Robot interaction control method and system based on scene

Abstract

The invention discloses a scene-based robot interaction control method and system, and relates to the technical field of robot control, and the method comprises the steps: constructing a universal interaction data set; training a universal interaction control model of the robot; continuously collecting real-time interaction data of the robot in an actual scene; constructing a scene targeted interaction data sample set; the intensified training robot intensifies an interaction control model aiming at a scene in an actual deployment scene and stores the interaction control model locally; when the robot needs to execute a new interaction task in the target application scene, a scene enhancement interaction control model is called; and a final control instruction is generated to drive the robot to execute actions. The method has the advantages that the robot has the continuous self-optimization capability through the combination of the pre-trained universal interaction control model covering multiple scenes and the real-time scene targeted intensified training, and the long-term interaction performance and decision accuracy in a complex dynamic environment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, in particular to a robot interaction control method and system based on scenes. BACKGROUND

[0002] In the field of robot interaction control, the existing technology usually adopts two schemes: one is to independently train a special model for each specific scene, which can guarantee the performance of the scene, but requires prepositioning a large number of scene models and cannot adapt to unforeseen scenes, resulting in high deployment cost and poor flexibility; the other is to rely on a single general model to cover multiple scenes, which has basic generalization ability, but performs poorly in specific scenes such as dynamically changing environments or high-precision tasks, and requires frequent manual parameter adjustment or retraining, making it difficult to achieve true adaptive learning. For example, although the traditional continuous learning method supports model updating, it has the problem of catastrophic forgetting - new scene data will cover old knowledge, causing the model to degrade in performance in historical scenes. In addition, there is a lack of dynamic management mechanism for training data, and old samples and noise samples continue to affect the model, restricting the adaptability optimization of the robot in long-term operation.

[0003] To solve the above problems, the industry has explored a method combining transfer learning and incremental learning, but there are still limitations: first, the pre-trained general model often weakens the generalization ability for low-frequency scenes due to data distribution deviation, such as the dominance of high-frequency scene data; second, real-time incremental training lacks an efficient data update strategy, and simple sample replacement can easily lose key experience or introduce noise, and frequent global retraining leads to excessive computational overhead. The current solution has not yet achieved the collaborative mechanism of "lightweight dynamic data management + phased adaptive training", and there is an urgent need for a robust interaction control framework that can balance scene individual optimization and system resource constraints SUMMARY

[0004] To solve the above technical problems, a robot interaction control method and system based on scenes are provided, which solve at least one of the technical problems mentioned in the background art.

[0005] To achieve the above purposes, the technical solution adopted by the present application is: A robot interaction control method based on scenes, comprising: Based on at least one target application scene of the robot, a robot interaction dataset covering all different target application scenes is constructed, denoted as a general-purpose interaction dataset; Training a general-purpose interaction control model for the robot based on the general-purpose interaction dataset; After the robot is deployed and runs, real-time interaction data of the robot in actual scenes is continuously collected during actual operation; Based on the real-time interaction data of the robot in the actual scene, a scene-specific interaction data sample set is constructed; based on the general-purpose interaction control model of the robot and the scene-specific interaction data samples, the robot is reinforced to train a scene-reinforced interaction control model for the actual deployment scene of the robot, and the scene-reinforced interaction control model is stored locally; When the robot needs to perform a new interaction task in the target application scene, the scene-reinforced interaction control model that is stored locally and is updated after reinforcement training is called; The new interaction instruction and the current environment perception information are input into the scene-reinforced interaction control model to generate a final control instruction to drive the robot to perform an action.

[0006] Preferably, the scene-specific interaction data sample set is constructed based on real-time interaction data of the robot in the actual scene, and specifically includes: A fixed-capacity circular data buffer is initialized to store interaction training data samples for the target application scene; During actual operation of the target application scene, real-time interaction data samples conforming to the characteristics of the scene are continuously collected; The newly collected real-time interaction data samples are added to the circular data buffer using a preset forgetting mechanism: The real-time interaction data samples stored in the circular data buffer are combined into a scene-specific interaction data sample set.

[0007] Preferably, the forgetting mechanism specifically includes: If the circular data buffer is not full, the new interaction data sample is directly stored; If the circular data buffer is full, at least one old interaction data sample in the buffer is removed according to a preset rule, and then the new interaction data sample is added, so that the total amount of data in the buffer always remains at a preset fixed capacity; The preset rule specifically includes: Based on the real-time interaction data samples stored in the circular data buffer, the real-time interaction data samples are classified, and based on the total number of types of the real-time interaction data samples, a type contribution value is added to each type; Based on the type contribution value of each type, the number of real-time interaction data samples under each type, and the storage time of each real-time interaction data sample, the importance of each real-time interaction data sample is analyzed; The real-time interaction data sample with the lowest importance is removed, and the new interaction data sample is added.

[0008] Preferably, the robot is reinforced to train a scene-reinforced interaction control model for the actual deployment scene of the robot based on the general-purpose interaction control model of the robot and the scene-specific interaction data samples, and specifically includes: A cumulative threshold is set, and when the real-time interaction data in the scene-specific interaction data sample set is replaced to reach the cumulative threshold, the small update mechanism is triggered; When the real-time interaction data in the scene-specific interaction data sample set is completely replaced compared with the scene-specific interaction data sample set after the last large update, the large update mechanism is triggered; The small update mechanism is as follows: taking the scene-specific interaction control model after the last small update as an initial model, and based on the initial model, the real-time interaction data in the scene-specific interaction data sample set is combined for reinforcement training to obtain a scene-specific reinforcement interaction control model. The large update mechanism is as follows: taking the scene-specific reinforcement interaction control model after the last large update as a first initial model, and based on the first initial model, the scene-specific interaction data sample set is combined for reinforcement training to obtain a first scene-specific reinforcement interaction control model. Taking the scene-specific reinforcement interaction control model after the last small update as a second initial model, and based on the initial model, the real-time interaction data in the scene-specific interaction data sample set is combined for reinforcement training to obtain a second scene-specific reinforcement interaction control model. The performances of the first scene-specific reinforcement interaction control model and the second scene-specific reinforcement interaction control model on the scene-specific interaction data sample set are verified respectively, and the one with a smaller loss value is selected as the scene-specific reinforcement interaction control model.

[0009] Preferably, the cumulative threshold is set to 1 / 4-1 / 8 of the fixed capacity of the cyclic data buffer.

[0010] Preferably, the at least one target application scenario of the robot comprises: Based on the actual deployment data of the robot, the application proportion of the robot actually deployed to various target application scenarios is determined. According to the application proportion of the robot actually deployed to various target application scenarios, the robot interaction data of each target application scenario is extracted to form a general-purpose interaction data set, so that the quantity proportion of the robot interaction data of each target application scenario in the general-purpose interaction data set is the same as the application proportion of the robot actually deployed to various target application scenarios.

[0011] Preferably, the training of the general-purpose interaction control model of the robot based on the general-purpose interaction data set comprises: Based on the quantity of the robot interaction data of each target scene in the general-purpose interaction data set, the robot interaction data of the target scene with a small quantity is introduced into the cost-sensitive learning to train the general-purpose interaction control model of the robot.

[0012] Further, a scene-based robot interaction control system is proposed for implementing the scene-based robot interaction control method as described above, comprising: a general-purpose dataset construction module for constructing a general-purpose interaction dataset covering all different target application scenarios based on at least one target application scenario of the robot; a general-purpose model training module for training a general-purpose interaction control model of the robot based on the general-purpose interaction dataset; a real-time data collection module deployed at the robot end for continuously collecting real-time interaction data of the robot during actual operation of the robot; a scene data buffer module including a fixed-capacity circular data buffer for dynamically updating and storing scene-specific interaction data sample sets through a preset forgetting mechanism; a scene reinforcement training module for executing reinforcement training based on the general-purpose interaction control model and the scene-specific interaction data sample sets to generate a scene-reinforced interaction control model; a model storage and calling module for storing the trained scene-reinforced interaction control model locally on the robot and calling the model when performing a task; an interaction control execution module for inputting user instructions and environmental perception information into the scene-reinforced interaction control model to generate control instructions to drive robot actions.

[0013] Optionally, the scene reinforcement training module comprises: a two-stage update triggering unit for triggering training according to the update degree of the scene-specific data sample sets: when the sample replacement amount reaches 1 / 4-1 / 8 of the fixed-capacity buffer capacity, a small update mechanism is triggered, and when the samples are completely replaced, a large update mechanism is triggered; a model iteration optimization unit for performing: in the small update mechanism, incremental training based on the previous scene-reinforced model and the new samples, in the large update mechanism, full-sample training based on the previous large update model, and loss comparison with the small update model, and the better one is saved as the final scene-reinforced model.

[0014] Optionally, the general-purpose dataset construction module comprises: a scene proportion analysis unit for counting the proportion of the robot actually deployed to each target application scenario; a data weighted sampling unit for extracting interaction data from each scene according to the proportion to construct a general-purpose interaction dataset with consistent number proportion and actual deployment proportion; a cost-sensitive training unit for introducing a cost-sensitive learning algorithm to the data-scarce target scenarios in the general-purpose model training to improve the generalization ability of the model to the rare scenarios.

[0015] Compared with the prior art, the application has the beneficial effects that: The application effectively solves the contradiction between the insufficient cross-scene migration ability of the robot and the adaptability to specific scenes by combining the pre-training of the general-purpose interactive control model covering multiple scenes and the real-time scene-specific reinforcement training; innovatively uses a fixed-capacity circular buffer and a dynamic forgetting mechanism to manage real-time interactive data, combines a phased model updating strategy, significantly suppresses the risk of catastrophic forgetting while ensuring the optimal fitting of the model to the current scene; by constructing a weighted data set and a cost-sensitive learning mechanism according to the actual deployment ratio, the data skew problem is overcome from the source, so that the robot has the ability of continuous self-optimization, and the long-term interactive performance and decision accuracy in a complex dynamic environment are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the scene-based robot interactive control method for embodiment one is shown in the figure; Figure 2 The flowchart of the method for constructing scene-specific interactive data sample set for embodiment two is shown in the figure; Figure 3 The flowchart of the method for constructing a robot interactive data set covering all different target application scenes for embodiment three is shown in the figure. DETAILED DESCRIPTION

[0017] The following description is used to disclose the application so that those skilled in the art can implement the application. The preferred embodiments in the following description are only as examples, and other obvious variants can be thought of by those skilled in the art.

[0018] Embodiment one:

[0019] Referring to Figure 1 The embodiment proposes a scene-based robot interactive control method, which includes: Based on at least one target application scene of the robot, a robot interactive data set covering all different target application scenes is constructed, denoted as a general-purpose interactive data set. By aggregating multiple scene data, a unified general-purpose data set is constructed to avoid single scene data bias from the source, providing a cross-scene common feature learning basis for the model, and significantly improving the initial generalization ability of the pre-trained model The general-purpose interactive control model of the robot is trained based on the general-purpose interactive data set. The trained general-purpose model has basic adaptability to unknown scenes, reduces the cold start time of the robot deployment, and provides high starting point parameters for subsequent scene optimization, greatly reducing the convergence difficulty of the scene-specific model; After the robot is deployed and runs, real-time interaction data of the robot in the actual scene is continuously collected during actual running, real scene dynamic data is collected, unique noise and interaction mode of the deployment environment are directly captured, a domain gap problem between simulation data and real environment is solved, and high-quality input is provided for model customization optimization; Based on the real-time interaction data of the robot in the actual scene, a scene-specific interaction data sample set is constructed. Based on the general interaction control model of the robot and the scene-specific interaction data sample, a scene reinforcement interaction control model of the robot for the actual deployment scene is trained, and is stored locally. When the robot needs to perform a new interaction task in a target application scene, the scene reinforcement interaction control model updated by the reinforcement training stored locally is called, the scene-specific model optimized locally is directly called, cloud communication delay is eliminated and offline execution reliability is guaranteed, and the scene environment changes are dynamically adapted. The new interaction instruction and the current environment perception information are input into the scene reinforcement interaction control model to generate the final control instruction to drive the robot to perform actions, accurate control instructions are generated in real time under the driving of the specific model, the task execution success rate and response speed are significantly improved, and the anti-interference ability and human-machine cooperation adaptability are stronger in complex dynamic scenes.

[0020] Embodiment two:

[0021] Referring to Figure 2 Based on the embodiment one, the embodiment further proposes constructing a scene-specific interaction data sample set based on real-time interaction data of the robot in the actual scene, which specifically includes: An initialized fixed-capacity circular data buffer is used to store interaction training data samples for the target application scene; During actual running of the target application scene, real-time interaction data samples conforming to the characteristics of the scene are continuously collected; The newly collected real-time interaction data samples are added to the circular data buffer by using a preset forgetting mechanism: The real-time interaction data samples stored in the circular data buffer are combined into a scene-specific interaction data sample set; Specifically, the forgetting mechanism specifically includes: If the circular data buffer is not full, the new interaction data sample is directly stored; If the circular data buffer is full, at least one old interaction data sample in the buffer is removed according to a preset rule, and then the new interaction data sample is added, so that the total amount of data in the buffer always remains at a preset fixed capacity. The preset rule specifically includes: Type classification is performed based on real-time interaction data samples stored in the cyclic data buffer, and a type contribution value is added to each type based on the total number of types of the real-time interaction data samples; Based on the type contribution value of each type, the number of real-time interaction data samples under each type, and the storage time of each real-time interaction data sample, the importance of each real-time interaction data sample is analyzed. The calculation formula of importance is: ; In the formula, is the importance of the ith real-time interaction data sample, N is the total number of types of real-time interaction data samples stored in the cyclic data buffer, M is the number of real-time interaction data samples under the type corresponding to the ith real-time interaction data sample, is the storage time of the ith real-time interaction data sample, and e is a natural constant. The real-time interaction data sample with the lowest importance is removed, and a new interaction data sample is added.

[0022] The intelligent forgetting mechanism proposed in this embodiment realizes dynamic optimization management of scene interaction data under the fixed capacity limit by fusing the quantitative evaluation framework of interaction type balance protection, time decay effect and intra-class sample density adaptive adjustment. The mechanism actively identifies and retains diverse interaction samples with long-term value for model evolution, especially strengthens the representativeness of low-frequency key scene patterns, while accurately eliminates obsolete or redundant data; the mathematical decision model coupled with multi-dimensional weights significantly improves the interpretability and stability of data purification, so that the buffer always maintains a high information density of scene state distribution. This provides a high-quality sample set with timeliness and generalization potential for subsequent reinforcement training, suppresses the risk of catastrophic forgetting from the data source, endows the robot with continuous learning ability and lifelong adaptability in dynamic environment, and finally achieves the coordinated leap of interaction accuracy and scene robustness.

[0023] Embodiment three:

[0024] Based on the real-time interaction data of the robot in the actual scene, the embodiment further proposes to construct a scene-specific interaction data sample set, which specifically includes: Based on the general-purpose interaction control model of the robot and the scene-specific interaction data sample, the reinforcement training robot performs a scene reinforcement interaction control model specific to its actual deployment scene, which specifically includes: A cumulative threshold is set, and when the replacement amount of real-time interaction data in the scene-specific interaction data sample set reaches the cumulative threshold, a small update mechanism is triggered; When the scene-specific interaction data sample set is completely replaced by real-time interaction data compared with the scene-specific interaction data sample set after the last large update, a large update mechanism is triggered. The small update mechanism is as follows: taking the scene reinforcement interaction control model after the last small update as an initial model, performing reinforcement training based on the initial model and real-time interaction data replaced in the scene-specific interaction data sample set to obtain a scene reinforcement interaction control model; The large update mechanism is as follows: taking the scene reinforcement interaction control model after the last large update as a first initial model, performing reinforcement training based on the first initial model and the scene-specific interaction data sample set to obtain a first scene reinforcement interaction control model; The scene reinforcement interaction control model after the last small update is taken as a second initial model, and reinforcement training is performed based on the initial model and real-time interaction data replaced in the scene-specific interaction data sample set to obtain a second scene reinforcement interaction control model; The performances of the first scene reinforcement interaction control model and the second scene reinforcement interaction control model on the scene-specific interaction data sample set are verified respectively, and the one with a smaller loss value is selected as the scene reinforcement interaction control model.

[0025] In some preferred embodiments, the accumulation threshold is set to 1 / 4-1 / 8 of the fixed capacity of the circular data buffer.

[0026] The two-stage update triggering and optimization mechanism designed in this embodiment brings significant technical improvement. Specifically, by accurately associating the small update threshold with the local replacement order of magnitude of the fixed capacity of the buffer, it ensures that incremental training captures the dynamic changes of the scene with minimal computational resource consumption. The large update mechanism started when the data is completely replaced relies on parallel training of double models and horizontal selection, which not only inherits the integrity of historical knowledge but also integrates the latest scene features, completely avoiding the performance fluctuation risk in traditional continuous learning. This strategy maintains the real-time response capability of the model while using adaptive matching of buffer size and update threshold to achieve efficient coordination of training resource allocation and model evolution needs, ultimately outputting a stable optimal scene reinforcement model, significantly improving the interaction robustness and decision accuracy of the robot in long-term deployment environments.

[0027] Embodiment Four

[0028] Referring to Figure 3 On the basis of any one of Embodiment One, Embodiment Two and Embodiment Three, the embodiment further proposes at least one target application scenario based on a robot, and constructs a robot interaction data set covering all different target application scenarios, which specifically includes: Based on the actual deployment data of the robot, the application proportion of the robot actually deployed to various target application scenarios is determined; According to the application proportion of the robot actually deployed to various target application scenarios, the robot interaction data of each target application scenario is extracted to form a general-purpose interaction dataset, and the number proportion of the robot interaction data of each target application scenario in the general-purpose interaction dataset is ensured to be the same as the application proportion of the robot actually deployed to various target application scenarios.

[0029] The training of the general-purpose interaction control model of the robot based on the general-purpose interaction dataset specifically includes: According to the number of the robot interaction data of each target scene in the general-purpose interaction dataset, the cost-sensitive learning is introduced for the robot interaction data of the target scene with a small number, and the general-purpose interaction control model of the robot is trained. Specifically, the cost-sensitive learning mechanism is that: the target scene with the largest number of robot interaction data in the general-purpose interaction dataset is set as a main target scene, the target scene with the number of robot interaction data less than 1 / 2 of the number of robot interaction data of the main target scene in the general-purpose interaction dataset is recorded as a low-frequency target scene, the loss value of the general-purpose interaction control model under the low-frequency target scene is compensated in weight, and the compensation multiple calculation formula is: ; wherein, is the loss value compensation weight multiple of the jth low-frequency target scene, k is a compensation coefficient, and the compensation coefficient is in the range of 1 / 4-1 / 2, is the number of robot interaction data of the jth low-frequency target scene, is the number of robot interaction data of the main target scene.

[0030] The proportion-weighted dataset construction and the cost-sensitive learning mechanism proposed in this embodiment form a double guarantee, which significantly improves the universality and fairness of the model in actual deployment: by strictly aligning the scene distribution of the dataset with the actual deployment proportion, it is ensured that the general-purpose model deeply internalizes the high-frequency scene requirements and low-frequency scene characteristics during the pre-training stage; the compensation weight formula with logarithmic smoothing processing is innovatively introduced, which intelligently identifies data scarce scenes and applies moderate gradient enhancement, effectively bridging the decision-making short board of the low-frequency scene while preventing overfitting of noise. This mechanism enables the final general-purpose model to break through the constraints of traditional data skew, and has full-scene seamless migration capability in the cold start stage, so that the robot can maintain optimal performance in high-frequency scenes while achieving systematic accuracy leap in response to sudden low-frequency tasks, thereby fundamentally building a robust interaction base that can adapt to a dynamic deployment ecosystem.

[0031] Embodiment five:

[0032] Based on the embodiments one to four, this embodiment further proposes a scene-based robot interaction control system, which includes: a general-purpose dataset construction module configured to construct a general-purpose interaction dataset covering all different target application scenarios based on at least one target application scenario of the robot; a general-purpose model training module configured to train a general-purpose interaction control model of the robot based on the general-purpose interaction dataset; a real-time data collection module deployed at the robot and configured to continuously collect real-time interaction data of the robot during actual operation of the robot; a scenario-specific data buffer module including a fixed-capacity circular data buffer and configured to dynamically update and store a scenario-specific interaction data sample set through a preset forgetting mechanism; a scenario-enhanced training module configured to perform enhanced training based on the general-purpose interaction control model and the scenario-specific interaction data sample set to generate a scenario-enhanced interaction control model; a model storage and calling module configured to store the trained scenario-enhanced interaction control model locally on the robot and to call the model when performing a task; an interaction control execution module configured to input a user instruction and environmental perception information into the scenario-enhanced interaction control model to generate a control instruction to drive an action of the robot.

[0033] The scenario-enhanced training module includes: a two-stage update triggering unit configured to trigger training according to an update degree of the scenario-specific data sample set: when a sample replacement amount reaches 1 / 4-1 / 8 of a capacity of the fixed-capacity buffer, a small update mechanism is triggered, and when the sample is completely replaced, a large update mechanism is triggered; a model iteration optimization unit configured to perform: in the small update mechanism, incremental training based on a previous scenario-enhanced model and new samples, in the large update mechanism, full-sample training based on a previous large update model, and loss comparison with the small update model, and optimally save as a final scenario-enhanced model.

[0034] The general-purpose dataset construction module includes: a scenario proportion analysis unit configured to statistically analyze proportions of the robot actually deployed to different target application scenarios; a data weighted sampling unit configured to extract interaction data from different scenarios in proportion to construct a general-purpose interaction dataset with consistent number proportions and actual deployment proportions; a cost-sensitive training unit configured to introduce a cost-sensitive learning algorithm to a target scenario with less data in general-purpose model training to improve generalization ability of the model to the rare scenario.

[0035] In summary, the application has the advantages that: through the combination of pre-training of a general-purpose interactive control model covering multiple scenes and real-time scene-specific reinforcement training, the contradiction between insufficient cross-scene migration ability of the robot and adaptability to specific scenes is effectively solved; the fixed-capacity circular buffer and dynamic forgetting mechanism are innovatively used to manage real-time interactive data, combined with a phased model updating strategy, which significantly suppresses the risk of catastrophic forgetting while ensuring optimal fitting of the model to the current scene; through the construction of a weighted data set and a cost-sensitive learning mechanism according to the actual deployment ratio, the data skew problem is overcome from the source, so that the robot has the ability of continuous self-optimization, and the long-term interactive performance and decision accuracy in a complex dynamic environment are significantly improved.

[0036] The basic principles, main features and advantages of the application are shown and described above. Those skilled in the art should understand that the application is not limited by the above examples, and the above examples and descriptions in the specification are only the principles of the application. Without departing from the spirit and scope of the application, various changes and improvements can be made to the application, and these changes and improvements all fall within the scope of the claimed application. The scope of protection required by the application is defined by the appended claims and their equivalents.

Claims

1. A method for controlling robot interaction based on a scene, characterized by, The method comprises the following steps: constructing a robot interaction dataset covering all different target application scenarios based on at least one target application scenario of the robot, and denoted as a general-purpose interaction dataset; training a general-purpose interaction control model of the robot based on the general-purpose interaction dataset; continuously collecting real-time interaction data of the robot in an actual scenario during actual operation of the robot after the robot is deployed and runs; constructing a scenario-specific interaction data sample set based on the real-time interaction data of the robot in the actual scenario; reinforcing a scenario reinforcement interaction control model of the robot for the actual deployment scenario based on the general-purpose interaction control model of the robot and the scenario-specific interaction data sample, and storing the scenario reinforcement interaction control model in the local; calling the scenario reinforcement interaction control model stored in the local and updated through reinforcement training when the robot needs to perform a new interaction task in the target application scenario; inputting a new interaction instruction and current environment perception information into the scenario reinforcement interaction control model to generate a final control instruction to drive the robot to perform an action.

2. The method of claim 1, wherein, The method of constructing a scenario-specific interaction data sample set based on real-time interaction data of the robot in an actual scenario comprises the following steps: initializing a fixed-capacity circular data buffer for storing interaction training data samples for the target application scenario; continuously collecting real-time interaction data samples conforming to the characteristics of the scenario during actual operation of the target application scenario; adding the newly collected real-time interaction data samples to the circular data buffer by using a preset forgetting mechanism: combining the real-time interaction data samples stored in the circular data buffer into a scenario-specific interaction data sample set.

3. The method of claim 2, wherein, The forgetting mechanism specifically comprises the following steps: if the circular data buffer is not full, directly storing the new interaction data sample; if the circular data buffer is full, removing at least one old interaction data sample in the buffer according to a preset rule, and then adding the new interaction data sample, so that the total amount of data in the buffer always remains at a preset fixed capacity; wherein the preset rule specifically comprises the following steps: dividing the real-time interaction data samples stored in the circular data buffer into types, and evenly adding a type contribution value to each type based on the total number of types of the real-time interaction data samples; analyzing the importance of each real-time interaction data sample based on the type contribution value of each type, the number of real-time interaction data samples under each type, and the storage time of each real-time interaction data sample; removing the real-time interaction data sample with the lowest importance and adding the new interaction data sample.

4. The method of claim 3, wherein, The method of reinforcing a scenario reinforcement interaction control model of the robot for the actual deployment scenario based on the general-purpose interaction control model of the robot and the scenario-specific interaction data sample specifically comprises the following steps: setting a cumulative threshold, and triggering a small update mechanism when the replacement amount of real-time interaction data in the scenario-specific interaction data sample set reaches the cumulative threshold; triggering a large update mechanism when the real-time interaction data in the scenario-specific interaction data sample set is completely replaced compared with the scenario-specific interaction data sample set after the last large update. The small update mechanism is as follows: taking the scene reinforcement interaction control model after the last small update as an initial model, performing reinforcement training based on the initial model and real-time interaction data replaced in the scene-specific interaction data sample set to obtain a scene reinforcement interaction control model; The large update mechanism is as follows: taking the scene reinforcement interaction control model after the last large update as a first initial model, performing reinforcement training based on the first initial model and the scene-specific interaction data sample set to obtain a first scene reinforcement interaction control model; Taking the scene reinforcement interaction control model after the last small update as a second initial model, performing reinforcement training based on the initial model and real-time interaction data replaced in the scene-specific interaction data sample set to obtain a second scene reinforcement interaction control model; The performances of the first scene reinforcement interaction control model and the second scene reinforcement interaction control model on the scene-specific interaction data sample set are verified respectively, and the one with a smaller loss value is selected as the scene reinforcement interaction control model.

5. The method of claim 4, wherein, The cumulative threshold is set to 1 / 4-1 / 8 of the fixed capacity of the circular data buffer.

6. The method of claim 1, wherein, The robot interaction data set covering all different target application scenarios is constructed based on at least one target application scenario of the robot, and the robot interaction data set covering all different target application scenarios is constructed based on at least one target application scenario of the robot. The application proportion of the robot actually deployed to various target application scenarios is determined based on the actual deployment data of the robot. The robot interaction data of each target application scenario is extracted according to the application proportion of the robot actually deployed to various target application scenarios to form a general-purpose interaction data set, so as to ensure that the quantity proportion of the robot interaction data of each target application scenario in the general-purpose interaction data set is the same as the application proportion of the robot actually deployed to various target application scenarios.

7. The method of claim 6, wherein, The general-purpose interaction control model of the robot is trained based on the general-purpose interaction data set, and the general-purpose interaction control model of the robot is trained based on the general-purpose interaction data set. The general-purpose data set construction module is configured to construct a general-purpose interaction data set covering all different target application scenarios based on at least one target application scenario of the robot.

8. A scene-based robot interaction control system, characterized by, The general-purpose model training module is configured to train a general-purpose interaction control model of the robot based on the general-purpose interaction data set. The real-time data collection module is deployed on the robot side and is configured to continuously collect real-time interaction data of the robot during actual operation of the robot. The scene data buffer module includes a circular data buffer with a fixed capacity and is configured to dynamically update and store the scene-specific interaction data sample set through a preset forgetting mechanism. The scene reinforcement training module is configured to perform reinforcement training based on the general-purpose interaction control model and the scene-specific interaction data sample set to generate a scene reinforcement interaction control model. The model storage and calling module is configured to store the trained scene reinforcement interaction control model in the robot locally and to call the model when performing a task. ​ ​ The interactive control execution module is configured to input the user instruction and the environment perception information into the scene reinforcement interactive control model to generate a control instruction to drive a robot action.

9. The scene-based robot interaction control system of claim 8, wherein, The scene reinforcement training module comprises: A double-stage update triggering unit is configured to trigger training according to an update degree of the scene-specific data sample set, wherein when a sample replacement amount reaches 1 / 4-1 / 8 of a capacity of a fixed buffer, a small update mechanism is triggered, and when the sample is completely replaced, a large update mechanism is triggered. A model iteration optimization unit is configured to perform: in the small update mechanism, incremental training based on a previous scene reinforcement model and new samples, in the large update mechanism, full-sample training based on a previous large update model, and loss comparison with the small update model, and the better one is saved as a final scene reinforcement model.

10. The scene-based robot interaction control system of claim 8, wherein, The general-purpose data set construction module comprises: A scene proportion analysis unit is configured to count a proportion of a robot actually deployed to each target application scene. A data weighted sampling unit is configured to extract interaction data from each scene according to the proportion to construct a general-purpose interaction data set with consistent quantity proportion and actual deployment proportion. A cost-sensitive training unit is configured to introduce a cost-sensitive learning algorithm to a target scene with less data in general-purpose model training to improve the generalization ability of the model to the rare scene.

Images