A robot imitation learning execution method, system and storage medium

Abstract

The application relates to a robot imitation learning execution method, a system and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring an original demonstration trajectory, and constructing an accelerated demonstration trajectory; acquiring a robot joint state, and combining the robot joint state with a robot action to construct an original trajectory frame-by-frame fusion feature; constructing an accelerated trajectory frame-by-frame fusion feature corresponding to the accelerated demonstration trajectory according to the feature; constructing a cost matrix according to the two fusion features; calculating a trajectory optimal transmission deviation score by using the cost matrix; screening all the accelerated demonstration trajectories based on the score to obtain an accelerated demonstration dataset; and performing multi-target iterative training by using the original demonstration dataset and the accelerated demonstration dataset, constantly updating shared model parameters until an imitation learning strategy model represented by the shared model parameters converges, obtaining optimal shared model parameters, and thus completing acceleration optimization deployment of the robot imitation learning. The application effectively improves the execution efficiency of a robot strategy.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a robot imitation learning execution method, system, and storage medium. Background Technology

[0002] In recent years, with the development of artificial intelligence, robotics, and sensor technology, learning control methods for robot operation tasks have become an important direction for research and industrial application. Traditional robot operation control schemes heavily rely on manually designed rule algorithms and model-driven control strategies. The core of these schemes relies on precise environmental modeling, complex parameter calibration, and pre-defined task prior knowledge to implement control logic. While these schemes can achieve basic operational results in structured, simple, and controllable scenarios, they generally suffer from poor environmental adaptability, insufficient control robustness, high scene migration costs, and weak generalization ability in complex, unstructured scenarios with object occlusion, dynamic lighting changes, cluttered backgrounds, and diverse object shapes. To reduce engineering development costs and improve system adaptability, data-driven learning methods have gradually become mainstream. Among these, end-to-end policy learning schemes, represented by imitation learning, are widely used in visual servoing operations such as grasping, placement, and assembly due to their advantages such as stable training and rapid convergence through expert demonstrations.

[0003] In the field of imitation learning, early methods were primarily based on supervised learning paradigms such as behavior cloning. Subsequent developments have incorporated algorithmic frameworks with interactive error correction to mitigate distribution offset issues, and network structures and training paradigms more suited to high-dimensional visual input and continuous control have emerged. Meanwhile, with the maturation of large-scale datasets, simulation platforms, and pre-trained visual representations, visual-motor policies have continuously improved in terms of generalization ability and task coverage. Addressing the needs of real-world robot deployment, researchers are further focusing on indicators such as the real-time performance, control frequency, and task timing of policies in actual execution. They are attempting to reduce redundant control steps, shorten the decision horizon, and improve execution efficiency while ensuring task success rates through methods such as action segmentation, temporal modeling, policy compression, and multi-objective training, thereby meeting the stability and throughput requirements of industrial and service scenarios.

[0004] Current mainstream imitation learning technologies encompass various classic implementation schemes, including supervised learning-based behavior cloning methods, optimization methods that rely on environmental interaction and error correction to expand demonstration data, and generative imitation learning methods based on distributed matching. These methods have advantages in training stability and utilization of expert priors, achieving good success rates in multi-stage tasks such as grasping, placing, and stacking. However, these methods often assume that the temporal granularity of the demonstration data is the control rhythm to be adopted during deployment, while actual demonstrations are often collected at a conservative pace, resulting in the learned strategy inheriting a "slow-paced" action sequence during deployment. Specifically, the decision-making horizon is too long, the number of control steps is redundant, and the completion time is too large, making it difficult to meet the requirements of factory and production line scenarios for cycle time, throughput, and real-time performance. In other words, while existing imitation learning strategies can "learn," they still suffer from the engineering bottleneck of "not being fast enough in execution."

[0005] To address the shortcomings of robot imitation learning strategies in terms of execution efficiency, some existing technologies attempt to reduce the burden of stepwise decision-making by improving policy structure or temporal modeling. For example, they employ policy structures that output action blocks / action sequences, or generative sequence modeling strategies to enhance the modeling capabilities for long-term tasks. Other approaches attempt to achieve faster execution by increasing control frequency, reducing repetitive actions, or pruning action sequences during deployment. These approaches can, to some extent, reduce the cumulative error of policies on long sequences or improve the completion time of specific tasks, but their effectiveness is usually still strongly constrained by the granularity of the demonstration time. Specifically, when the demonstration itself is slow, the model tends to reproduce this slow-paced trajectory; however, relying solely on heuristic pruning or forced acceleration during deployment can easily introduce discontinuous actions, abrupt state transitions, or missing key intermediate steps, leading to decreased success rate, unstable control, or safety risks. Furthermore, existing technologies also accelerate policies by reconstructing training datasets. This includes processing the original demonstration trajectories with time downsampling, temporal retiming, and trajectory compression to construct high-speed demonstration samples as training targets, or introducing reinforcement learning / reward optimization after imitation learning for post-training, aiming to further pursue shorter durations or fewer steps under the constraint of the reward function. While these methods conceptually exert direct pressure on "speed," they have significant drawbacks in engineering applications: Firstly, time downsampling / retiming can cause a shift in the distribution of demonstration trajectories, easily introducing anomalous segments that differ significantly from the original demonstration (e.g., skipped key transitions, sparsified contact processes, and disrupted action-state correspondences). Without reliable quality assessment and removal mechanisms, training can significantly sacrifice accuracy and success rate, even leading to policy jitter and failure modes. Secondly, reinforcement learning or post-training methods often rely on reward design and numerous interactions, resulting in high training costs, significant impact on stability and reproducibility from environmental noise, and safety and cost issues in real robots, making them unsuitable as a general, low-cost acceleration path.

[0006] To alleviate the conflict between "speed and accuracy", there also exist trade-off methods for multi-task / multi-objective learning in the prior art, for example, solving the update direction through fixed-weight weighted summation, dynamic weight allocation or gradient conflict-based processing, so as to achieve a compromise among multiple loss objectives. Such methods can mitigate objective conflicts at the optimization level, but they generally belong to general training frameworks: on the one hand, the objective weight or update direction may still be sensitive to data noise. When the accelerated demonstration itself contains strong distribution shifts or abnormal trajectories, it is difficult to fundamentally avoid erroneous gradients driven by bad data only by relying on weight balancing; on the other hand, existing solutions often lack a closed-loop mechanism matching the construction of accelerated demonstration data, making it difficult to simultaneously generate faster demonstrations to improve efficiency, perform interpretable evaluation and filtering on the distribution shift introduced by acceleration, and adaptively and stably seek optimization between accuracy and efficiency during training.

[0007] In conclusion, the existing robot operation imitation learning technology cannot balance operation accuracy and execution efficiency without increasing the demonstration collection cost and sacrificing the operation success rate, and is difficult to adapt to the core requirements of industrial high-takt deployment scenarios. Contents of the Invention

[0008] Therefore, the technical problem to be solved by the present invention is to overcome the problems of slow policy operation rhythm, conflict between accuracy and efficiency, and poor training stability in the prior art.

[0009] In a first aspect, to solve the above technical problem, the present invention provides a robot imitation learning execution method, comprising: S1. Acquire an original demonstration trajectory; perform time-series downsampling processing on the original demonstration trajectory using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein the original demonstration trajectory includes robot observation states and robot actions. S2. Extract robot joint states from the robot observation states, and perform normalization processing on the time sequence of the original demonstration trajectory to obtain normalized time; construct frame-by-frame fusion features of the original trajectory according to the robot joint states, the robot actions and the normalized time. S3. Construct frame-by-frame fusion features of the accelerated trajectory corresponding to the accelerated demonstration trajectory according to the frame-by-frame fusion features of the original trajectory; construct a cost matrix according to the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the accelerated trajectory. S4. Calculate the optimal transport plan of the original demonstration trajectory and the accelerated demonstration trajectory according to the cost matrix, so as to obtain an optimal transport deviation score of the trajectory; screen all the accelerated demonstration trajectories according to a preset screening threshold and the optimal transport deviation score of the trajectory, eliminate abnormal accelerated trajectories with excessive deviation, and obtain an accelerated demonstration dataset. S5. Based on the preset shared model parameters, and using the original demonstration dataset and the accelerated demonstration dataset, perform multi-objective iterative training, continuously update the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, and obtain the optimal shared model parameters. S6. Based on the optimal shared model parameters, complete the accelerated execution optimization deployment of robot imitation learning.

[0010] In one embodiment of the present invention, in step S3, a cost matrix is ​​constructed using squared Euclidean distance based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory; wherein, each element in the cost matrix satisfies the following relationship: ; in, The squared Euclidean distance cost represents the frame-by-frame fusion feature of the original trajectory and the frame-by-frame fusion feature of the accelerated trajectory. This represents the frame-by-frame fusion features of the original trajectory. This indicates the frame-by-frame fusion features of the acceleration trajectory. This represents the time downsampling factor.

[0011] In one embodiment of the present invention, in step S4, the step of calculating the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory based on the cost matrix to obtain the optimal transmission deviation score of the trajectory is as follows: Based on the cost matrix and preset discrete weights on both sides, the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory is calculated using entropy-regularized optimal transmission; wherein, the expression for the optimal transmission plan is: ; in, This represents the optimal transmission plan. This represents the transmission plan to be solved. This represents the inner product of the transport plan and the cost matrix, indicating the transport cost. Represents the discrete weights on both sides; Represents a set of transmission plans. Represents the entropy regularity coefficient. Represents the cost matrix; Based on the optimal transmission plan, calculate the optimal transmission deviation score; wherein, the expression for the optimal transmission deviation score is: ; in, The score represents the deviation from the optimal transmission trajectory. Indicates the original demonstration trajectory. Indicates an accelerated demonstration trajectory. This represents the time downsampling factor.

[0012] In one embodiment of the present invention, the expression for the accelerated demonstration dataset is: ; in, This indicates an accelerated demonstration dataset. The score represents the deviation from the optimal transmission trajectory. This represents the original demonstration dataset. Indicates an accelerated demonstration trajectory. This indicates the preset filtering threshold.

[0013] In one embodiment of the present invention, the expression for the accelerated demonstration trajectory is: ; in, Indicates an accelerated demonstration trajectory. Indicates the time-sampling factor. Indicates the first The robot's observation state at any given moment. Indicates the first The robot's movements at all times This indicates the total time series length of the accelerated demonstration trajectory.

[0014] In one embodiment of the present invention, the mathematical expression for the frame-by-frame fusion feature of the acceleration trajectory is: ; ; in, This indicates the frame-by-frame fusion features of the acceleration trajectory. Indicates the first The state of the robot's joints at any given moment. Indicates the first The robot's movements at all times Indicates the weight of the time dimension. This represents the total time series length of the original demonstration trajectory.

[0015] In one embodiment of the present invention, step S5, which involves performing multi-objective iterative training based on preset shared model parameters and using the original demonstration dataset and the accelerated demonstration dataset to continuously update the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, is as follows: Preset shared model parameters, and calculate the first imitation learning loss function and the second imitation learning loss function for the original demonstration dataset and the accelerated demonstration dataset respectively during each round of iterative training; Calculate the first gradient and the second gradient of the first imitation learning loss function and the second imitation learning loss function with respect to the shared model parameters, respectively. Stack the first gradient and the second gradient to construct a gradient matrix; construct a Gram matrix based on the gradient matrix; The positive weight vector corresponding to the Gram matrix is ​​solved using the Nash bargaining equilibrium condition; the Nash fusion loss function is constructed based on the positive weight vector. Based on the Nash fusion loss function, the shared model parameters are iteratively updated using the gradient descent method to complete the optimization of model parameters in a single iteration, and the iteration continues until the imitation learning strategy model represented by the shared model parameters converges.

[0016] In one embodiment of the present invention, the expression for the Nash fusion loss function is: ; in, This represents the Nash fusion loss function. Indicates shared model parameters. This represents the Nash coordination weight corresponding to the first objective. This represents the Nash coordination weight corresponding to the second objective. This represents the first imitation learning loss function. This represents the second imitation learning loss function.

[0017] Secondly, to solve the above-mentioned technical problems, the present invention provides a robot imitation learning execution system for implementing the above-mentioned robot imitation learning execution method, comprising: The first construction module is used to acquire the original demonstration trajectory; and to perform temporal downsampling processing on the original demonstration trajectory using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein, the original demonstration trajectory includes the robot's observation state and robot actions; The second construction module is used to extract the robot joint state from the robot observation state, and normalize the time sequence of the original demonstration trajectory to obtain normalized time; and construct the frame-by-frame fusion feature of the original trajectory based on the robot joint state, the robot action and the normalized time. The cost matrix construction module is used to construct the frame-by-frame fusion features of the acceleration demonstration trajectory corresponding to the original trajectory based on the frame-by-frame fusion features of the original trajectory; and to construct the cost matrix based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory. The filtering module is used to calculate the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory based on the cost matrix, and obtain the optimal transmission deviation score of the trajectory; and to filter all the accelerated demonstration trajectories according to the preset filtering threshold and the optimal transmission deviation score of the trajectory, and remove abnormal accelerated trajectories with excessive deviation, so as to obtain the accelerated demonstration dataset. The iterative training module is used to perform multi-objective iterative training based on preset shared model parameters and using the original demonstration dataset and the accelerated demonstration dataset, continuously updating the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, thus obtaining the optimal shared model parameters. The deployment module is used to complete the accelerated execution optimization deployment of robot imitation learning based on the optimal shared model parameters.

[0018] Thirdly, to solve the above-mentioned technical problems, the present invention provides a computer-readable storage medium storing one or more programs, which can be executed by one or more processors to implement the steps in the above-mentioned robot imitation learning execution method.

[0019] Compared with the prior art, the above-described technical solution of the present invention has the following advantages: (1) The robot imitation learning execution method, system, and storage medium described in this invention effectively solve the technical defects existing in the robot accelerated imitation learning process, such as poor downsampled trajectory quality, data distribution deviation, solidification of multi-objective training weights, and difficulty in balancing policy accuracy and execution efficiency. This invention constructs an accelerated demonstration dataset by time downsampling the original expert demonstrations and introduces a demonstration quality evaluation and filtering mechanism based on the optimal transmission distance. This mechanism does not rely on manual annotation and can measure the degree of distribution deviation between the accelerated demonstration trajectory and the original demonstration trajectory in an unsupervised manner. It accurately filters and removes abnormal trajectories and low-quality demonstration data generated during time downsampling, effectively suppresses distribution noise in the demonstration data, improves the overall quality of training data, and avoids the negative interference caused by low-quality accelerated demonstration data to robot policy learning from the data source.

[0020] (2) This invention unifies the accuracy optimization objective corresponding to the original demonstration and the efficiency optimization objective corresponding to the accelerated demonstration into a multi-objective optimization problem under a shared model parameter space, and relies on the Nash weight balancing module to adaptively solve the optimal weight ratio of the two optimization objectives. Compared with the traditional fixed-weight linear weighting optimization method, this invention can dynamically adjust the training optimization direction according to the geometric correlation of the gradients of different optimization objectives, effectively solving the problem of a single objective dominating model training for a long time, and significantly improving the stability and robustness of multi-objective joint training.

[0021] (3) This invention achieves deep integration of quality control of demonstration data layer and weight coordination of model optimization layer, decoupling the two core problems of training efficiency gain and performance degradation caused by distribution offset brought about by accelerated demonstration. This enables the robot strategy obtained by training to retain fine control capability and high task success rate, and to learn efficient action rhythm, effectively reducing redundant control steps in robot deployment process and greatly improving overall execution efficiency.

[0022] (4) The present invention is decoupled from the specific imitation learning model structure, which can be widely adapted to various behavior cloning algorithms and vision-motor strategy network structures. It has excellent versatility and scalability, and can be quickly adapted to different robot hardware platforms and diverse task scenarios. Attached Figure Description

[0023] To make the content of this invention easier to understand, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0024] Figure 1 This is a flowchart of a robot imitation learning execution method according to a preferred embodiment of the present invention; Figure 2 This is a representative simulation task example diagram in a preferred embodiment of the present invention; Figure 3 This is a representative real-world failure analysis example diagram in a preferred embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, so that those skilled in the art can better understand and implement the present invention. However, the embodiments described are not intended to limit the present invention.

[0026] Example 1: Reference Figure 1 As shown, the present invention provides a robot imitation learning execution method, including but not limited to the following steps: S1. Obtain the original demonstration trajectory; perform temporal downsampling on the original demonstration trajectory using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein, the original demonstration trajectory includes the robot's observation state and robot actions; S2. Extract robot joint states from robot observation states and normalize the timing of the original demonstration trajectory to obtain normalized time; construct frame-by-frame fusion features of the original trajectory based on robot joint states, robot actions, and normalized time. S3. Based on the frame-by-frame fusion features of the original trajectory, construct the frame-by-frame fusion features of the acceleration trajectory corresponding to the acceleration demonstration trajectory; construct the cost matrix based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory. S4. Based on the cost matrix, calculate the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory to obtain the optimal transmission deviation score of the trajectory; based on the preset screening threshold and the optimal transmission deviation score of the trajectory, screen all accelerated demonstration trajectories, remove abnormal accelerated trajectories with excessive deviation, and obtain the accelerated demonstration dataset. S5. Based on the preset shared model parameters, and using the original demonstration dataset and the accelerated demonstration dataset, perform multi-objective iterative training, continuously update the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, and obtain the optimal shared model parameters. S6. Based on the optimal shared model parameters, complete the accelerated execution optimization deployment of robot imitation learning.

[0027] This invention discloses a robot imitation learning execution method. It constructs accelerated demonstrations by temporally downsampling expert demonstrations and introduces an optimal transmission distance-based demonstration quality assessment and filtering (OTF) mechanism. This method can unsupervisedly measure the distribution deviation between the accelerated demonstration trajectory and the original demonstration trajectory without additional manual annotation, and filter out abnormal trajectories or poor-quality demonstration data caused by downsampling. This effectively reduces distribution noise in the demonstration data, improves the overall quality of training data, and avoids the negative impact of low-quality accelerated demonstrations on policy learning. Secondly, this method unifies the accuracy objective induced by the original demonstration and the efficiency objective induced by the accelerated demonstration into a multi-objective optimization problem in a shared parameter space, and adaptively solves the weight relationships between different objectives through a Nash weight balancing module. Compared with the traditional linear weighting method using fixed weights, this method can dynamically adjust the optimization direction according to the geometric relationship between the gradients of different objectives, thereby avoiding a single objective dominating the training process for a long time and improving the stability and robustness of multi-objective training. Furthermore, this invention decouples the efficiency gains from accelerated demonstration from the performance degradation caused by distribution shift by combining quality control at the demonstration data level with weight coordination at the optimization level. This allows the strategy to learn a more compact action rhythm while maintaining fine control capabilities and task success rate, thereby reducing redundant control steps during deployment and improving overall execution efficiency. In addition, the method proposed in this invention is decoupled from specific imitation learning model structures, making it applicable to various behavior cloning methods and vision-motion policy network structures. Therefore, it has good versatility and scalability, and can be easily applied to different robot platforms and task scenarios. This invention can effectively improve the execution efficiency of robot strategies and reduce the training risks caused by demonstration data distribution shift while ensuring strategy stability and control accuracy, demonstrating good engineering application value and promising prospects for widespread adoption.

[0028] This invention aims to address the common problems of insufficient execution efficiency and difficulty in maintaining stable accuracy under accelerated conditions in the real-world deployment of existing imitation learning robot operation strategies. Specifically, the technical problems to be solved by this invention include: First, existing strategies often inherit the conservative time granularity of expert demonstrations, resulting in excessively long decision-making horizons, redundant control steps, and excessively long completion times, making it difficult to meet the requirements of cycle time and real-time performance in industrial scenarios; Second, existing schemes for constructing accelerated demonstrations through time downsampling and other methods lack reliable quality assessment and screening mechanisms. Accelerated demonstrations are prone to introducing trajectory distribution deviations and abnormal segments, and directly using them for training will cause the strategy to deviate from the original expert distribution, resulting in decreased success rate or unstable control; Third, facing the inherent conflict between success rate and efficiency, existing methods mostly rely on fixed weights or manual parameter tuning for target merging, lacking an optimization mechanism that can adaptively balance accuracy and efficiency during training, making it difficult to consistently obtain better compromise results in different tasks and training stages. Therefore, the purpose of this invention is to provide an imitation learning execution acceleration method for deployment timing. By performing interpretable quality assessment and filtering on accelerated demonstrations and combining adaptive accuracy and efficiency trade-off training, the number of control steps / completion time is significantly reduced while maintaining task success rate and control stability, thereby improving the deployability and throughput efficiency of the strategy in real robot scenarios.

[0029] Specifically, MTC (Multi-Telecom) Timescale Coordination (MTC) is a general training framework for visual motion imitation learning, aiming to improve policy execution efficiency without sacrificing fine-grained control. The framework includes two modules for data selection / training mechanisms, as detailed below: First, an accelerated demonstration is constructed through temporal downsampling, and an unsupervised quality assessment is performed based on the trajectory distance of the Optimal Transport (OT) to filter out samples that deviate significantly from the original distribution, thereby reducing the distribution noise introduced by downsampling. This part is the OTF module.

[0030] Secondly, the accuracy objective of the original demonstration-induced training and the efficiency objective of the accelerated demonstration-induced training are explicitly modeled as a multi-objective optimization problem with shared parameters. Specifically, the loss and gradient corresponding to the accuracy / efficiency objectives are calculated simultaneously, and then an adaptive weight is obtained using a Nash solver, so that the training process continuously approaches the Pareto front of the speed-accuracy tradeoff. This part is the NWB module.

[0031] It should be noted that the MTC framework described in this embodiment of the invention is decoupled from specific imitation learning paradigms. Specifically, the OTF module performs quality control on accelerated demonstrations at the data level, while the NWB module dynamically balances multiple objectives at the optimization level. Therefore, it can be naturally transferred to different behavior cloning methods and sequence decision networks. Furthermore, the introduction of the OTF and NWB modules decomposes the efficiency gains from accelerated demonstrations and the performance degradation caused by potential distribution shifts into two controllable stages. These two controllable stages specifically include: the former is obtained through downsampling, and the latter is suppressed through OT filtering and Nash weight coordination. This makes it more suitable for the stability and maintainability requirements of real-world deployments and more easily mitigates the gap in sim-to-real scenarios.

[0032] Specifically, temporal downsampling can significantly shorten the decision-making time domain of the demonstration, but it may also skip critical intermediate states, thus requiring the introduction of distribution offsets in the state and action spaces. Furthermore, accelerated demonstrations exhibit significant quality variation: most trajectories remain usable after downsampling, but a small number show abrupt jumps or mismatches, causing disproportionate negative impacts during training. Therefore, in this embodiment of the invention, before using accelerated demonstrations, data screening first performs unsupervised quality assessment and filtering steps to control the distribution noise introduced by downsampling.

[0033] Furthermore, in this embodiment of the invention, accelerated demonstration quality assessment and filtering based on OT distance is preferred because: when downsampling causes changes in the time scale, simple time-by-time alignment is insufficient to characterize changes in trajectory shape and state-action distribution; while OT can measure the minimum overall rearrangement cost between two sets of points while allowing cross-time-time matching, thus more sensitively capturing local mutations and global drift caused by skipping steps. This allows the OTF module to provide stable filtering criteria without additional supervision labels. The specific steps of accelerated demonstration quality assessment and filtering based on OT distance are described in detail below.

[0034] Specifically, in step S1, the original demonstration trajectory is obtained, given a time downsampling factor. Construct the corresponding acceleration demonstration trajectory. The mathematical expression for the original demonstration trajectory is: ; in, Indicates the original demonstration trajectory. Indicates the first The robot's observation state at any given moment. Indicates the first The robot's movements at all times This represents the total temporal length of the original demonstration trajectory. In this embodiment of the invention, the time downsampling factor... The preferred value is 2.

[0035] Furthermore, the expression for the accelerated demonstration trajectory is: ; ; in, Indicates an accelerated demonstration trajectory. Indicates the first The robot's observation state at any given moment. Indicates the first The robot's movements at all times This indicates the total time series length of the accelerated demonstration trajectory. This represents the largest integer not greater than the value within the parentheses.

[0036] Specifically, in step S2, the robot observes the state. Extracting robot joint states (e.g., qpos). Define normalized time. The mathematical expression is: .

[0037] Furthermore, based on the robot's joint states Robot Actions and normalized time Construct frame-by-frame fusion features of the original trajectory Its mathematical expression is: ; in, This represents the weight of the time dimension. It is used when the task relies more heavily on strict time alignment. Only want to take the geometric deviation when measuring During the implementation process, Perform global standardization, time component After normalization, the parts are spliced ​​together.

[0038] Specifically, in step S3, based on the frame-by-frame fusion features of the original trajectory, the first frame of the downsampled trajectory is... For each point, a frame-by-frame fusion feature of the acceleration trajectory is constructed, and its mathematical expression is: ; ; in, This indicates the frame-by-frame fusion features of the acceleration trajectory. Indicates the first The state of the robot's joints at any given moment. Indicates the first The robot's movements at all times This indicates the weight of the time dimension.

[0039] Furthermore, based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the accelerated trajectory, a cost matrix is ​​constructed using squared Euclidean distance; where each element in the cost matrix satisfies the following relationship: ; in, The squared Euclidean distance cost represents the frame-by-frame fusion feature of the original trajectory and the frame-by-frame fusion feature of the accelerated trajectory. This represents the frame-by-frame fusion features of the original trajectory. This represents the frame-by-frame fusion feature of the acceleration trajectory.

[0040] Specifically, in step S4, the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory is calculated based on the cost matrix, and the specific steps for obtaining the optimal transmission deviation score of the trajectory are as follows: S410. Calculate the matching cost of the two trajectories using entropy-regularized optimal transfer (Sinkhorn). Specifically, based on the cost matrix and preset discrete weights on both sides, calculate the optimal transfer plan for the original demonstration trajectory and the accelerated demonstration trajectory using entropy-regularized optimal transfer; where the expression for the optimal transfer plan is: ; in, This represents the optimal transmission plan. This represents the transmission plan to be solved. This represents the inner product of the transport plan and the cost matrix, indicating the transport cost. The weights on both sides are discrete, and in this embodiment of the invention, uniform weights are preferred. Let represent the set of non-negative transmission plans that satisfy the edge constraints. Represents the entropy regularity coefficient. This represents the cost matrix.

[0041] S420. Based on the optimal transmission plan, calculate the optimal transmission deviation score (i.e., OT deviation score); whereby the expression for the optimal transmission deviation score is: ; in, The score represents the deviation from the optimal transmission trajectory. Indicates the original demonstration trajectory. This indicates an accelerated demonstration trajectory.

[0042] S430. Based on the optimal transmission deviation score, a fixed filtering threshold is used. The accelerated demonstration trajectories are filtered to obtain the filtered accelerated demonstration dataset, whose expression is: ; in, This indicates an accelerated demonstration dataset. The score represents the deviation from the optimal transmission trajectory. This represents the original demonstration dataset. Indicates an accelerated demonstration trajectory. This indicates the preset filtering threshold.

[0043] Specifically, after completing the OT-based filtering, this embodiment of the invention uses two types of data (i.e., the original demonstration dataset). Compared with the filtered accelerated demonstration dataset Training is performed on two datasets, one for accuracy and the other for efficiency, which often conflict within a shared parameter space. Using a linear weighted average with fixed weights can easily lead to one objective dominating due to differences in loss scale or gradient magnitude, resulting in unstable trade-offs between different tasks and training stages. Therefore, this invention formulates training as a multi-objective optimization problem and employs a Nash / Pareto solver for adaptive balancing. Multi-objective training is viewed as a "bargaining" process to obtain a more stable compromise update direction. This bargaining criterion favors directions that improve proportionally and is more robust to scale changes. The specific steps of multi-objective collaborative training based on Nash–Pareto solutions are described in detail below.

[0044] Specifically, in step S5, based on the preset shared model parameters, and using the original demonstration dataset and the accelerated demonstration dataset for multi-objective iterative training, the shared model parameters are continuously updated until the imitation learning strategy model represented by the shared model parameters converges. The specific steps are as follows: S510, Preset Shared Model Parameters In each iteration, the imitation learning loss is calculated on both classes of data. Specifically, during each round of training iteration, the first and second imitation learning loss functions are calculated for the original demonstration dataset and the accelerated demonstration dataset, respectively. The expressions for the first and second imitation learning loss functions are: ; ; in, This represents the first imitation learning loss function. This represents the second imitation learning loss function. This indicates the imitation learning strategy model in the robot's observed state. Predicted actions and demonstrated actions The loss function values ​​between Indicates the robot's observation state With robot movement Expectation calculation under the sample distribution.

[0045] S520. Calculate the first gradient and the second gradient of the first imitation learning loss function with respect to the shared model parameters, respectively. The expressions for the first gradient and the second gradient are: ; ; in, Indicates the first gradient. Indicates the second gradient. Indicates shared model parameters The gradient operator.

[0046] S530. Stack the task gradients into a gradient matrix. Specifically, stack the first and second gradients to construct a gradient matrix. Construct the Gram matrix based on the gradient matrix. The expression for the Gram matrix is: ; in, Represents the Gram matrix. The symbol for transpose of a matrix.

[0047] The S540 and Nash bargaining solutions induce a positive weight vector. Specifically, the positive weight vector corresponding to the Gram matrix is ​​obtained using the Nash bargaining equilibrium condition. ,in This represents a set of vectors in the two-dimensional real number field where all elements are positive. Positive weight vectors. The relationship that satisfies this is: (Take the reciprocal of each element).

[0048] Furthermore, based on the given positive weight vector We construct a scalarized objective, namely the Nash fusion loss function; the expression for the Nash fusion loss function is: ; in, This represents the Nash fusion loss function. Indicates shared model parameters. This represents the Nash coordination weight corresponding to the first objective. This represents the Nash coordination weight corresponding to the second objective.

[0049] S550, based on Nash fusion loss function The gradient descent method is used to iteratively update the shared model parameters, completing the optimization of the model parameters in a single iteration, and then iterating repeatedly until the imitation learning strategy model represented by the shared model parameters converges.

[0050] Furthermore, gradient descent is used to update the shared model parameters. The expression is: ; in, This represents the learning rate for gradient descent.

[0051] The strategy described in this invention can dynamically coordinate the contributions of the original demonstration and the accelerated demonstration without repeated manual parameter tuning, maintaining a stable trade-off in training across different tasks and stages, and achieving more consistent performance improvements with minimal additional computational overhead. Subsequent experiments will verify the generalization effect of this adaptive trade-off in multi-task scenarios under different simulation environments and compare different strategies, and provide corresponding ablation experiments and analysis. Details are as follows: In ablation experiments, the role of the OT filtering mechanism and adaptive multi-objective coordination was verified in a typical mixed-timescale training framework. Table 1 summarizes the success rate ablation results of the complete MTC pipeline and four representative variants. Four control variants were set up in this experiment: removing the OT filter, removing the adaptive solver, training only with downsampled examples filtered by the OT filter, and training only with downsampled examples. Experimental results show that the complete method achieves the best performance on both the Action Chunking with Transformers (ACT) and Diffusion Policy (DP) backbone policies. Removing the OT filter results in a moderate performance drop (ACT baseline success rate drops from 51% to 47%, and DP baseline from 62% to 57%), indicating that quality filtering can reduce harmful distribution biases in accelerated examples. In contrast, replacing adaptive coordination with naive mixed training has a more severe impact, with success rates dropping to 12% and 14%, respectively. This indicates that multi-timescale supervision must be coordinated and cannot be directly merged. For the two control groups using only downsampled samples, the scheme introducing OT filtering outperformed the scheme without filtering, but both were significantly worse than the complete MTC pipeline. This phenomenon proves that relying solely on accelerated downsampled demo data is insufficient to support stable training; the shared supervision signal provided by the original complete demo is the key support for model performance.

[0052] Table 1:

[0053] Furthermore, referring to Figure 2As shown in the figure, this diagram illustrates representative scenarios from the evaluated static and dynamic simulation benchmarks. In particular, the two Leju examples on the right demonstrate the two-stage coordination and time-sensitive object transfer process required in the dynamic benchmark. In this experiment, representative scenarios from the simulation benchmarks used for evaluation included the RoboTwin static task, the ManiSkill static task, and the Leju dynamic task. In the Leju setup, Task 1 corresponds to moving target grasping: objects on the conveyor belt are first transferred to an intermediate area before reaching the final target area; Task 2 corresponds to moving object weighing: objects are first placed on a scale, weighed, and then transferred to the target area.

[0054] Furthermore, referring to Table 2, a compact backbone-level summary is provided to supplement the quantitative summary results of control steps and power in the static baseline above. Compared to the Origin baseline, MTC improves the average success rate of ACT from 45% to 55% and the average success rate of DP from 49% to 63%; at the same time, the average number of control steps for ACT decreases from 213 to 130, and the average number of control steps for DP decreases from 210 to 145, corresponding to speedup ratios of 1.64× and 1.45×, respectively. These gains indicate that MTC does not simply trade shorter trajectories for a decrease in accuracy.

[0055] Table 2:

[0056] Furthermore, referring to Figure 3 As shown in the figure, this illustrates representative failure cases. The main failure modes include contact mismatch caused by the gap between the training and real-world environment, and surface slippage after the object is lifted. Figure 3 (a) shows the grasping effect after the simulation domain is migrated to the real scene. There is a deviation between the support conditions of the simulation and the real scene, which causes the object contact area to shift and the stability of the lifting process to drop significantly. Figure 3 (b) is an example of unstable gripping on a curved surface. After the object is lifted off the support surface, the curved surface contact interface is prone to relative sliding and rotation, which eventually leads to gripping failure.

[0057] Furthermore, the method of the present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments. This embodiment is based on a robot imitation learning task, and verifies the effectiveness of the proposed MTC training framework in the RoboTwin simulation environment. The experimental process includes steps such as environment configuration, training data preparation, policy training, and policy evaluation. The specific content is as follows: First, install the Vulkan graphical runtime environment on a Linux system (such as Ubuntu 20.04 or Ubuntu 22.04) to ensure that the simulation environment runs normally.

[0058] Then, Conda was used to create a Python runtime environment and install the relevant dependency libraries, including: conda create -n MTC python=3.10 -y; conda activate MTC.

[0059] Furthermore, the experiment requires downloading relevant model and data resources. If you encounter Hugging Face access rate limitations, you can log in to your Hugging Face account using the following command.

[0060] Next, navigate to the policy training directory and install the necessary dependency libraries for training: cd policy / ACT; pip install pyquaternion pyyaml ​​rospkg pexpect mujoco==2.3.7 dm_control==1.0.14 opencv-python matplotlib einops packaging h5py ipython.

[0061] Next, enter the DETR submodule and install the relevant components, while downloading the resource files required for training: cd detr&&pip install -e .&&cd ..; bash script / _download_assets.sh.

[0062] After completing the environment configuration, the raw RoboTwin 2.0 data needs to be preprocessed to convert it into the data format required for ACT model training. The data preprocessing process is performed using the `process_data.sh` script, where the parameter `expert_data_num` represents the number of expert demonstration trajectories used for training.

[0063] This step involves parsing the original demonstration trajectory and generating a dataset that conforms to the training format. Simultaneously, in the method of this invention, the accelerated demonstration trajectory obtained through time downsampling is subjected to quality assessment and filtering based on the optimal transmission distance to obtain filtered training data.

[0064] After data preparation is complete, the policy training process begins. The training script train.sh will start training based on the specified task, data configuration, and random seed. Under default configuration, the model typically trains for approximately 6000 steps.

[0065] During training, the method of this invention uses both original demonstration data and accelerated demonstration data for joint training, and dynamically adjusts the weights between the accuracy target and the efficiency target through the Nash weight balancing mechanism, thereby improving the policy execution efficiency while maintaining control accuracy.

[0066] After training, the learned policy needs to be evaluated in a simulation environment. The evaluation script eval.sh is used to load the trained policy model and execute the test in the specified environment. The specific script is as follows: bash eval.sh ${task_name} ${task_config} ${ckpt_setting} ${expert_data_num} ${seed} ${gpu_id}; Here, `task_config` represents the evaluation environment configuration, while `ckpt_setting` represents the training data configuration used during the policy training phase. By statistically analyzing metrics such as task success rate, trajectory execution steps, and policy stability, it can be verified that the method of this invention effectively shortens the policy execution time domain while maintaining task success rate, thereby improving the robot's task execution efficiency.

[0067] The RoboTwin simulation platform allows for the selection of various robot operation tasks as experimental subjects, such as object-tapping tasks, object-grabbing tasks, and assembly tasks. Taking the beat_block_hammer task as an example, the robot needs to use a robotic arm to tap a target object. Experimental results show that, compared with traditional imitation learning methods, the training framework proposed in this invention can significantly reduce the number of control steps required for policy execution while maintaining the task success rate, thereby achieving more efficient policy execution.

[0068] In summary, the above embodiments demonstrate that the imitation learning-based accelerated execution method proposed in this invention, based on optimal transmission filtering and Nash weight balancing, can operate stably in various robot simulation tasks and significantly improve policy execution efficiency while ensuring control accuracy.

[0069] The robot imitation learning execution method described in this embodiment of the invention has the following advantages: First, this invention proposes an accelerated execution strategy training method for visual-motor imitation learning. By constructing accelerated demonstrations through time downsampling of expert demonstrations, and simultaneously using the original demonstrations and accelerated demonstrations for joint learning during the training process, the strategy learns a more compact movement rhythm while maintaining its original fine control capabilities. This reduces control steps and improves task execution efficiency during actual deployment.

[0070] Secondly, this invention proposes a demonstration data filtering scheme based on optimal transport. By constructing feature representations between the original demonstration trajectory and the accelerated demonstration trajectory, and calculating the optimal transport distance between them, unsupervised quality assessment and filtering of the accelerated demonstration are performed. This eliminates abnormal trajectories or data with large distribution shifts caused by time downsampling, thereby improving the quality of training data and the stability of policy learning.

[0071] Third, this invention proposes a multi-objective optimization module based on Nash balancing. The accuracy objective corresponding to the original demonstration and the efficiency objective corresponding to the accelerated demonstration are modeled as a unified multi-objective optimization problem. The weights are adaptively calculated based on the relationship between the gradients of different objectives using a Nash weighting mechanism, thereby dynamically coordinating the trade-off between accuracy and efficiency during training.

[0072] Fourth, the method described in this invention has good versatility and practicality. The proposed training framework is decoupled from the specific imitation learning model structure, making it applicable to various behavior cloning algorithms and vision-motion policy networks. It can be stably applied in different robot operation tasks and simulation environments, and has good engineering promotion value.

[0073] Example 2: Based on the same inventive concept, this embodiment provides a robot imitation learning execution system. The principle of solving the problem is similar to that of the robot imitation learning execution method provided in Embodiment 1, and the repeated parts will not be described again.

[0074] This embodiment provides a robot imitation learning execution system for implementing the robot imitation learning execution method described in Embodiment 1, including: The first construction module is used to acquire the original demonstration trajectory; the original demonstration trajectory is processed by temporal downsampling using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein, the original demonstration trajectory includes the robot's observation state and robot actions; The second construction module is used to extract the robot joint state from the robot's observation state and normalize the time sequence of the original demonstration trajectory to obtain the normalized time; based on the robot joint state, robot action and normalized time, the original trajectory frame-by-frame fusion features are constructed. The cost matrix construction module is used to construct the frame-by-frame fusion features of the accelerated demonstration trajectory based on the frame-by-frame fusion features of the original trajectory; and to construct the cost matrix based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the accelerated trajectory. The filtering module is used to calculate the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory based on the cost matrix, and obtain the optimal transmission deviation score of the trajectory; based on the preset filtering threshold and the optimal transmission deviation score of the trajectory, all accelerated demonstration trajectories are filtered to remove abnormal accelerated trajectories with excessive deviation, and the accelerated demonstration dataset is obtained. The iterative training module is used to perform multi-objective iterative training based on preset shared model parameters and using the original demonstration dataset and accelerated demonstration dataset. It continuously updates the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, thus obtaining the optimal shared model parameters. The deployment module is used to accelerate the execution and optimize the deployment of robot imitation learning based on the optimal shared model parameters.

[0075] Example 3: This embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the robot imitation learning execution method as described in Embodiment 1.

[0076] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0077] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0078] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0079] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0080] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A robot imitation learning execution method, characterized in that, include: S1. Obtain the original demonstration trajectory; perform temporal downsampling processing on the original demonstration trajectory using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein, the original demonstration trajectory includes the robot's observation state and robot actions; S2. Extract the robot joint state from the robot observation state, and normalize the timing of the original demonstration trajectory to obtain normalized time; construct frame-by-frame fusion features of the original trajectory based on the robot joint state, the robot action, and the normalized time. S3. Based on the frame-by-frame fusion features of the original trajectory, construct the frame-by-frame fusion features of the acceleration trajectory corresponding to the acceleration demonstration trajectory; construct the cost matrix based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory. S4. Based on the cost matrix, calculate the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory to obtain the optimal transmission deviation score of the trajectory; filter all the accelerated demonstration trajectories according to the preset screening threshold and the optimal transmission deviation score of the trajectory, and remove abnormal accelerated trajectories with excessive deviation to obtain the accelerated demonstration dataset. S5. Based on the preset shared model parameters, and using the original demonstration dataset and the accelerated demonstration dataset, perform multi-objective iterative training, continuously update the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, and obtain the optimal shared model parameters. S6. Based on the optimal shared model parameters, complete the accelerated execution optimization deployment of robot imitation learning.

2. The robot imitation learning execution method according to claim 1, characterized in that, In step S3, a cost matrix is ​​constructed using squared Euclidean distance based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory; wherein, each element in the cost matrix satisfies the following relationship: ； in, The squared Euclidean distance cost represents the frame-by-frame fusion feature of the original trajectory and the frame-by-frame fusion feature of the accelerated trajectory. This represents the frame-by-frame fusion features of the original trajectory. This indicates the frame-by-frame fusion features of the acceleration trajectory. This represents the time downsampling factor.

3. The robot imitation learning execution method according to claim 1, characterized in that, In step S4, the step of calculating the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory based on the cost matrix, and obtaining the optimal transmission deviation score of the trajectory, is as follows: Based on the cost matrix and preset discrete weights on both sides, the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory is calculated using entropy-regularized optimal transmission; wherein, the expression for the optimal transmission plan is: ； in, This represents the optimal transmission plan. This represents the transmission plan to be solved. This represents the inner product of the transport plan and the cost matrix, indicating the transport cost. Represents the discrete weights on both sides; Represents a set of transmission plans. Represents the entropy regularity coefficient. Represents the cost matrix; Based on the optimal transmission plan, calculate the optimal transmission deviation score; wherein, the expression for the optimal transmission deviation score is: ； in, The score represents the deviation from the optimal transmission trajectory. Indicates the original demonstration trajectory. Indicates an accelerated demonstration trajectory. This represents the time downsampling factor.

4. The robot imitation learning execution method according to claim 1, characterized in that, The expression for the accelerated demonstration dataset is: ； in, This indicates an accelerated demonstration dataset. The score represents the deviation from the optimal transmission trajectory. This represents the original demonstration dataset. Indicates an accelerated demonstration trajectory. This indicates the preset filtering threshold.

5. The robot imitation learning execution method according to claim 1, characterized in that, The expression for the accelerated demonstration trajectory is: ; in, Indicates an accelerated demonstration trajectory. Indicates the time-sampling factor. Indicates the first The robot's observation state at any given moment. Indicates the first The robot's movements at all times This indicates the total time series length of the accelerated demonstration trajectory.

6. The robot imitation learning execution method according to claim 1, characterized in that, The mathematical expression for the frame-by-frame fusion feature of the acceleration trajectory is: ； ； in, This indicates the frame-by-frame fusion features of the acceleration trajectory. Indicates the first The state of the robot's joints at any given moment. Indicates the first The robot's movements at all times Indicates the weight of the time dimension. This represents the total time series length of the original demonstration trajectory.

7. The robot imitation learning execution method according to claim 1, characterized in that, In step S5, the step of performing multi-objective iterative training based on preset shared model parameters and using the original demonstration dataset and the accelerated demonstration dataset to continuously update the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges is as follows: Preset shared model parameters, and calculate the first imitation learning loss function and the second imitation learning loss function for the original demonstration dataset and the accelerated demonstration dataset respectively during each round of iterative training; Calculate the first gradient and the second gradient of the first imitation learning loss function and the second imitation learning loss function with respect to the shared model parameters, respectively; Stack the first gradient and the second gradient to construct a gradient matrix; Construct the Gram matrix based on the gradient matrix; The positive weight vector corresponding to the Gram matrix is ​​solved using the Nash bargaining equilibrium condition; the Nash fusion loss function is constructed based on the positive weight vector. Based on the Nash fusion loss function, the shared model parameters are iteratively updated using the gradient descent method to complete the optimization of model parameters in a single iteration, and the iteration continues until the imitation learning strategy model represented by the shared model parameters converges.

8. The robot imitation learning execution method according to claim 7, characterized in that, The expression for the Nash fusion loss function is: ； in, This represents the Nash fusion loss function. Indicates shared model parameters. This represents the Nash coordination weight corresponding to the first objective. This represents the Nash coordination weight corresponding to the second objective. This represents the first imitation learning loss function. This represents the second imitation learning loss function.

9. A robot imitation learning execution system, used to implement the robot imitation learning execution method according to any one of claims 1 to 8, characterized in that, include: The first construction module is used to acquire the original demonstration trajectory; and to perform temporal downsampling processing on the original demonstration trajectory using a preset time downsampling factor to construct an accelerated demonstration trajectory; wherein, the original demonstration trajectory includes the robot's observation state and robot actions; The second construction module is used to extract the robot joint state from the robot observation state, and normalize the time sequence of the original demonstration trajectory to obtain normalized time; and construct the frame-by-frame fusion feature of the original trajectory based on the robot joint state, the robot action and the normalized time. The cost matrix construction module is used to construct the frame-by-frame fusion features of the acceleration demonstration trajectory corresponding to the original trajectory based on the frame-by-frame fusion features of the original trajectory; and to construct the cost matrix based on the frame-by-frame fusion features of the original trajectory and the frame-by-frame fusion features of the acceleration trajectory. The filtering module is used to calculate the optimal transmission plan for the original demonstration trajectory and the accelerated demonstration trajectory based on the cost matrix, and obtain the optimal transmission deviation score of the trajectory; and to filter all the accelerated demonstration trajectories according to the preset filtering threshold and the optimal transmission deviation score of the trajectory, and remove abnormal accelerated trajectories with excessive deviation, so as to obtain the accelerated demonstration dataset. The iterative training module is used to perform multi-objective iterative training based on preset shared model parameters and using the original demonstration dataset and the accelerated demonstration dataset, continuously updating the shared model parameters until the imitation learning strategy model represented by the shared model parameters converges, thus obtaining the optimal shared model parameters. The deployment module is used to complete the accelerated execution optimization deployment of robot imitation learning based on the optimal shared model parameters.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the robot imitation learning execution method as described in any one of claims 1 to 8.

Images