the robot. According to the method, noise estimation of the pre-trained single-mode diffusion strategy can be combined during reasoning, and information of different modes can be efficiently fused, so that on the premise of avoiding complexity of multi-mode data collection and joint training, multi-mode strategy fusion is realized, and meanwhile, the multi-mode data collection and joint training efficiency is improved. And the flexibility, the real-time adaptability, the robustness and the expandability of the action generation process are enhanced. Moreover, by avoiding multi-modal joint training, the training cost and the computing resource demand are greatly reduced.

Description

Technical Field

[0001] The present application relates to the field of robot control technology, and in particular to a robot motion generation method and a robot control system. Background Art

[0002] In robot policy learning, diffusion models have gradually become the mainstream method for generating robust policies due to their powerful multimodal action distribution modeling capabilities.

[0003] At present, diffusion models generally support the processing of unimodal data. If the diffusion model is required to support the processing of multimodal data, it is necessary to learn joint representations through end-to-end training to obtain a diffusion model that supports multimodality.

[0004] However, when training a multimodal diffusion model using this approach, the multimodal data needs to be mapped into a shared latent space. This results in the mixing of the distributional characteristics of each modality (such as the mean and variance of Gaussian noise), making them impossible to control individually. Furthermore, the weights of the distribution combinations cannot be adjusted during inference based on real-time requirements (such as modal confidence), leading to the problem of ignoring the dynamic dependencies between different modalities at the distribution level. Furthermore, this processing approach requires a large amount of aligned multimodal data, and the number of model parameters increases linearly with the number of modalities, resulting in high training costs. Furthermore, it relies on complex architectural designs, making it difficult to adapt to the incremental expansion of new modalities and resulting in low flexibility. Summary of the Invention

[0005] The purpose of this application is to provide a robot motion generation method and a robot control system to address the deficiencies in the above-mentioned prior art, so as to solve the problems in the prior art of ignoring the dynamic dependencies between different modalities at the distribution level, high training costs and low flexibility.

[0006] To achieve the above objectives, the technical solutions adopted in the embodiments of the present application are as follows: In a first aspect, an embodiment of the present application provides a method for generating robot motions, the method comprising: Acquire multimodal observation data at the current moment, wherein the multimodal observation data includes: observation data under multiple modes; Obtaining a modal weight and an initial noise trajectory of each mode, wherein the initial noise trajectory is used to represent an initial motion trajectory before motion control of the robot; Based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse chronological order to generate a motion trajectory sequence of the robot. The noise estimation model corresponding to each modality is obtained through pre-training. The motion trajectory sequence is used to indicate the motion sequence of the robot within a preset time period after the current moment.

[0007] In one possible implementation, the iterative exploration of the initial noise trajectory in reverse chronological order based on the multimodal observation data, the modal weight of each modality, and the noise estimation model corresponding to each modality to generate the motion trajectory sequence of the robot includes: Iteratively exploring the initial noise trajectory in reverse chronological order based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality to sequentially generate the motion trajectory of the robot at each time step; According to the motion trajectory of the robot at each time step, a motion trajectory sequence of the robot is generated.

[0008] In one possible implementation, the iterative exploration of the initial noise trajectory in reverse time order based on the multimodal observation data, the modal weight of each modality, and the noise estimation model corresponding to each modality to sequentially generate the motion trajectory of the robot at each time step includes: Obtaining the motion trajectory of the current time step, wherein if the current time step is the first time step, the initial noise trajectory is used as the motion trajectory of the current time step; otherwise, the motion trajectory of the current time step is determined according to the motion trajectory of the next time step after the current time step; Determining noise information of each modality based on the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory of the current time step; Determining the total noise at the current time step according to the modal weights of the modes and the noise information of the modes; The motion trajectory of the robot at a time step before the current time step is generated according to the motion trajectory of the current time step and the total noise at the current time step.

[0009] In one possible implementation, determining the noise information of each modality based on the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory of the current time step includes: The observation data of the first modality and the motion trajectory of the current time step are input into the noise estimation model corresponding to the first modality to predict the noise information of the first modality, wherein the first modality is any modality in the multi-modality.

[0010] In a possible implementation, determining the total noise at the current time step according to the modal weights of the modalities and the noise information of the modalities includes: Calculate the product of the model weight of each mode and the corresponding noise information respectively to obtain the product result of each mode; The product results of the modes are summed to obtain the total noise at the current time step.

[0011] In a possible implementation, generating the motion trajectory of the robot at a time step previous to the current time step based on the motion trajectory at the current time step and the total noise at the current time step includes: Get the gradient step coefficient, denoising intensity coefficient and additional Gaussian noise; Calculating a first product of the gradient step coefficient and the total noise at the current time step; Calculating the difference between the motion trajectory of the current time step and the first product; Calculating a second product of the difference and the denoising strength coefficient; The sum of the second product and the additional Gaussian noise is calculated as the motion trajectory of the robot in the time step before the current time step.

[0012] In one possible implementation, the process of determining the initial noise trajectory includes: Get the trajectory dimension of the action trajectory sequence; According to the trajectory dimension, the initial noise trajectory is generated by sampling independently element by element from the standard Gaussian distribution.

[0013] In a possible implementation, the process of determining the modal weight of each modality includes: Obtain the current task success rate and current confidence of each modality; The modality weight of each modality is calculated based on the current task success rate of each modality and the current confidence of each modality.

[0014] In one possible implementation, the training process of the noise estimation model corresponding to each modality includes: Acquire a first training data set in a first modality, wherein the first training data set includes a real action sequence and observation data in the first modality; According to the preset noise scheduling coefficient, the real action sequence is iteratively processed at each time step to generate the action trajectory at each time step in sequence; According to the motion trajectory at each time step, the noise estimation model corresponding to the first modality is trained.

[0015] In the second aspect, another embodiment of the present application provides a control system for a robot, comprising: a processor, a storage medium and a bus, wherein the storage medium stores machine-readable instructions executable by the processor. When the control system of the robot is running, the processor communicates with the storage medium through the bus, and the processor executes the machine-readable instructions to perform the steps of any method described in the first aspect above.

[0016] The beneficial effects of the present application are: by obtaining the multimodal observation data at the current moment, the modal weights of each modality, and the initial noise trajectory, and based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to generate a sequence of robot motion trajectories, and the noise estimation of the pre-trained single-modal diffusion strategy can be combined during reasoning to efficiently fuse information from different modalities, thereby achieving the fusion of multimodal strategies while avoiding the complexity of multimodal data collection and joint training. At the same time, it also supports dynamic adjustment of modal weights and multimodal expansion. When adding a new modality, only the corresponding noise estimation model needs to be trained for seamless integration, without the need to reconstruct the multimodal model, thereby enhancing the flexibility, real-time adaptability, robustness, and scalability of the motion generation process. Moreover, by avoiding multimodal joint training, the training cost and computing resource requirements are greatly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A schematic diagram of a flow chart of a robot action generation method provided in an embodiment of the present application; Figure 2 A schematic diagram of a flow chart for generating a robot's motion trajectory sequence in the robot motion generation method provided in an embodiment of the present application; Figure 3 A schematic diagram of a process for sequentially generating the robot's motion trajectory at each time step in the robot motion generation method provided in an embodiment of the present application; Figure 4 A schematic diagram of a flow chart for determining the total noise at the current time step in the robot motion generation method provided in an embodiment of the present application; Figure 5 A schematic diagram of a flow chart for generating a robot's motion trajectory at a time step preceding a current time step in a robot motion generation method provided in an embodiment of the present application; Figure 6 A schematic diagram of a process for determining an initial noise trajectory in the robot motion generation method provided in an embodiment of the present application; Figure 7 A schematic diagram of a process for determining modal weights in the robot motion generation method provided in an embodiment of the present application; Figure 8 A schematic diagram of a process for training noise estimation models corresponding to each modality in the robot motion generation method provided in an embodiment of the present application; Figure 9 A schematic diagram of the structure of the control system of the robot provided in an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. It should be understood that the drawings in the present application only serve the purpose of illustration and description and are not used to limit the scope of protection of the present application. In addition, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate the operations implemented according to some embodiments of the present application. It should be understood that the operations of the flowcharts can be implemented out of sequence, and steps without logical context can be reversed or implemented simultaneously. In addition, those skilled in the art, under the guidance of the contents of this application, can add one or more other operations to the flowchart, or remove one or more operations from the flowchart.

[0020] In addition, the described embodiments are only a part of the embodiments of the present application, rather than all of the embodiments. The components of the embodiments of the present application generally described and shown in the drawings here can be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present application.

[0021] It should be noted that the term "comprising" will be used in the embodiments of the present application to indicate the existence of the features declared thereafter, but does not exclude the addition of other features.

[0022] At present, diffusion models generally support the processing of unimodal data. If the diffusion model is required to support the processing of multimodal data, it is necessary to learn joint representations through end-to-end training to obtain a diffusion model that supports multimodality.

[0023] However, when training a multimodal diffusion model using this approach, the multimodal data needs to be mapped into a shared latent space. This results in the mixing of the distributional characteristics of each modality (such as the mean and variance of Gaussian noise), making them impossible to control individually. Furthermore, the weights of the distribution combinations cannot be adjusted during inference based on real-time requirements (such as modal confidence), leading to the problem of ignoring the dynamic dependencies between different modalities at the distribution level. Furthermore, this processing approach requires a large amount of aligned multimodal data, and the number of model parameters increases linearly with the number of modalities, resulting in high training costs. Furthermore, it relies on complex architectural designs, making it difficult to adapt to the incremental expansion of new modalities and resulting in low flexibility.

[0024] Based on the above problems, the embodiment of the present application proposes a robot motion generation method. By obtaining the multimodal observation data, the modal weights of each modality, and the initial noise trajectory at the current moment, and based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to generate a sequence of robot motion trajectories. The noise estimation of the pre-trained single-modal diffusion strategy can be combined during inference to achieve the fusion of multimodal strategies. At the same time, it also supports dynamic weights and modal expansion. When adding a new modality, there is no need to reconstruct the multimodal model, which enhances the flexibility, real-time adaptability, and scalability of the motion generation process. In addition, it also reduces training costs and improves the robustness of the motion generation process.

[0025] First, the relevant background involved in the robot action generation method provided in the embodiment of the present application is described in detail.

[0026] It can be understood that the robot motion generation method provided in the embodiments of the present application can be deployed in a robot, for example, can be deployed in a robot or a robot-like device.

[0027] For example, a humanoid robot is a type of robot that mimics a human's appearance, movement, and behavior. It typically possesses a head, torso, arms, legs, and other human-like structures, and can perform human-like functions such as running, dancing, jumping, swimming, fighting, grasping, and expressing emotions. Furthermore, each arm is composed of multiple joints, enabling it to perform different actions depending on the scenario.

[0028] Optionally, the robot may include a mechanical structure, a drive system, a sensor system, a control system, and a power supply system.

[0029] The mechanical structure includes at least one robotic arm, an end effector corresponding to each robotic arm, two legs and two feet, and the mechanical structure is used to perform specific actions.

[0030] Among them, the drive system includes: servo motors, etc., which are used to control the rotation or linear motion of each joint in the robotic arm and provide precise power output.

[0031] Optionally, the sensing system may include sensor devices of different modalities such as torque sensors, visual cameras, inertial measurement units, and tactile sensors. The sensing system is used to detect environmental information in real time to obtain multimodal observation data to achieve control of the robot.

[0032] Among them, the control system includes: a processor and a memory, etc. The memory stores machine-readable instructions executable by the processor. When the robot's control system is running, the processor executes the machine-readable instructions to execute the steps of the robot motion generation method provided in the embodiment of the present application, thereby controlling the operation of the robot.

[0033] The following describes in detail the robot motion generation method provided in the embodiments of the present application in combination with multiple embodiments.

[0034] Figure 1 A flow chart of a robot action generation method provided in an embodiment of the present application, referring to Figure 1 As shown, the execution subject of the method can be any electronic device with processing capabilities, such as the control system of the above-mentioned robot, and the method includes: S101. Obtain multimodal observation data at the current moment.

[0035] It can be understood that during the robot motion generation process, the robot can predict the motion trajectory sequence in the future period of time through the multimodal observation data at the current moment.

[0036] The current moment refers to the decision moment when the robot starts to generate the action sequence, that is, the current time point when multimodal observation data is acquired and action generation is started.

[0037] Exemplarily, obtaining multimodal observation data at the current moment can be understood as obtaining the observation data of the first modality at the current moment, and obtaining the observation data of the second modality at the moment corresponding to the current moment, and using the obtained observation data of the first modality and the observation data of the second modality as the multimodal observation data at the current moment.

[0038] Among them, multimodal observation data includes: observation data under multiple modes.

[0039] Exemplarily, the modality types of the multimodal observation data may include: one or more of visual modalities (Visual Modalities), non-visual modalities (Non-Visual Modalities), and advanced modalities (Advanced Modalities).

[0040] Visual modalities include RGB images, depth maps / point clouds, infrared / thermal imaging, and event cameras. Non-visual modalities include touch / force, inertial measurement units (IMUs), sound / speech, and proprioception. Specifically, proprioception refers to internal sensor data such as joint angles and motor currents.

[0041] Among them, advanced modalities include: language commands, radar / lidar, and brain-computer interface (BCI).

[0042] S102: Obtain the modal weight of each mode and the initial noise trajectory.

[0043] Optionally, the modality weight of each modality is used to quantify the relative importance of different modalities to the final generated motion trajectory.

[0044] Exemplarily, the modal weight of each modality can be a fixed modal weight of each modality obtained in advance, or a dynamic modal weight of each modality calculated by a predefined modal weight calculation formula, or a dynamic modal weight of each modality predicted by a pre-trained modal weight prediction model through multimodal observation data.

[0045] In one example, the modality weight of each modality may be any constant between 0 and 1.

[0046] In another example, the modal weight of each modality may be any constant between 0 and 1, and the sum of the modal weights of each modality may also be 1.

[0047] Optionally, the initial noise trajectory is used to characterize the initial motion trajectory before the robot is controlled. The initial noise trajectory is used to provide a random starting point motion for the robot, which is the starting point of the generated motion trajectory sequence, so that the initial noise trajectory can be converted into a physically feasible motion sequence through multimodal observation data in subsequent steps.

[0048] Exemplarily, the initial noise trajectory can be sampled from a standard Gaussian distribution, so that the initial noise trajectory can be quickly obtained while improving the reversibility and convergence of the generated robot motion trajectory sequence.

[0049] For example, the initial noise trajectory can be sampled from a uniform distribution, which is simple and easy to implement and avoids obtaining extreme values.

[0050] For example, the initial noise trajectory can be sampled from a mixture of Gaussian distributions, thereby being able to capture multimodal motion distributions.

[0051] Exemplarily, the initial noise trajectory can also be sampled from the robot's historical motion data, so that the obtained initial noise trajectory is closer to the real motion distribution, accelerating the denoising convergence.

[0052] Exemplarily, the initial noise trajectory can also be obtained by a pre-trained initial noise trajectory generation network. The multimodal observation data at the current moment is input into the initial noise trajectory generation network to generate the initial noise trajectory, thereby accelerating the denoising convergence.

[0053] S103 , based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, iteratively explore the initial noise trajectory in reverse time order to generate a motion trajectory sequence of the robot.

[0054] Optionally, after obtaining the model weights of each modality and the noise estimation model corresponding to each modality, the initial noise trajectory can be iteratively explored in reverse time order based on the multimodal observation data, the modal weights of each modality and the noise estimation model corresponding to each modality to generate a sequence of robot motion trajectories.

[0055] The noise estimation model corresponding to each mode is obtained through pre-training, and the motion trajectory sequence is used to indicate the motion sequence of the robot within a preset time period after the current moment.

[0056] For example, a diffusion strategy (DP) can be pre-trained for each modality, thereby using the diffusion strategy (DP) for each modality as the noise estimation model corresponding to each modality. The noise estimation model corresponding to each modality is used to capture the semantic information of each modality and model the geometric constraints of each modality.

[0057] For example, at each time step, noise prediction can be performed based on multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, and the predicted noise can be added to the initial noise trajectory in reverse order of the time steps to generate a sequence of robot motion trajectories.

[0058] Exemplarily, at each time step, noise prediction can be performed based on the multimodal observation data, the modal weights of each modality, the noise estimation model corresponding to each modality, and the initial noise trajectory to generate the noise trajectory of each time step. Noise prediction can be performed in reverse time order based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, and the predicted noise can be added to the noise trajectory of each time step in sequence to generate a sequence of motion trajectories of the robot.

[0059] For example, taking the robot grasping task as an example, the multimodal observation data may include the RGB image, point cloud data and proprioception data at the current moment. By executing S101-S103, a motion trajectory sequence can be generated, which can represent the robot arm posture and the end effector posture every 0.1 second in the next 2 seconds.

[0060] For example, after obtaining a motion trajectory sequence representing the robot arm posture and the end effector posture every 0.1 seconds in the next 2 seconds, the robot can execute the motion trajectory sequence according to the motion trajectory sequence.

[0061] In this embodiment, by obtaining the multimodal observation data at the current moment, the modal weights of each modality, and the initial noise trajectory, and based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to generate a sequence of robot motion trajectories. During inference, the noise estimation of the pre-trained single-modal diffusion strategy can be combined to efficiently fuse information from different modalities, thereby achieving the fusion of multimodal strategies while avoiding the complexity of multimodal data collection and joint training. At the same time, it also supports dynamic adjustment of modal weights and multimodal expansion. When adding a new modality, only the corresponding noise estimation model needs to be trained for seamless integration without the need to reconstruct the multimodal model, thereby enhancing the flexibility, real-time adaptability, robustness, and scalability of the motion generation process. In addition, by avoiding multimodal joint training, the training cost and computing resource requirements are greatly reduced.

[0062] In one possible implementation, Figure 2 A schematic diagram of a flow chart for generating a robot's motion trajectory sequence in the robot motion generation method provided in an embodiment of the present application, referring to Figure 2 As shown, in the above S103, based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to generate the robot's motion trajectory sequence, including: S201. Based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to sequentially generate the robot's motion trajectory at each time step.

[0063] Optionally, at each time step, noise prediction can be performed based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, and the predicted noise can be added to the initial noise trajectory in reverse order of the time steps to generate the robot's motion trajectory at each time step.

[0064] Optionally, at each time step, noise prediction can be performed based on the multimodal observation data, the modal weights of each modality, the noise estimation model corresponding to each modality, and the initial noise trajectory to generate the noise trajectory for each time step. Noise prediction can be performed in reverse time order based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, and the predicted noise can be added to the noise trajectory of each time step in sequence to generate the motion trajectory of the robot at each time step.

[0065] Among them, the motion trajectory of the robot at each time step refers to the control instructions of the robot at each time step, which can be used to control the joint angles and end effector posture of the robot.

[0066] For example, continuing with the above-mentioned robot grasping task as an example, the multimodal observation data includes the RGB image, point cloud data and proprioception data at the current moment. By executing S201, assuming that the generated motion trajectory sequence is 2 seconds long and the control frequency is 10 Hz, the motion trajectory of the robot in 20 time steps can be generated.

[0067] S202: Generate a motion trajectory sequence of the robot according to the motion trajectory of the robot at each time step.

[0068] Optionally, after obtaining the motion trajectory of the robot at each time step, the motion trajectory of the robot at each time step may be sorted in reverse chronological order to generate a motion trajectory sequence of the robot.

[0069] For example, continuing with the above-mentioned robot grasping task as an example, assuming that the motion trajectory duration is 2 seconds and the control frequency is 10 Hz, the motion trajectory of the robot in 20 time steps is generated. By executing S202, a motion trajectory sequence including the motion trajectory in 20 time steps can be generated.

[0070] Through multimodal observation data, the modal weights of each mode and the noise estimation model corresponding to each mode, the initial noise trajectory is iteratively explored in reverse time order, and the robot's motion trajectory at each time step is generated in sequence. Based on the robot's motion trajectory at each time step, the robot's motion trajectory sequence is generated. The motion trajectory can be gradually refined by diffusing the time step, thereby achieving global optimization of the initial noise trajectory. In addition, the time step of motion generation can be decoupled from the time step of physical control, thereby efficiently fusing information from different modalities.

[0071] In one possible implementation, Figure 3 A schematic diagram of a process for sequentially generating the robot's motion trajectory at each time step in the robot motion generation method provided in the embodiment of the present application, referring to Figure 3As shown, in the above S201, based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse time order to sequentially generate the robot's motion trajectory at each time step, including: S301: Obtain the motion trajectory of the current time step.

[0072] Optionally, during the iterative exploration of the initial noise trajectory, the motion trajectory of the current time step can be obtained at each time step. The number of time steps can be determined in combination with the control frequency and the duration of the generated motion trajectory sequence.

[0073] Among them, if the current time step is the first time step, the initial noise trajectory is used as the action trajectory of the current time step; otherwise, the action trajectory of the current time step is determined according to the action trajectory of the next time step after the current time step.

[0074] S302 : Determine the noise information of each modality based on the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory of the current time step.

[0075] Optionally, at the current time step, based on the observation data of each modality in the multimodal observation data, the motion trajectory at the current time step, and the noise estimation model corresponding to each modality, the noise information of each modality is estimated respectively to obtain the noise information of each modality.

[0076] Among them, the noise information of each modality can be understood as the quantitative correction information of the motion trajectory deviation implicit in the observation data of the current modality. Specifically, when the initial noise trajectory is the initial noise trajectory sampled from the standard Gaussian distribution, the noise information of each modality can be expressed in the form of a Gaussian distribution to express the direction and uncertainty of the correction.

[0077] S303: Determine the total noise at the current time step according to the modal weight of each mode and the noise information of each mode.

[0078] Optionally, after obtaining the noise information of each modality and the modal weight of each modality, the noise information of each modality can be weightedly combined by the modal weight of each modality to determine the total noise at the current time step, thereby achieving robust motion generation.

[0079] Optionally, after obtaining the noise information of each modality and the modal weight of each modality, the noise information of each modality and the modal weight of each modality can be combined through the attention mechanism to determine the total noise at the current time step, thereby achieving robust action generation.

[0080] Optionally, after obtaining the noise information of each modality and the modal weight of each modality, the noise information of each modality and the modal weight of each modality can be probabilistically fused through covariance to determine the total noise at the current time step, thereby achieving robust motion generation.

[0081] Among them, the total noise at the current time step can indicate the optimization direction of the motion trajectory sequence jointly driven by multimodal observation data.

[0082] For example, when the initial noise trajectory is the initial noise trajectory sampled from a standard Gaussian distribution, and the noise information of each modality is expressed in the form of a Gaussian distribution to represent the direction and uncertainty of the correction, by determining the total noise, the noise information of each modality can be integrated to generate the optimal action adjustment strategy, thereby converting the "hard constraints" in traditional motion planning into probabilistic soft constraints, and naturally handling multimodal uncertainty through Gaussian distribution.

[0083] S304: Generate the motion trajectory of the robot in the time step before the current time step according to the motion trajectory of the current time step and the total noise in the current time step.

[0084] Optionally, after obtaining the total noise at the current time step, the motion trajectory of the current time step can be optimized using the total noise at the current time step, thereby generating the motion trajectory of the robot at the time step before the current time step.

[0085] For example, the total noise at the current time step can be subtracted from the motion trajectory at the current time step to obtain the motion trajectory of the robot at the time step before the current time step.

[0086] By obtaining the motion trajectory at the current time step and determining the noise information of each modality using multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory at the current time step, the total noise at the current time step is determined using the modal weights and noise information of each modality. This generates a noise correction signal that integrates the advantages of each modality. This allows the robot's motion trajectory at the previous time step to be generated based on the motion trajectory at the current time step and the total noise at the current time step. This fused motion trajectory simultaneously meets both high-level task requirements and low-level motion constraints. Furthermore, modal weight adjustment automatically mitigates the impact of unreliable modal data, significantly improving robustness in unstructured environments. Furthermore, the time-step iterative trajectory generation method ensures spatiotemporal continuity at the motion sequence level. Furthermore, the computational overhead of explicit constraint processing is eliminated, significantly reducing the iteration cost associated with modal expansion, achieving a balance between computational efficiency and real-time performance.

[0087] In one possible implementation, determining the noise information of each modality in S302 according to the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory at the current time step includes: The observation data of the first mode and the motion trajectory of the current time step are input into the noise estimation model corresponding to the first mode to predict the noise information of the first mode.

[0088] The first mode is any mode in the multi-mode.

[0089] Exemplarily, the observation data of the first modality and the motion trajectory of the current time step are input into the noise estimation model corresponding to the first modality, the observation features in the observation data of the first modality are extracted through the encoder in the noise estimation model corresponding to the first modality, and the trajectory features in the motion trajectory of the current time step are extracted through the time series encoder in the noise estimation model corresponding to the first modality, and the observation features and the trajectory features are spliced ​​or cross-attention fused to generate fused features.

[0090] Exemplarily, based on the fused features, the estimated noise is calculated through the network body in the noise estimation model corresponding to the first modality.

[0091] Exemplarily, the current time step is embedded through sinusoidal position encoding or learnable embedding, and the intensity of the estimated noise is adjusted to obtain noise information of the first modality.

[0092] For example, when the initial noise trajectory is the initial noise trajectory sampled from the standard Gaussian distribution, the noise information corresponding to each mode can be expressed as ,in, , For modal The corresponding noise information, For modal The corresponding noise estimation model, is the current time step The trajectory of the movement, For modal The observation data, is a Gaussian distribution, is the mean of the Gaussian distribution, is the expected value of the noise information, indicating the mode Current action trajectory The optimal correction direction and amplitude of is the covariance matrix of the noise information, indicating the mode Uncertainty assessment of the corrections.

[0093] By inputting the observed data of the first modality and the motion trajectory of the current time step into the noise estimation model corresponding to the first modality, the noise information of the first modality is predicted. This can preserve the specificity of the noise information of each modality and avoid interference between modalities. At the same time, each modality focuses on its area of ​​strength, and through combination, it maximizes its effectiveness, improving task adaptability and the robustness of the motion generation process. Furthermore, by predicting the noise information of the first modality through the motion trajectory of the current time step, the spatiotemporal consistency of the motion sequence can be maintained. Furthermore, parallel computing can be performed to infer the noise information of multiple modalities in real time. Furthermore, it can reduce data and computing requirements, facilitating deployment and scalability.

[0094] In one possible implementation, Figure 4 A schematic diagram of a flow chart for determining the total noise at the current time step in the robot motion generation method provided in an embodiment of the present application, referring to Figure 4 As shown, in the above S303, the total noise at the current time step is determined according to the modal weight of each mode and the noise information of each mode, including: S401 , respectively calculating the product of the model weight of each mode and the corresponding noise information to obtain the product result of each mode.

[0095] Alternatively, the product of the model weights of each modality and the noise information corresponding to each modality can be calculated separately to obtain the product results of each modality. The product results of each modality are used to indicate the strength and direction of each modality's contribution to the robot's motion trajectory correction in the current task scenario, and can represent the local correction parameters of each modality to the motion trajectory at the current time step.

[0096] For example, when the initial noise trajectory is the initial noise trajectory sampled from the standard Gaussian distribution, the noise information corresponding to each mode can be expressed as , then the product of each mode can be expressed as ,in, , For modal The modal weight of For modal The product result is Inherits Gaussianity and reflects the uncertainty of the correction amount, For modal The corresponding noise information, For modal The corresponding noise estimation model, is the current time step The trajectory of the movement, For modal The observation data, is a Gaussian distribution, is the mean of the Gaussian distribution, is the expected value of the noise information, indicating the mode Current action trajectory The optimal correction direction and amplitude of is the covariance matrix of the noise information, indicating the mode Uncertainty assessment of the corrections.

[0097] S402: Sum the product results of each mode to obtain the total noise at the current time step.

[0098] Optionally, the sum of the product results of each mode is calculated to obtain the total noise at the current time step.

[0099] For example, when the initial noise trajectory is the initial noise trajectory sampled from the standard Gaussian distribution, the noise information corresponding to each mode is , modal The product of can be expressed as , then the total noise at the current time step can be expressed as .in, For modal The modal weight of For modal The product result is Inherits Gaussianity and reflects the uncertainty of the correction amount, For modal The corresponding noise information, For modal The corresponding noise estimation model, is the current time step The trajectory of the movement, For modal The observation data, is a Gaussian distribution, is the mean of the Gaussian distribution, is the expected value of the noise information, indicating the mode Current action trajectory The optimal correction direction and amplitude of is the covariance matrix of the noise information, indicating the mode Uncertainty assessment of the corrections.

[0100] By calculating the product of each modality's model weight and the corresponding noise information, the resulting product is summed to obtain the total noise at the current time step. This weighted expression quantifies the strength and direction of each modality's contribution to action correction, and linear combinations resolve conflicts and propagate uncertainty. Furthermore, due to its lightweight computation, it also supports high-frequency policy updates.

[0101] In one possible implementation, Figure 5 A schematic diagram of a flow chart of generating a robot's motion trajectory at the time step before the current time step in the robot motion generation method provided in the embodiment of the present application, referring to Figure 5 As shown, in the above S304, the motion trajectory of the robot at the time step before the current time step is generated according to the motion trajectory of the current time step and the total noise at the current time step, including: S501: Obtain a gradient step coefficient, a denoising intensity coefficient, and additional Gaussian noise.

[0102] Optionally, you can get the gradient step coefficient , denoising strength coefficient and additive Gaussian noise .

[0103] Among them, the gradient step coefficient Used to adjust the update step size of noise estimation and affect the convergence speed.

[0104] Among them, the denoising strength coefficient Used to control the contribution of total noise to trajectory updates.

[0105] Among them, the additional Gaussian noise Used to prevent deterministic updates from falling into local optimality and enhance exploration.

[0106] For example, , , .

[0107] The gradient step size coefficient and the denoising strength coefficient can be used to control the process of correcting the motion trajectory of the current time step by the total noise. By adding Gaussian noise, it can be ensured that the strategy retains a certain degree of randomness when generating the trajectory, preventing the robot arm's motion from falling into a repetitive pattern.

[0108] S502: Calculate the first product of the gradient step coefficient and the total noise at the current time step.

[0109] Optionally, calculate the gradient step size coefficient and the total noise at the current time step The first product of .

[0110] S503: Calculate the difference between the motion trajectory of the current time step and the first product.

[0111] Optionally, calculate the current time step Movement trajectory and the first product The difference .

[0112] S504: Calculate the second product of the difference and the denoising strength coefficient.

[0113] Optionally, calculate the difference and denoising strength coefficient The second product of .

[0114] S505 : Calculate the sum of the second product and the additional Gaussian noise as the motion trajectory of the robot in the time step before the current time step.

[0115] Optionally, calculate the second product With additive Gaussian noise of , as the robot's motion trajectory in the previous time step of the current time step.

[0116] The robot's motion trajectory in the previous time step of the current time step is calculated by the gradient step coefficient, denoising intensity coefficient, additional Gaussian noise, motion trajectory of the current time step and total noise in the current time step, which can generate robust and adaptive motion sequences under multimodal constraints.

[0117] In one possible implementation, Figure 6 A schematic diagram of a flow chart for determining the initial noise trajectory in the robot motion generation method provided in an embodiment of the present application, referring to Figure 6 As shown in Figure 2, the process of determining the initial noise trajectory includes: S601: Obtain the trajectory dimension of the action trajectory sequence.

[0118] Optionally, the trajectory dimension of the motion trajectory sequence to be generated is obtained, wherein the trajectory dimension includes the motion dimension and the number of time steps.

[0119] For example, suppose we need to generate a motion trajectory sequence consisting of 20 time steps, and each motion is 7-dimensional (e.g., a 7-DOF robotic arm), then The action dimension is 7 and the number of time steps is 20.

[0120] S602 : Based on the trajectory dimension, independently sample element by element from the standard Gaussian distribution to generate an initial noise trajectory.

[0121] Optionally, according to the trajectory dimension, Each element of The initial noise trajectory is obtained by sampling independently from the standard Gaussian distribution, .in, Indicates the The first joint time steps.

[0122] For example, taking the robot grasping task, the motion trajectory sequence is the motion trajectory sequence of a 7-joint robotic arm within 2 seconds, and the control frequency is 10Hz. The initial noise trajectory is a 140-dimensional vector, and each element value of the vector is independently sampled from ,For example: .

[0123] By taking the trajectory dimension of the motion trajectory sequence and sampling independently element by element from the standard Gaussian distribution according to the trajectory dimension, the initial noise trajectory is generated. This ensures that the obtained initial noise trajectory allows the strategy to generate a diverse sequence of motion trajectories, avoiding falling into local optimality. At the same time, it can also provide a random starting point for the robot's motion generation.

[0124] In one possible implementation, Figure 7 A flow chart of determining modal weights in the robot action generation method provided in the embodiment of the present application, with reference to Figure 7 As shown in FIG, the process of determining the modal weight of each mode includes: S701: Obtain the current task success rate and current confidence of each modality.

[0125] Optionally, for each mode, a sliding window method may be used to maintain a task success rate ri within a sliding window, where the task success rate may be, for example, the success rate of the most recent 10 actions.

[0126] Optionally, for each modality, the current confidence rj of each modality can also be obtained based on the historical action success rate, the inverse variance of the noise estimate, or the modality-specific reward.

[0127] S702: Calculate the modality weight of each modality based on the current task success rate of each modality and the current confidence of each modality.

[0128] Optionally, the modality weight of each modality may be calculated based on the current task success rate of each modality and the current confidence of each modality.

[0129] For example, for mode i, the modal weight of mode i can be calculated by referring to the following formula: :

[0130] in, is the total number of modes.

[0131] By obtaining the current task success rate and current confidence of each modality, and calculating the modal weight of each modality based on the current task success rate and current confidence of each modality, it is possible to respond to changes in the quality of the observed data in real time, and always give priority to trusting high-confidence modalities.

[0132] In one possible implementation, Figure 8 A flow chart of the process of training the noise estimation model corresponding to each mode in the robot motion generation method provided in the embodiment of the present application, referring to Figure 8 As shown in Figure 2, the training process of the noise estimation model corresponding to each mode includes: S801: Obtain a first training data set in a first modality.

[0133] Optionally, a first training data set in a first modality may be obtained, wherein the first training data set includes a real action sequence and observation data in the first modality.

[0134] Optionally, a diffusion model may be used as the initial noise estimation model, for example, a Denoising Diffusion Probabilistic Model (DDPM).

[0135] S802: According to the preset noise scheduling coefficient, the real action sequence is iteratively processed at each time step to generate the action trajectory at each time step in sequence.

[0136] Optionally, according to a preset noise scheduling coefficient αt, Gaussian noise is gradually added to the real action sequence, thereby corrupting the data through Gaussian noise and constructing a diffusion chain.

[0137] For example, The motion trajectory under time steps It can be:

[0138] in, , is the noise scheduling coefficient.

[0139] S803: Training a noise estimation model corresponding to the first modality according to the motion trajectory at each time step.

[0140] Optionally, the action trajectory at each time step can be , train the noise estimation model corresponding to the first mode.

[0141] For example, the training objective may be to minimize the noise estimation error.

[0142] For example, after training, each modality is saved The noise estimation model ϵθi and noise scheduling parameters {αt,σt}.

[0143] The present application also provides a robot, such as Figure 9As shown, Figure 9 This is a schematic diagram of the structure of a robot control system provided in an embodiment of the present application, comprising: a processor 901, a memory 902, and optionally, a bus 903. The memory 902 stores machine-readable instructions executable by the processor 901. When the robot control system is running, the processor 901 communicates with the memory 902 via the bus 903. When the machine-readable instructions are executed by the processor 901, the steps of the above-described robot motion generation method are performed.

[0144] The above is only a specific implementation method of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the protection scope of the present application.

Claims

1. A robot motion generation method, characterized in that: include: Acquire multimodal observation data at the current moment, wherein the multimodal observation data includes: observation data under multiple modes; Obtaining a modal weight and an initial noise trajectory of each mode, wherein the initial noise trajectory is used to represent an initial motion trajectory before motion control of the robot; Based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, the initial noise trajectory is iteratively explored in reverse chronological order to generate a motion trajectory sequence of the robot. The noise estimation model corresponding to each modality is obtained through pre-training. The motion trajectory sequence is used to indicate the motion sequence of the robot within a preset time period after the current moment.

2. The robot motion generation method according to claim 1, characterized in that: The method of iteratively exploring the initial noise trajectory in reverse time order based on the multimodal observation data, the modal weight of each modality, and the noise estimation model corresponding to each modality to generate the motion trajectory sequence of the robot includes: Iteratively exploring the initial noise trajectory in reverse time order based on the multimodal observation data, the modal weights of each modality, and the noise estimation model corresponding to each modality, and sequentially generating the motion trajectory of the robot at each time step; According to the motion trajectory of the robot at each time step, a motion trajectory sequence of the robot is generated.

3. The robot motion generation method according to claim 2, characterized in that: The method of iteratively exploring the initial noise trajectory in reverse time order based on the multimodal observation data, the modal weight of each modality, and the noise estimation model corresponding to each modality to sequentially generate the motion trajectory of the robot at each time step includes: Obtaining the motion trajectory of the current time step, wherein if the current time step is the first time step, the initial noise trajectory is used as the motion trajectory of the current time step; otherwise, the motion trajectory of the current time step is determined according to the motion trajectory of the next time step after the current time step; Determining noise information of each modality based on the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory of the current time step; Determining the total noise at the current time step according to the modal weights of the modes and the noise information of the modes; The motion trajectory of the robot at a time step before the current time step is generated according to the motion trajectory of the current time step and the total noise at the current time step.

4. The robot motion generation method according to claim 3, characterized in that: The determining of noise information of each modality according to the multimodal observation data, the noise estimation model corresponding to each modality, and the motion trajectory of the current time step includes: The observation data of the first modality and the motion trajectory of the current time step are input into the noise estimation model corresponding to the first modality to predict the noise information of the first modality, wherein the first modality is any modality in the multi-modality.

5. The robot motion generation method according to claim 3, characterized in that: Determining the total noise at the current time step according to the modal weights of the modes and the noise information of the modes includes: Calculate the product of the model weight of each mode and the corresponding noise information respectively to obtain the product result of each mode; The product results of the modes are summed to obtain the total noise at the current time step.

6. The robot motion generation method according to claim 3, characterized in that: Generating the motion trajectory of the robot at a time step before the current time step according to the motion trajectory of the current time step and the total noise at the current time step comprises: Get the gradient step coefficient, denoising intensity coefficient and additional Gaussian noise; Calculating a first product of the gradient step coefficient and the total noise at the current time step; Calculating the difference between the motion trajectory of the current time step and the first product; Calculating a second product of the difference and the denoising strength coefficient; The sum of the second product and the additional Gaussian noise is calculated as the motion trajectory of the robot in the time step before the current time step.

7. The robot motion generation method according to claim 1, characterized in that: The process of determining the initial noise trajectory includes: Get the trajectory dimension of the action trajectory sequence; According to the trajectory dimension, the initial noise trajectory is generated by sampling independently element by element from the standard Gaussian distribution.

8. The robot motion generation method according to claim 1, characterized in that: The process of determining the modal weight of each modality includes: Obtain the current task success rate and current confidence of each modality; The modality weight of each modality is calculated based on the current task success rate of each modality and the current confidence of each modality.

9. The robot motion generation method according to claim 1, characterized in that: The training process of the noise estimation model corresponding to each mode includes: Acquire a first training data set in a first modality, wherein the first training data set includes a real action sequence and observation data in the first modality; According to the preset noise scheduling coefficient, the real action sequence is iteratively processed at each time step to generate the action trajectory at each time step in sequence; According to the motion trajectory at each time step, the noise estimation model corresponding to the first modality is trained.

10. A robot control system, characterized in that: include: A processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the control system of the robot is running, the processor executes the machine-readable instructions to perform the steps of the robot motion generation method according to any one of claims 1 to 9.

Images