A scene-based robot interaction control method and system

Abstract

The application discloses a kind of scene-based robot interaction control method and system, it is related to robot control technical field, including: constructing general-purpose interaction dataset;Train the general-purpose interaction control model of robot;Continuously collect real-time interaction data of robot in actual scene;Construct scene-specific interaction data sample set;Strengthen the training of robot scene reinforcement interaction control model in its actual deployment scene, and store in local;When robot needs to execute new interaction task in the target application scene, call scene reinforcement interaction control model;Generate final control instruction to drive robot to execute action.The application has the advantages that: by the combination of pre-training covering multi-scene general-purpose interaction control model and real-time scene-specific reinforcement training, robot has sustained self-optimization capability, significantly improves long-term interaction performance and decision accuracy in complex dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a scenario-based robot interaction control method and system. Background Technology

[0002] In the field of robot interactive control, existing technologies typically employ two approaches: First, training dedicated models independently for each specific scenario. While this ensures scenario performance, it requires a large number of pre-built scenario models and cannot adapt to unforeseen scenarios, resulting in high deployment costs and poor flexibility. Second, relying on a single general-purpose model to cover multiple scenarios. While possessing basic generalization capabilities, this approach falls short in specific scenarios, such as dynamically changing environments or high-precision tasks, requiring frequent manual parameter adjustments or retraining, making true adaptive learning difficult. For example, while traditional continuous learning methods support model updates, they suffer from catastrophic forgetting—new scenario data overwrites old knowledge, causing model performance degradation in historical scenarios. Furthermore, the lack of dynamic management mechanisms for training data means that old and noisy samples continuously influence the model, hindering the adaptive optimization of robots during long-term operation.

[0003] To address these issues, the industry has explored methods combining transfer learning and incremental learning, but limitations remain: First, pre-trained general models often suffer from data distribution biases, such as high-frequency scene data dominating, weakening their generalization ability to low-frequency scenes; second, real-time incremental training lacks efficient data update strategies, and simple sample replacement easily leads to the loss of key experience or the introduction of noise, while frequent global retraining results in excessive computational overhead. Current solutions have not yet achieved a collaborative mechanism of "lightweight dynamic data management + phased adaptive training," and a robust interactive control framework that can balance scene-specific optimization with system resource constraints is urgently needed. Summary of the Invention

[0004] To address the aforementioned technical problems, a scenario-based robot interaction control method and system are provided. This technical solution solves at least one of the technical problems mentioned in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A scenario-based robot interaction control method, comprising:

[0007] Based on at least one target application scenario of the robot, construct a robot interaction dataset covering all different target application scenarios, denoted as the general-purpose interaction dataset;

[0008] A general-purpose interactive control model for robots is trained based on a general-purpose interactive dataset;

[0009] After the robot is deployed and runs, real-time interaction data of the robot in an actual scene is continuously collected during actual running;

[0010] Based on the real-time interaction data of the robot in the actual scene, a scene-specific interaction data sample set is constructed;

[0011] Based on the general-purpose 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 stored locally through reinforcement training;

[0012] When the robot needs to perform a new interaction task in the target application scene, the scene reinforcement interaction control model updated through reinforcement training stored locally is called;

[0013] The new interaction instruction and the current environment perception information are input into the scene reinforcement interaction control model to generate a final control instruction to drive the robot to perform an action.

[0014] Preferably, the construction of the scene-specific interaction data sample set based on the real-time interaction data of the robot in the actual scene specifically includes:

[0015] A fixed-capacity circular data buffer is initialized to store interaction training data samples for the target application scene;

[0016] During actual running of the target application scene, real-time interaction data samples conforming to the characteristics of the scene are continuously collected;

[0017] A preset forgetting mechanism is used to add the newly collected real-time interaction data samples to the circular data buffer:

[0018] The real-time interaction data samples stored in the circular data buffer are combined into a scene-specific interaction data sample set.

[0019] Preferably, the forgetting mechanism specifically includes:

[0020] If the circular data buffer is not full, the new interaction data sample is directly stored;

[0021] 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;

[0022] The preset rule specifically includes:

[0023] Based on the type division of the real-time interaction data samples stored in the circular data buffer, a type contribution value is evenly added to each type based on the total number of types of the real-time interaction data samples.

[0024] analyze the importance of each real-time interaction data sample based on the type contribution value of each category, the number of real-time interaction data samples under each category, and the storage time of each real-time interaction data sample;

[0025] remove the real-time interaction data sample with the lowest importance and add a new interaction data sample.

[0026] Preferably, the robot-based general-purpose interaction control model and the scene-specific interaction data sample are subjected to reinforcement training to obtain a scene reinforcement interaction control model specific to the actual deployment scene of the robot, which specifically comprises:

[0027] A cumulative threshold is set, and when the amount of real-time interaction data replacement in the scene-specific interaction data sample set reaches the cumulative threshold, a small update mechanism is triggered;

[0028] 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;

[0029] The small update mechanism is: taking the scene reinforcement interaction control model after the last small update as an initial model, and based on the initial model, the real-time interaction data replaced in the scene-specific interaction data sample set is subjected to reinforcement training to obtain the scene reinforcement interaction control model.

[0030] The large update mechanism is: taking the scene 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 subjected to reinforcement training to obtain a first scene reinforcement interaction control model.

[0031] Taking the scene 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 replaced in the scene-specific interaction data sample set is subjected to reinforcement training to obtain a second scene reinforcement interaction control model.

[0032] 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.

[0033] Preferably, the cumulative threshold is set to 1 / 4-1 / 8 of the fixed capacity of the circular data buffer.

[0034] Preferably, the at least one target application scene based on the robot comprises a robot interaction data set covering all different target application scenes, which specifically comprises:

[0035] determine an application proportion of the robots actually deployed to various target application scenarios based on actual deployment data of the robots;

[0036] extract robot interaction data of each target application scenario according to the application proportion of the robots actually deployed to various target application scenarios, to form a general-purpose interaction dataset, and ensure that the quantity proportion of the robot interaction data of each target application scenario in the general-purpose interaction dataset is the same as the application proportion of the robots actually deployed to various target application scenarios.

[0037] Preferably, the training of the general-purpose interaction control model of the robot based on the general-purpose interaction dataset specifically includes:

[0038] based on the quantity of the robot interaction data of each target scenario in the general-purpose interaction dataset, introducing cost-sensitive learning for the robot interaction data of a target scenario with a small quantity, and training the general-purpose interaction control model of the robot.

[0039] Further, a robot interaction control system based on scenarios is proposed to implement the robot interaction control method based on scenarios as described above, and includes:

[0040] 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;

[0041] 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;

[0042] a real-time data collection module deployed at the robot end and configured to continuously collect real-time interaction data of the robot during actual operation of the robot;

[0043] a scenario 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;

[0044] a scenario reinforcement training module configured to perform reinforcement training based on the general-purpose interaction control model and the scenario-specific interaction data sample set to generate a scenario-reinforced interaction control model;

[0045] a model storage and calling module configured to store the trained scenario-reinforced interaction control model in the robot locally and call the model when performing a task;

[0046] an interaction control execution module configured to input a user instruction and environmental perception information into the scenario-reinforced interaction control model to generate a control instruction to drive the robot to act.

[0047] Optionally, the scenario reinforcement training module includes:

[0048] A double-stage update triggering unit is configured to trigger training according to the update degree of the scenario-specific data sample set; when the sample replacement amount reaches 1 / 4-1 / 8 of the 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.

[0049] A model iteration optimization unit is configured to perform: in the small update mechanism, incremental training based on the previous scenario reinforcement model and the new sample; 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 optimal one is saved as the final scenario reinforcement model.

[0050] Optionally, the general-purpose data set construction module comprises:

[0051] A scenario proportion analysis unit is configured to count the proportion of the robot actually deployed to each target application scenario;

[0052] A data weighted sampling unit is configured to extract interaction data from each scenario according to the proportion, and construct a general-purpose interaction data set with consistent quantity proportion and actual deployment proportion;

[0053] A cost-sensitive training unit is configured to introduce a cost-sensitive learning algorithm for the target scenario with less data in the general-purpose model training, so as to improve the generalization ability of the model to the rare scenario.

[0054] Compared with the prior art, the beneficial effects of the present application are:

[0055] The present application effectively solves the contradiction between the insufficient cross-scenario migration ability of the robot and the adaptability to the specific scenario by combining the pre-trained general-purpose interaction control model covering multiple scenarios with real-time scenario-specific reinforcement training; innovatively uses a fixed-capacity circular buffer and a dynamic forgetting mechanism to manage real-time interaction data, and combines a phased model update strategy to significantly suppress the risk of catastrophic forgetting while ensuring the optimal fitting of the model to the current scenario; by constructing a weighted data set according to the actual deployment proportion and a cost-sensitive learning mechanism, the data skew problem is overcome from the source, so that the robot has a continuous self-optimization ability, and the long-term interaction performance and decision accuracy in a complex dynamic environment are significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A flowchart of the scenario-based robot interaction control method for embodiment one is shown;

[0057] Figure 2 A flowchart of the method for constructing a scenario-specific interaction data sample set for embodiment two is shown;

[0058] Figure 3 A flowchart of the method for constructing a robot interaction data set covering all different target application scenarios for embodiment three is shown. DETAILED DESCRIPTION

[0059] The following description is presented to enable any person skilled in the art to practice the application as claimed. The preferred embodiments disclosed herein are only examples of the many possible variations of the present application.

[0060] Embodiment one:

[0061] Referring to Figure 1 The embodiment proposes a scene-based robot interaction control method, which includes:

[0062] Based on at least one target application scenario of the robot, a robot interaction dataset covering all different target application scenarios is constructed, denoted as a general-purpose interaction dataset, a unified general-purpose dataset is constructed by aggregating multi-scene data, the single-scene data deviation problem is avoided from the source, the cross-scene common feature learning basis is provided for the model, and the initial generalization ability of the pre-training model is significantly improved

[0063] The general-purpose interaction control model of the robot is trained based on the general-purpose interaction dataset. The trained general-purpose model has basic adaptability to unknown scenes, reduces the cold start time when the robot is deployed, provides high starting point parameters for subsequent scene optimization, and greatly reduces the convergence difficulty of scene-specific models;

[0064] After the robot is deployed and runs, real-time interaction data of the robot in the actual scene is continuously collected during actual operation, real scene data is dynamically collected based on the real scene, and the unique noise and interaction mode of the deployment environment are directly captured, solving the domain gap problem between simulation data and real environment, and providing high-quality input for model customization optimization;

[0065] Based on the real-time interaction data of the robot in the actual scene, a scene-specific interaction data sample set is constructed;

[0066] Based on the general-purpose interaction control model of the robot and the scene-specific interaction data sample, the robot is reinforced to train a scene reinforcement interaction control model for its actual deployment scene, and is stored locally;

[0067] When the robot needs to perform a new interaction task in a target application scenario, the scene reinforcement interaction control model stored locally and updated after reinforcement training is called, the scene-specific model optimized locally is directly called, the cloud communication delay is eliminated and the offline execution reliability is guaranteed, and the scene environment changes are dynamically adapted;

[0068] The new interaction instruction and the current environment perception information are input into the scene reinforcement interaction control model to generate a final control instruction to drive the robot to perform an action. Precise control instructions are generated in real time under the driving of the dedicated model, which significantly improves the success rate of task execution and response speed, and especially shows stronger anti-interference ability and human-machine cooperation adaptability in complex dynamic scenes.

[0069] Embodiment two:

[0070] Referring to Figure 2 The embodiment is based on embodiment one, and further proposes constructing a scene-specific interaction data sample set based on real-time interaction data of the robot in an actual scene, which specifically includes:

[0071] An initialized fixed-capacity circular data buffer is used to store interaction training data samples for a target application scenario;

[0072] In the actual running process of the target application scenario, real-time interaction data samples conforming to the characteristics of the scene are continuously collected;

[0073] The newly collected real-time interaction data samples are added to the circular data buffer using a preset forgetting mechanism:

[0074] The real-time interaction data samples stored in the circular data buffer are combined into a scene-specific interaction data sample set;

[0075] Specifically, the forgetting mechanism is specifically:

[0076] If the circular data buffer is not full, the new interaction data sample is directly stored;

[0077] 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;

[0078] The preset rule is specifically:

[0079] Based on the real-time interaction data samples stored in the circular data buffer, type classification is performed, and based on the total number of types of real-time interaction data samples, a type contribution value is added to each type in a balanced manner;

[0080] 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;

[0081] The calculation formula of importance is:

[0082] ;

[0083] In the formula, Let N represent the importance of the i-th real-time interactive data sample, N be the total number of types of real-time interactive data samples stored in the circular data buffer, and M be the number of real-time interactive data samples of the type corresponding to the i-th real-time interactive data sample. Let be the storage time of the i-th real-time interactive data sample, and e be a natural constant;

[0084] Remove the least important real-time interaction data samples and add new interaction data samples.

[0085] The intelligent forgetting mechanism proposed in this embodiment achieves dynamic optimization management of scene interaction data under a fixed capacity limit by integrating a quantitative evaluation framework that combines interaction type balance protection, time decay effect, and adaptive adjustment of intra-class sample density. This mechanism actively identifies and retains diverse interaction samples with long-term value for model evolution, especially enhancing the representativeness of low-frequency key scene patterns, while accurately eliminating outdated or redundant data. Its mathematical decision-making model, coupled with multi-dimensional weights, significantly improves the interpretability and stability of data purification, ensuring that the buffer always maintains a high-information-density scene state distribution. This provides a high-quality sample set with both timeliness and generalization potential for subsequent reinforcement training, suppressing the risk of catastrophic forgetting from the data source, endowing the robot with continuous learning ability and lifelong adaptability to dynamic environments, and ultimately achieving a synergistic leap in interaction accuracy and scene robustness.

[0086] Example 3:

[0087] This embodiment, based on Embodiment 2, further proposes to construct a scenario-specific interaction data sample set based on the robot's real-time interaction data in actual scenarios, specifically including:

[0088] Based on the robot's general-purpose interactive control model and scenario-specific interactive data samples, the robot's interactive control model is enhanced through training to adapt to its actual deployment scenarios. This specifically includes:

[0089] Set a cumulative threshold. When the amount of real-time interactive data replacement in the scenario-specific interactive data sample set reaches the cumulative threshold, a small update mechanism is triggered.

[0090] When the real-time interaction data is completely replaced by the scene-specific interaction data sample set after the last major update, the major update mechanism is triggered.

[0091] The small update mechanism is as follows: the scene-enhanced interaction control model after the previous small update is used as the initial model. Based on the initial model, real-time interaction data replaced by scene-specific interaction data sample set is used for reinforcement training to obtain the scene-enhanced interaction control model.

[0092] The major update mechanism is as follows: the scene-enhanced interaction control model after the last major update is used as the first initial model. Based on the first initial model and combined with the scene-specific interaction data sample set, the first scene-enhanced interaction control model is obtained through enhanced training.

[0093] Using the scene-enhanced interaction control model updated in the previous minor update as the second initial model, the second scene-enhanced interaction control model is obtained by enhancing the training based on the initial model and the real-time interaction data replaced in the scene-specific interaction data sample set.

[0094] The performance of the first scenario-based enhanced interaction control model and the second scenario-based enhanced interaction control model were verified on the scenario-specific interaction data sample set, and the model with the smaller loss value was selected as the scenario-based enhanced interaction control model.

[0095] In some preferred embodiments, the cumulative threshold is set to 1 / 4 to 1 / 8 of the fixed capacity of the circular data buffer.

[0096] The dual-level update triggering and optimization mechanism designed in this embodiment brings significant technical improvements. Specifically, by accurately associating the small update threshold with the local rotation level of the fixed buffer capacity, incremental training is ensured to capture dynamic changes in the scene with minimal computational overhead. The large update mechanism, which is initiated when the data is completely rotated, relies on parallel training of dual models and lateral selection to both inherit the integrity of historical knowledge and integrate the latest scene features, completely avoiding the performance fluctuation risks in traditional continuous learning. While maintaining the real-time response capability of the model, this strategy utilizes the adaptive matching of the buffer size and the update threshold to achieve efficient coordination between training resource allocation and model evolution requirements, ultimately outputting a stable and optimal scene reinforcement model, significantly improving the robot's interaction robustness and decision-making accuracy in long-term deployment environments.

[0097] Example 4:

[0098] Reference Figure 3 As shown, this embodiment, based on any one of Embodiment 1, Embodiment 2, and Embodiment 3, further proposes at least one target application scenario based on a robot, and constructs a robot interaction dataset covering all different target application scenarios, specifically including:

[0099] Based on actual robot deployment data, determine the application ratio of robots actually deployed in various target application scenarios;

[0100] Based on the actual deployment ratio of robots in various target application scenarios, robot interaction data for each target application scenario is extracted to form a general-purpose interaction dataset. This ensures that the proportion of robot interaction data for each target application scenario in the general-purpose interaction dataset is the same as the actual deployment ratio of robots in various target application scenarios.

[0101] The general interactive control model for training robots based on general interactive datasets specifically includes:

[0102] Based on the number of robot interaction data for each target scenario in the generalized interaction dataset, cost-sensitive learning is introduced for robot interaction data of a small number of target scenarios to train the robot's generalized interaction control model.

[0103] Specifically, the cost-sensitive learning mechanism is as follows: The target scenario with the largest number of robot interaction data points in the generalized interaction dataset is designated as the primary target scenario. Target scenarios where the number of robot interaction data points in the generalized interaction dataset is less than half that in the primary target scenario are designated as low-frequency target scenarios. The loss value of the generalized interaction control model in low-frequency target scenarios is then weighted and compensated. The compensation ratio is calculated using the following formula:

[0104] ;

[0105] in, The loss value of the j-th low-frequency target scene is supplemented with a weight multiplier, where k is the compensation coefficient, and the value of the compensation coefficient ranges from 1 / 4 to 1 / 2. Let J be the number of robot interaction data points for the j-th low-frequency target scenario. The amount of robot interaction data in the main target scenario.

[0106] The proportionally weighted dataset construction and cost-sensitive learning mechanism proposed in this embodiment provide dual guarantees, significantly improving the model's universality and fairness in actual deployment: By strictly aligning the dataset's scene distribution with the actual deployment ratio, it ensures that the general-purpose model deeply internalizes the needs of high-frequency scenes and the characteristics of low-frequency scenes during the pre-training stage; the innovative introduction of a compensation weight formula with logarithmic smoothing intelligently identifies data-scarce scenes and applies appropriate gradient enhancement, effectively bridging the decision-making shortcomings of low-frequency scenes while preventing overfitting noise. This mechanism enables the final general-purpose model to break through the constraints of traditional data skew, possessing seamless transfer capabilities across all scenes from the cold start stage. This allows the robot to maintain optimal performance in high-frequency scenes while achieving a systematic leap in response accuracy for sudden low-frequency tasks, fundamentally building a robust interactive foundation adaptable to dynamic deployment ecosystems.

[0107] Example 5:

[0108] This embodiment, based on embodiments one through four, further proposes a scenario-based robot interaction control system, including:

[0109] A general-purpose dataset building module is used to build a general-purpose interactive dataset covering all different target application scenarios for at least one robot-based application scenario.

[0110] The general-purpose model training module is used to train a general-purpose interactive control model for the robot based on a general-purpose interactive dataset.

[0111] The real-time data acquisition module, deployed on the robot, is used to continuously collect real-time interaction data during the actual operation of the robot;

[0112] The scene data buffer module includes a fixed-capacity circular data buffer, which is used to dynamically update and store scene-specific interaction data sample sets through a preset forgetting mechanism;

[0113] The scenario reinforcement training module is used to perform reinforcement training based on the generalized interaction control model and the scenario-specific interaction data sample set to generate a scenario reinforcement interaction control model.

[0114] The model storage and retrieval module is used to store the trained scene-enhanced interactive control model locally on the robot and retrieve the model when performing tasks.

[0115] The interactive control execution module is used to input user commands and environmental perception information into the scene to enhance the interactive control model and generate control commands to drive the robot's actions.

[0116] The scene reinforcement training module includes:

[0117] A two-level update triggering unit is used to trigger training based on the update degree of the scenario-specific data sample set: when the sample replacement amount reaches 1 / 4 to 1 / 8 of the fixed capacity buffer capacity, a small update mechanism is triggered; when the sample is completely replaced, a large update mechanism is triggered.

[0118] The model iteration optimization unit is used to perform the following: In the small update mechanism, incremental training is performed based on the previous scene enhancement model and the newly added samples; in the large update mechanism, full sample training is performed based on the previous large update model, and the loss is compared with the small update model. The best model is selected and saved as the final scene enhancement model.

[0119] The general-purpose dataset building module includes:

[0120] The scenario ratio analysis unit is used to calculate the proportion of robots actually deployed to each target application scenario;

[0121] The data weighted sampling unit is used to extract interactive data from each scenario proportionally and build a general interactive dataset whose quantity ratio is consistent with the actual deployment ratio.

[0122] Cost-sensitive training units are used to introduce cost-sensitive learning algorithms into target scenarios with limited data during general model training, thereby improving the model's generalization ability to scarce scenarios.

[0123] In summary, the advantages of this invention are as follows: By combining a pre-trained general-purpose interactive control model covering multiple scenarios with real-time scenario-specific reinforcement training, the contradiction between insufficient cross-scenario transfer capability and specific scenario adaptability of the robot is effectively resolved; by innovatively adopting a fixed-capacity circular buffer and a dynamic forgetting mechanism to manage real-time interactive data, combined with a phased model update strategy, the risk of catastrophic forgetting is significantly suppressed while ensuring the optimal fit of the model to the current scenario; by constructing a weighted dataset according to the actual deployment ratio and a cost-sensitive learning mechanism, the data skew problem is overcome from the source, enabling the robot to have continuous self-optimization capability and significantly improving long-term interactive performance and decision-making accuracy in complex dynamic environments.

[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A scene-based robot interaction control method, characterized by, The method comprises the following steps: Based on at least one target application scenario of the robot, a robot interaction dataset covering all different target application scenarios is constructed, which is denoted as a general-purpose interaction dataset; Based on the general-purpose interaction dataset, a general-purpose interaction control model of the robot is trained; After the robot is deployed and runs, real-time interaction data of the robot in the actual scene is continuously collected during actual running; 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 sample, a scene reinforcement interaction control model of the robot for the actual deployment scene is trained by reinforcement, and is stored locally; When the robot needs to perform a new interaction task in the target application scenario, the scene reinforcement interaction control model updated by reinforcement and stored locally is called; The new interaction instruction and the current environment perception information are input into the scene reinforcement interaction control model to generate a final control instruction to drive the robot to perform an action; The method further comprises the following steps: A fixed-capacity circular data buffer is initialized to store interaction training data samples for the target application scenario; During actual running of the target application scenario, real-time interaction data samples conforming to the characteristics of the scene are continuously collected; A preset forgetting mechanism is used to add the newly collected real-time interaction data samples to the circular data buffer: The real-time interaction data samples stored in the circular data buffer are combined into a scene-specific interaction data sample set; The forgetting mechanism specifically comprises the following steps: 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 comprises the following steps: Based on the real-time interaction data samples stored in the circular data buffer, type division is performed, and based on the total number of types of the real-time interaction data samples, a type contribution value is added to each type in a balanced manner; 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.

2. The method of claim 1, wherein, The method further comprises the following steps: 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 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, 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; 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.

3. The method of claim 2, wherein, The cumulative threshold is set to 1 / 4-1 / 8 of the fixed capacity of the circular data buffer.

4. 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.

5. The method of claim 4, 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.

6. 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.

7. The scene-based robot interaction control system of claim 6, 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.

8. The scene-based robot interaction control system of claim 6, 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