Body-aware data processing method and system, model training method and related equipment

Abstract

The embodiment of the application provides a body intelligent data processing method and system, a model training method and related equipment, the body intelligent data processing method comprises the following steps: inputting observation data, motion control input data, motion state data and body intelligent task information corresponding to a target body intelligent agent into a trained body intelligent model, so as to execute the following processing based on the body intelligent model: joint modeling based on the observation data and the motion control input data to obtain joint feature information; determining future observation information based on the joint feature information; determining future motion state information based on the motion state data and the joint feature information; and determining task evaluation information based on at least two of the future observation information, the future motion state information, the joint feature information and the body intelligent task information. The embodiment of the application improves the problem that the body intelligent task processing process is prone to over-reliance on a single observation result and is difficult to simultaneously consider multiple data processing targets.

Description

Technical Field

[0001] This application relates to the fields of embodied intelligence and model training technology, and in particular to embodied intelligence data processing methods and systems, model training methods and related equipment. Background Technology

[0002] With the development of artificial intelligence, robotics, multimodal perception, and data-driven modeling technologies, embodied intelligence has gradually become an important research direction in the fields of intelligent control, intelligent interaction, and autonomous decision-making. Embodied intelligent agents need to possess the ability to perceive, move, interact, and perform tasks in real or simulated environments. They should be able to analyze, predict, and judge subsequent actions and execution results by receiving environmental information, their own state information, and task objective information. However, in related technologies, the embodied intelligence task processing process tends to over-rely on single observation results, making it difficult to simultaneously achieve multiple data processing objectives.

[0003] Based on this, embodiments of this application provide a method and system for embodied intelligent data processing, a model training method and related equipment, to improve related technologies. Summary of the Invention

[0004] The purpose of this application is to provide an embodied intelligence data processing method and system, a model training method and related equipment, to improve the problem in the related technology that the embodied intelligence task processing process is prone to over-reliance on a single observation result and is difficult to simultaneously take into account environmental evolution prediction, ontological motion trend analysis and task execution effect judgment.

[0005] In a first aspect, embodiments of this application provide an embodied intelligence data processing method, the method comprising: receiving observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to a target embodied intelligence agent; inputting the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligence agent into a trained embodied intelligence model, and performing the following processing based on the embodied intelligence model: performing joint modeling based on the observation data and motion control input data corresponding to the target embodied intelligence agent to obtain joint feature information corresponding to the target embodied intelligence agent; determining future observation information corresponding to the target embodied intelligence agent based on the joint feature information corresponding to the target embodied intelligence agent; determining future motion state information corresponding to the target embodied intelligence agent based on the motion state data and joint feature information corresponding to the target embodied intelligence agent; and determining task evaluation information corresponding to the target embodied intelligence agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the target embodied intelligence agent.

[0006] In some embodiments, the observation data includes multi-view observation data.

[0007] In some embodiments, the observation data includes current observation data.

[0008] In some embodiments, the motion control input data includes at least one of end-effector control input information and joint control input information.

[0009] In some embodiments, the motion state data includes at least one of end-effector state information and joint state information.

[0010] In some embodiments, the motion state data includes current motion state data and historical motion state data.

[0011] In some embodiments, the task evaluation information is determined based on the future observation information and the embodied intelligent task information.

[0012] In some embodiments, the task evaluation information includes at least one of the following: reward value, task success probability, task score, task completion rate, task completion tag, and task progress information.

[0013] Secondly, embodiments of this application provide a method for training an embodied intelligence model. The method includes: receiving observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to an embodied intelligence agent; inputting the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the embodied intelligence agent into the embodied intelligence model to be trained, and performing the following processing based on the embodied intelligence model to be trained: performing joint modeling based on the observation data and motion control input data corresponding to the embodied intelligence agent to obtain joint feature information corresponding to the embodied intelligence agent; and performing at least one of the following processing based on the embodied intelligence model to be trained: based on the observation data and motion control input data corresponding to the embodied intelligence agent... Based on the joint feature information corresponding to the embodied intelligent agent, determine the future observation information corresponding to the embodied intelligent agent; based on the motion state data and joint feature information corresponding to the embodied intelligent agent, determine the future motion state information corresponding to the embodied intelligent agent; based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligent task information corresponding to the embodied intelligent agent, determine the task evaluation information corresponding to the embodied intelligent agent; based on at least two of the future observation information, future motion state information, joint feature information, and task evaluation information corresponding to the embodied intelligent agent and the corresponding target information, determine the training loss; based on the training loss, update at least one model parameter of the embodied intelligent model to be trained.

[0014] In some embodiments, the embodied intelligence model to be trained includes a joint modeling module, a state expert module, and an evaluation expert module. The joint modeling module is used to perform joint modeling based on the observation data and motion control input data corresponding to the embodied intelligence agent to obtain joint feature information corresponding to the embodied intelligence agent, wherein the future observation information corresponding to the embodied intelligence agent is determined based on the joint feature information corresponding to the embodied intelligence agent. The state expert module is used to determine the future motion state information corresponding to the embodied intelligence agent based on the motion state data and joint feature information corresponding to the embodied intelligence agent. The evaluation expert module is used to determine the task evaluation information corresponding to the embodied intelligence agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the embodied intelligence agent.

[0015] In some embodiments, the training loss includes at least one of future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss; the future observation prediction loss is determined based on the future observation information corresponding to the embodied agent and its corresponding target future observation information.

[0016] In some embodiments, the future motion state prediction loss is determined based on the future motion state information corresponding to the embodied agent and the corresponding target future motion state information.

[0017] In some embodiments, the joint feature loss is determined based on the joint feature information corresponding to the embodied agent and its corresponding target joint feature information.

[0018] In some embodiments, the task evaluation loss is determined based on the task evaluation information corresponding to the embodied agent and its corresponding target task evaluation information.

[0019] Thirdly, embodiments of this application provide an embodied intelligence data processing system, comprising: a first receiving module, configured to receive observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to a target embodied intelligence agent; and a first processing module, configured to input the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligence agent into a trained embodied intelligence model, and to perform the following processing based on the embodied intelligence model: performing joint modeling based on the observation data and motion control input data corresponding to the target embodied intelligence agent to obtain joint feature information corresponding to the target embodied intelligence agent; determining future observation information corresponding to the target embodied intelligence agent based on the joint feature information corresponding to the target embodied intelligence agent; determining future motion state information corresponding to the target embodied intelligence agent based on the motion state data and joint feature information corresponding to the target embodied intelligence agent; and determining task evaluation information corresponding to the target embodied intelligence agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the target embodied intelligence agent.

[0020] Fourthly, embodiments of this application provide an embodied intelligence model training system, comprising: a second receiving module for receiving observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to an embodied intelligence agent; and a second processing module for inputting the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the embodied intelligence agent into the embodied intelligence model to be trained, and performing the following processing based on the embodied intelligence model to be trained: performing joint modeling based on the observation data and motion control input data corresponding to the embodied intelligence agent to obtain joint feature information corresponding to the embodied intelligence agent; and performing at least one of the following processing based on the embodied intelligence model to be trained: based on the observation data and motion control input data corresponding to the embodied intelligence agent... The system comprises: a joint feature information corresponding to the embodied intelligent agent; a future observation information corresponding to the embodied intelligent agent based on the motion state data and joint feature information corresponding to the embodied intelligent agent; and a task evaluation information corresponding to the embodied intelligent agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligent task information corresponding to the embodied intelligent agent. A loss determination module is used to determine a training loss based on at least two of the future observation information, future motion state information, joint feature information, and task evaluation information corresponding to the embodied intelligent agent and the corresponding target information. A parameter update module is used to update at least one model parameter of the embodied intelligent model to be trained based on the training loss.

[0021] Fifthly, embodiments of this application provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of any of the above methods.

[0022] Sixthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0023] In a seventh aspect, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the steps of any of the above methods.

[0024] This application provides an embodied intelligence data processing method and system, a model training method, and related equipment. By receiving observation data, motion control input data, motion state data, and embodied intelligence task information, and inputting multiple types of data into the trained embodied intelligence model, the trained embodied intelligence model can simultaneously utilize environmental information, control drive information, and ontological state information for analysis. This improves the completeness of the representation of the embodied intelligence task process and reduces the analytical bias caused by relying solely on single observation information. By jointly modeling based on observation data and motion control input data to obtain joint feature information, the correlation between environmental changes and action drives can be established, improving the trained embodied intelligence model's ability to understand the task execution context. By determining future observation information based on joint feature information, the potential environmental evolution results that the target embodied intelligence agent may face in the future can be characterized in advance, providing a forward-looking basis for subsequent task analysis. By determining future motion state information based on motion state data and joint feature information, the evolution trend of the target embodied intelligence agent's ontological actions at subsequent moments can be simultaneously depicted, improving the trained embodied intelligence model's ability to predict the action process and state changes of the embodied intelligence agent. By determining task evaluation information based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information, a comprehensive judgment can be made on task execution from multiple dimensions, including environmental results, ontology state, and task objectives. This improves the accuracy and stability of task evaluation results and reduces the reliance on a single information source in the task evaluation process. Integrating the determination of future observation information, future motion state information, and task evaluation information into the same trained embodied intelligent model reduces the process fragmentation caused by the separate deployment and processing of multiple independent modules, improving the consistency and processing efficiency of the embodied intelligent task analysis process. This embodiment addresses the problem in related technologies where the embodied intelligent task processing process tends to over-rely on a single observation result and struggles to simultaneously consider environmental evolution analysis, ontology motion trend analysis, and task execution effect judgment. It adopts a concept of jointly modeling observation data and motion control input data, and combining motion state data and embodied intelligent task information to determine future observation information, future motion state information, and task evaluation information separately. This improves the joint representation capability, forward-looking analysis capability, and task evaluation accuracy of the embodied intelligent task process. Attached Figure Description

[0025] The embodiments of this application are further described below with reference to the accompanying drawings and specific implementation details.

[0026] Figure 1 This is a flowchart illustrating an embodied intelligence data processing method provided in an embodiment of this application.

[0027] Figure 2This is a flowchart illustrating an embodied intelligence model training method provided in an embodiment of this application.

[0028] Figure 3 This is a flowchart of an embodied intelligent data processing embodiment provided in this application.

[0029] Figure 4 This is a structural block diagram of an embodied intelligent data processing system provided in an embodiment of this application.

[0030] Figure 5 This is a structural block diagram of an embodied intelligent model training system provided in an embodiment of this application.

[0031] Figure 6 This is a structural block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this application.

[0033] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0034] With the development of artificial intelligence, robotics, multimodal perception, and data-driven modeling technologies, embodied intelligence has gradually become an important research direction in the fields of intelligent control, intelligent interaction, and autonomous decision-making. Embodied intelligence emphasizes the perception, movement, interaction, and task execution capabilities of intelligent agents in real or simulated environments. It can analyze, predict, and judge subsequent actions and execution results by receiving environmental information, its own state information, and task objective information. Embodied intelligent agents can be robotic arms, mobile robots, dual-arm robots, humanoid robots, or other intelligent entities with environmental perception and action execution capabilities. Data processing and task modeling around embodied intelligent agents have led to research and application demands covering multiple directions, including environmental perception, motion control, state estimation, task understanding, future prediction, and task evaluation.

[0035] In related technologies, to enhance the performance of embodied agents in complex task scenarios, it is often necessary to comprehensively analyze the current environmental state, received control inputs, the agent's motion state, and the task objective. For example, in tasks such as grasping, carrying, assembling, opening doors, navigating, obstacle avoidance, and interactive services, the embodied agent not only needs to understand the state of the target object, obstacles, and interactive area in the current environment, but also needs to combine the current control drive and its own motion state to infer the subsequent environmental change trends and the trends of the agent's actions, and further determine whether the current task has a high probability of completion or to what extent it has been completed. Therefore, how to effectively process the multi-source heterogeneous data of embodied agents during task execution and form a data processing mechanism for future state prediction and task evaluation has become one of the important issues in embodied intelligence technology.

[0036] In related technologies, common data processing methods for embodied intelligent agents often focus on the individual utilization of a single type of information. For example, some methods primarily rely on observation data to identify, detect, or predict scene states to understand target objects and spatial relationships within the environment; others primarily use motion control input data or ontology motion state data to model the embodied intelligent agent's motion trajectory, posture changes, or control results; still others use task objectives or task labels as separate inputs to train task classification models, reward models, or success rate judgment models. While these methods can be effective in their respective areas of focus, when faced with embodied intelligent tasks, they often struggle to simultaneously consider the intrinsic relationships between environmental changes, ontology motion evolution, and task objective achievement.

[0037] Furthermore, in some related solutions, environmental perception, action prediction, and task evaluation are often separated into relatively independent processing modules. For example, one part of the model is responsible for processing observational information such as images, videos, or point clouds to predict future scene changes; another part is responsible for predicting future motion outcomes based on joint states, end-effector states, or control commands; and yet another part is responsible for outputting task evaluation results based on task labels, execution results, or reward rules. Because these processing processes are separated from each other, the correlation between different information is difficult to fully model within a unified representation space, which can easily lead to insufficient expression of the coupling relationship between environmental information, control information, and state information, thereby affecting the accuracy of future prediction results and task judgment results.

[0038] Furthermore, in embodied intelligence tasks, observational data often only reflects the external environmental state perceived by the embodied agent at a specific moment or within a certain time period, and is insufficient to express why the embodied agent performs a specific action or how subsequent actions will continue to evolve. Motion control input data can reflect the intention behind the action, but without correlation modeling with environmental observations, it is difficult to accurately determine the impact of control inputs on environmental evolution. Motion state data can reflect the current ontological state and execution feedback of the embodied agent, but without joint analysis of future environmental states and task objectives, it is difficult to accurately assess task completion. Therefore, relying solely on a single data source, or simply splicing different data sources, often fails to meet the comprehensive needs of complex embodied intelligence tasks for future observation prediction, future motion state prediction, and task evaluation.

[0039] On the other hand, as embodied intelligence technology is increasingly applied to real-world scenarios, the types of tasks are constantly increasing, and task constraints are becoming more stringent. Related data processing methods not only need to understand the current state but also require predictive capabilities for the future and evaluation capabilities for the task at hand. For example, in grasping and handling tasks, it is necessary to predict whether the target object will be successfully grasped and subsequently placed in the designated location; in assembly tasks, it is necessary to predict whether the components can be aligned and inserted into the target position; in movement and navigation tasks, it is necessary to predict whether the embodied agent will reach the target area and meet obstacle avoidance constraints. These application requirements indicate that embodied intelligence data processing methods should not only remain at the level of current state analysis but should also be able to combine environmental information, action-driven information, agent state information, and task information to collaboratively infer subsequent environmental changes, agent action changes, and task results.

[0040] In summary, it is necessary to propose a new embodied intelligence data processing method to improve the problem that the embodied intelligence task processing process in related technologies tends to rely too much on single observation results and is difficult to simultaneously take into account environmental evolution analysis, ontological motion trend analysis and task execution effect judgment.

[0041] See Figure 1 , Figure 1 This is a flowchart illustrating an embodied intelligence data processing method provided in an embodiment of this application.

[0042] To address the shortcomings of related technologies, this application provides an embodied intelligence data processing method, which includes steps S101 to S102.

[0043] Step S101: Receive the observation data, motion control input data, motion state data and embodied intelligent task information corresponding to the target embodied intelligent agent.

[0044] Step S102: Input the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligent agent into the trained embodied intelligent model, and perform the following processing based on the embodied intelligent model: Joint modeling is performed based on the observation data and motion control input data corresponding to the target embodied intelligent agent to obtain the joint feature information corresponding to the target embodied intelligent agent; Based on the joint feature information corresponding to the target embodied intelligent agent, the future observation information corresponding to the target embodied intelligent agent is determined; Based on the motion state data and joint feature information corresponding to the target embodied intelligent agent, the future motion state information corresponding to the target embodied intelligent agent is determined. Based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information corresponding to the target embodied intelligent agent, the task evaluation information corresponding to the target embodied intelligent agent is determined.

[0045] An embodied intelligent agent can refer to an intelligent entity possessing environmental perception, motion execution, and task interaction capabilities. It can be a robotic arm, a dual-arm robot, a mobile manipulation robot, a humanoid robot, a wheeled service robot, a legged robot, or other intelligent equipment with end effectors and sensing units. Unlike purely software-based intelligent models, embodied intelligent agents not only need to process environmental information but also need to combine their own motion control inputs and physical motion state to complete task decisions and execution. Therefore, in this embodiment, the embodied intelligent agent is both the primary source of observation data, motion control input data, and motion state data, and the corresponding object of future observation information, future motion state information, and task evaluation information.

[0046] Observational data refers to data used to characterize the external environment and the states of interacting objects within an embodied intelligent agent. Observational data can include one or more of the following: single-view image data, multi-view image data, video sequence data, depth map data, point cloud data, tactile perception data, force perception data, semantic segmentation results, object detection results, and scene reconstruction results. The role of observational data is to reflect the spatial layout, target locations, obstacle distribution, contact states, and changes in the task scenario within the current environment, thereby providing an input basis for subsequent joint feature extraction, future observation prediction, and task evaluation.

[0047] Motion control input data refers to the control information used to drive the embodied agent to perform actions. It can include one or more of the following: end-effector control input, joint control input, trajectory control input, velocity control input, acceleration control input, gripper opening / closing control input, and base movement control input. Motion control input data reflects "how the embodied agent is instructed to move," thus reflecting the action intention, execution direction, and action-driving trend. Using motion control input data in conjunction with observation data for joint modeling helps establish the correspondence between environmental changes and action driving forces, thereby improving the problem that static observations alone are insufficient to accurately characterize the task evolution process.

[0048] Motion state data refers to data used to characterize the motion state of an embodied intelligent agent, and may include one or more of the following: end-effector position, end-effector posture, end-effector velocity, end-effector acceleration, joint position, joint angle, joint velocity, joint acceleration, gripper state, base posture, and limb coordination state. Motion state data can include current motion state data and historical motion state data. Current motion state data may include current posture data corresponding to at least one of the end-effector and joints, while historical motion state data may include historical posture data corresponding to at least one of the end-effector and joints. It can be seen that motion state data can correspond to the end-effector motion space and / or joint motion space. Motion state data differs from motion control input data; motion control input data focuses more on driving commands (or control commands), while motion state data focuses more on the actual ontological state feedback after execution. Determining future motion state information based on motion state data and joint feature information enables the embodied intelligent model to understand not only how the external environment changes but also how the embodied intelligent agent itself evolves, thereby improving the completeness of the modeling of the task execution process.

[0049] Embodied intelligent task information refers to information used to characterize the current task objective, task semantic constraints, or task completion conditions. It can include one or more of the following: task type identifier, task text description, target object identifier, target position constraint, operation sequence constraint, interaction relationship constraint, success determination rules, and phased goal definition. For example, embodied intelligent task information could be "grab a cup from the table and place it on a tray," "pull a drawer open to a specified degree," "rotate a valve to a target angle," or "move a target object to a designated area." The introduction of embodied intelligent task information transforms task evaluation information from merely a mechanical judgment of action results into a purposeful assessment that combines specific task objectives with future observation information, future motion state information, and joint feature information.

[0050] A trained embodied intelligence model refers to a data processing model that has undergone sample learning, parameter optimization, and capability solidification. It receives multi-source input data and outputs future observation information, future motion state information, and task evaluation information. A trained embodied intelligence model can be one or more combined implementations of neural network models, temporal prediction models, multi-branch joint modeling models, multi-task learning models, expert collaborative models, and generative prediction models. The key to a trained embodied intelligence model lies not in the specific network name, but in its ability to jointly model environmental evolution, ontology state evolution, and task completion within a unified framework, thereby reducing the process fragmentation caused by separate processing of multiple modules.

[0051] Joint modeling refers to a modeling approach that correlates and unifies the representation of two or more data sources with different semantics or functions. In this embodiment, joint modeling integrates observation data and motion control input data into the same processing chain for collaborative analysis. Observation data can be used to characterize the environmental state of the embodied agent, the state of the target object, and changes in the interaction scene. Motion control input data can be used to characterize the action-driven information, control intentions, or trajectory requirements assigned to the embodied agent. These two data points reflect two different levels of information: "what the external environment is" and "how the embodied agent is required to move." The implementation methods of joint modeling can include one or more of the following: feature splicing modeling, attention interaction modeling, conditional modulation modeling, cross-modal alignment modeling, temporal correlation modeling, latent space fusion modeling, and multi-branch collaborative modeling. Through joint modeling, a correspondence between observation results and action drives can be established, enabling the obtained joint feature information to not only characterize the current environmental state but also the influence trend of action drives on environmental evolution. This enhances the understanding of the embodied intelligent task process and addresses the problem that it is difficult to accurately depict subsequent environmental changes and task context based solely on single observation information.

[0052] Joint feature information refers to intermediate representations obtained through joint modeling that characterize the current environmental state and its relationship with action-driven processes. Joint feature information can be one of the following: high-dimensional feature vectors, feature maps, temporal feature sequences, cross-modal fusion features, or latent space representations. Compared to the original observation data, joint feature information has extracted key task-related information, such as target object location, scene interaction area, obstacle distribution, contact trends, and operational context. Compared to the original motion control input data, joint feature information also reflects the trend of action-driven processes influencing environmental changes. Therefore, joint feature information is not only an important basis for predicting future observation information but also a crucial support for predicting future motion state information and determining task evaluation information.

[0053] Future observation information refers to environmental information inferred from current joint feature information, used to characterize the embodied agent's potential reception of environmental information in subsequent moments. Future observation information can include one or more of the following: future image results, future video clips, future depth distribution, future point cloud state, future target object positional relationships, future scene layout changes, and future contact state results. Essentially, future observation information provides a forward-looking characterization of environmental evolution trends, reflecting how the task scenario might change given a given action and current environmental context. For example, whether a target object will be picked up, whether a drawer will continue to open, or whether a door will rotate to a specified angle can all be reflected through future observation information.

[0054] Future motion state information refers to information inferred from motion state data and joint feature information, used to characterize the subsequent motion evolution of an embodied agent. Future motion state information can include one or more of the following: future end-effector state information, future joint state information, future base posture information, future gripper state information, and future limb coordination state information. As an example, future motion state information can include future posture information corresponding to at least one of the end-effector and joints. Future motion state information reflects the action state that the embodied agent may reach in subsequent moments, rather than the changes in the external environment itself. By combining joint feature information for future motion state prediction, the model can understand both environmental changes and the agent's motion evolution simultaneously, thereby overcoming the shortcomings of predicting only from the environmental perspective while ignoring the continuity of agent actions.

[0055] The task evaluation information is determined based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information. This embodies a multi-information joint judgment mechanism. The phrase "at least two" indicates that the determination of task evaluation information is not limited to a fixed combination of inputs. Instead, it allows for flexible selection of two, three, or four of the aforementioned information for comprehensive analysis, depending on the task type, model design, or data availability. For example, task evaluation information can be determined based on future observation information and embodied intelligent task information, or it can be determined based on future motion state information and joint feature information, or it can be determined based on a combination of future observation information, future motion state information, and embodied intelligent task information.

[0056] Task evaluation information refers to information used to characterize the performance, completion level, or probability of task achievement of an embodied agent. This information may include one or more of the following: reward value, task completion degree, task progress information, task success probability, task score, task completion label, and task stage judgment results. By combining at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information, the analysis examines whether the current task is progressing towards the target direction and whether it is ultimately likely to achieve the target result, from three levels: environmental outcome, ontology evolution, and task objective. This approach allows task evaluation to move beyond relying solely on a single observation result or a single state indicator, thereby improving the accuracy, stability, and task adaptability of the evaluation results.

[0057] The method for determining task evaluation information corresponds to a task judgment logic based on multi-source information fusion. Its core lies not in simply outputting a numerical result, but in purposefully interpreting and evaluating future observation results and future ontological state results according to the task objectives. In specific implementation, methods such as classification, regression, ranking, comparison, matching, or joint judgment can be used to generate information results that characterize the task execution effect. For example, in a grasping task, future observation information can be used to determine whether the target object has entered the grasping area; combined with future motion state information, it can be determined whether the end-effector posture meets the grasping conditions; and combined with embodied intelligence task information, it can be determined whether the action conforms to the target object and target position requirements, ultimately obtaining the task success probability or task completion score. This method of determining task evaluation information helps to improve the problem of separation between environmental prediction and task judgment in related solutions.

[0058] In this embodiment, observation data, motion control input data, motion state data, and embodied intelligence task information of the target embodied intelligent agent during task execution are jointly input into the trained embodied intelligent model. This allows the trained embodied intelligent model to no longer rely solely on the current observation image for subsequent judgments, but instead combine the control-driven information already received by the target embodied intelligent agent with its current motion state to jointly model the behavioral evolution process of the target embodied intelligent agent. Specifically, the trained embodied intelligent model first performs joint modeling based on the observation data and motion control input data, extracting joint feature information that characterizes the changing trends of the external environment and the relationship between action drives. Subsequently, based on the joint feature information, it infers the environmental observation results that the target embodied intelligent agent may encounter in the future, obtaining future observation information. Simultaneously, the trained embodied intelligent model also infers the motion evolution results of the target embodied intelligent agent in the future based on the motion state data and joint feature information, obtaining future motion state information. Building upon this foundation, the trained embodied intelligence model further integrates at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligence task information, to analyze the current task execution effect, task completion trend, or task completion degree, thereby obtaining task evaluation information. Thus, the trained embodied intelligence model can simultaneously complete environmental evolution prediction, ontology motion trend prediction, and task execution result evaluation within a unified data processing flow, thereby forming a joint reasoning mechanism oriented towards the embodied intelligence task process.

[0059] This embodiment receives observation data, motion control input data, motion state data, and embodied intelligence task information, and inputs these multiple types of data into the trained embodied intelligence model. This allows the trained model to simultaneously utilize environmental information, control-driven information, and ontological state information for analysis, improving the completeness of the representation of the embodied intelligence task process and reducing the analytical bias caused by relying solely on single observation information. By jointly modeling based on observation data and motion control input data to obtain joint feature information, the correlation between environmental changes and action drives can be established, enhancing the trained embodied intelligence model's understanding of the task execution context. By determining future observation information based on joint feature information, the potential environmental evolution outcomes faced by the target embodied intelligence agent can be characterized in advance, providing a forward-looking basis for subsequent task analysis. By determining future motion state information based on motion state data and joint feature information, the evolution trend of the target embodied intelligence agent's ontological actions at subsequent moments can be simultaneously depicted, improving the trained embodied intelligence model's predictive ability for the embodied intelligence agent's action process and state changes. By determining task evaluation information based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligence task information, a comprehensive assessment of task execution can be made from multiple dimensions, including environmental outcomes, ontological state, and task objectives. This improves the accuracy and stability of task evaluation results and reduces the reliance on a single information source in the task evaluation process. Furthermore, by integrating the determination of future observation information, future motion state information, and task evaluation information into the same trained embodied intelligence model, the fragmentation caused by deploying and processing multiple independent modules separately can be reduced, improving the consistency and processing efficiency of the embodied intelligence task analysis process.

[0060] This embodiment addresses the problem in related technologies that the embodied intelligence task processing process tends to over-rely on single observation results and struggles to simultaneously consider environmental evolution analysis, ontological motion trend analysis, and task execution effect judgment. It adopts a concept of jointly modeling observation data and motion control input data, and combining motion state data and embodied intelligence task information to determine future observation information, future motion state information, and task evaluation information, thereby improving the joint representation capability, forward-looking analysis capability, and task evaluation accuracy of the embodied intelligence task process.

[0061] In some embodiments, the observation data may include multi-view observation data. In some embodiments, the observation data may include current observation data.

[0062] Observational data can originate from sensing devices such as cameras, depth sensors, lidar, tactile sensors, force sensors, millimeter-wave sensors, and environmental sensing arrays, or from intermediate results obtained after further processing of raw sensing results. Multi-view observation data refers to data acquired from two or more different observation azimuths, positions, orientations, or devices. Multi-view observation data can include at least two of the following: front-view images, side-view images, top-view images, oblique-view images, end-effector view images (such as left-hand and right-hand camera images), fixed-view images, and moving-view images. It can also include multiple depth perspective results, multiple point cloud observation results, or multi-source perspective observation results formed by combining images and depth information. The significance of using multi-view observation data lies in its ability to supplement the information representation of target areas, operational areas, and obstacle areas from different spatial directions. This improves the information loss problems caused by occlusion, incomplete field of view, partial invisibility of targets, or difficulty in accurately judging spatial relationships under a single perspective, thereby enhancing the completeness of the representation of environmental states, target states, and interaction processes.

[0063] In some embodiments, multi-view observation data can be organized and used in a manner that combines time synchronization, spatial calibration, and viewpoint association. Time synchronization refers to aligning data collected from different perspectives to the same time reference, thereby ensuring that multiple perspectives reflect the environmental state at the same moment or during the same period. Spatial calibration refers to establishing spatial correspondences between different perspectives, enabling the correlation of target position, attitude information, and scene structure information from different perspectives. Viewpoint association refers to jointly analyzing key areas in multiple perspectives according to task requirements, such as simultaneously focusing on the manipulated object, the end effector, and the task target area. Multi-view observation data not only provides richer scene information but also provides a more reliable data foundation for subsequent joint feature information extraction, thereby improving the accuracy and stability of future observation information prediction, future motion state information prediction, and task evaluation information determination.

[0064] In some embodiments, the motion control input data may include at least one of end-effector control input information and joint control input information.

[0065] In some embodiments, the motion state data may include at least one of end-effector state information and joint state information.

[0066] In some embodiments, the motion state data may include current motion state data and historical motion state data.

[0067] Motion control input data may include one or more of the following: position control commands, posture control commands, speed control commands, force control commands, impedance control commands, trajectory tracking commands, grasping control commands, base movement control commands, and limb coordination control commands.

[0068] End-effector control input information refers to data used to control the end effector of an embodied intelligent agent, primarily characterizing the target motion requirements of the end effector in Cartesian space or task space. End-effector control input information can include one or more of the following: end-effector position target information, end-effector posture target information, end-effector velocity target information, end-effector acceleration target information, end-effector force control information, end-effector torque control information, gripping opening and closing control information, contact control information, and tracking path control information. For example, in a grasping task, end-effector control input information can correspond to the pose control command for the gripper to approach the target object; in a insertion task, end-effector control input information can correspond to the control command for the end effector to move along the target hole direction and maintain a defined posture. End-effector control input information can directly describe the execution actions most relevant to the task interaction, thus providing stronger semantic constraints for subsequent joint feature extraction and future observation prediction.

[0069] Joint control input information refers to data used to control the execution parts of each joint of an embodied intelligent agent, primarily used to characterize the target driving requirements of each joint within the joint space. Joint control input information can include one or more of the following: joint angle target value, joint position target value, joint velocity target value, joint acceleration target value, joint torque control information, joint current control information, joint synchronization control information, and joint sequence driving information. Compared to end-effector control input information, joint control input information focuses more on the direct driving of the degrees of freedom within the embodied intelligent agent itself. For example, when a robotic arm performs a handling task, joint control input information can correspond to a control sequence in which each joint is sequentially adjusted to a specified angle; in a dual-arm collaborative task, joint control input information can also be used to characterize the collaborative driving relationship between different limbs. By introducing joint control input information, the motion evolution process can be characterized at the ontological control level, thereby improving the problem that it is difficult to accurately infer the subsequent motion trend of the embodied intelligent agent based solely on external observation data.

[0070] Motion state data may include one or more of the following: end-effector state information, joint state information, base state information, gripper state information, limb coordination state information, contact state information, velocity state information, and acceleration state information.

[0071] End-effector state information refers to data characterizing the current actual state of the end effector of an embodied agent. End-effector state information can include one or more of the following: end-effector position, end-effector posture, end-effector velocity, end-effector acceleration, end-effector force state, end-effector torque state, end-effector contact state, gripper opening / closing state, and tool center point state. For example, in a grasping scenario, end-effector state information can characterize whether the gripper has contacted the target object and whether the gripper opening / closing range meets the grasping conditions; in an assembly scenario, end-effector state information can characterize whether the end tool has been aligned with the target assembly position. End-effector state information directly reflects the core interaction state between the embodied agent and the task object, therefore it has high judgment value in predicting future motion state information and determining task evaluation information, helping to improve the problem of missing interaction details caused by relying solely on far-field observation results.

[0072] Joint state information refers to data characterizing the current actual motion state of each joint of an embodied agent. Joint state information can include one or more of the following: joint angle, joint position, joint velocity, joint acceleration, joint torque, joint current, joint damping state, joint compliance state, and joint coordination relationship information. Joint state information reflects the internal degree-of-freedom configuration of the embodied agent, characterizing whether the current action is in the lifting, extension, grasping, or retraction phase, and whether the expected coordination relationship has been achieved between different joints. Compared to end-effector state information, joint state information provides a more granular description of the internal motion structure of the agent. Therefore, in some scenarios, even if external observation is obstructed or the end-effector is partially invisible, the action trend and subsequent evolution direction of the embodied agent can still be inferred based on the joint state information, thereby improving the robustness of future motion state prediction and task evaluation.

[0073] The motion control input data may include at least one of end-effector control input information and joint control input information, and the motion state data may include at least one of end-effector state information and joint state information. This indicates that the relevant data types are not limited to a single fixed form, but can be flexibly selected according to the embodied agent type, control architecture, and task scenario. For example, for robotic arm applications primarily involving task space operations, end-effector control input information and end-effector state information can be primarily used; for body control scenarios primarily involving joint-level control, joint control input information and joint state information can be primarily used; for complex tasks requiring coordination between end-effector execution effects and body degrees of freedom, end-effector control input information, joint control input information, end-effector state information, and joint state information can be used simultaneously. Using the expression "at least one" helps improve the adaptability of this embodiment to different hardware platforms, different control levels, and different task types.

[0074] Motion control input data and motion state data can form a "control-driven-state feedback" correspondence. Motion control input data describes the desired action of the embodied agent, while motion state data reflects the actual state feedback of the embodied agent after executing the desired action. Based on this relationship, the embodied intelligence model can use motion control input data to understand the action-driven intent, and motion state data to understand the current ontological state and execution deviations. Furthermore, it can combine observation data to extract joint feature information, thereby jointly predicting future observation information and future motion state information. For example, when the end-effector control input information indicates that the gripper is approaching the target object, but the end-effector state information indicates that the gripper has not yet reached the predetermined position, the embodied intelligence model can determine that the current action is still in the approaching stage, rather than the completed grasping stage. Through this control logic, action commands, ontological feedback, and environmental changes can be unified into the same reasoning chain, thereby improving problems such as incomplete understanding of the task process, inaccurate judgment of state evolution, and large fluctuations in task evaluation results.

[0075] In some embodiments, the task evaluation information may be determined based on the future observation information and the embodied intelligent task information.

[0076] In some embodiments, the task evaluation information may include at least one of the following: reward value, task success probability, task score, task completion rate, task completion tag, and task progress information.

[0077] In this embodiment, task evaluation information may not be a single numerical value, but may be numerical results, probabilistic results, label-based results, stage-based results, or combined results. For example, task evaluation information can be used to characterize whether the current action is moving towards the task goal, whether the future state meets the task completion conditions, whether the current task result meets the preset standard, and the priority differences between different candidate action schemes. Task evaluation information can establish a correspondence between future observation results and task goal constraints, so that the embodied intelligence model can not only predict "how the environment will change", but also further determine "whether the change meets the task requirements", thereby improving the problem of only being able to predict the state but having difficulty making task-level judgments. Future observation information can reflect the possible changes in target objects, obstacles, contact states, and spatial relationships in the subsequent scene driven by the embodied intelligence agent under the given action. Embodied intelligence task information can provide target basis and semantic constraints for task evaluation, so that task evaluation information is no longer just judging the environmental changes themselves, but can analyze whether the future observation results meet the task completion conditions in combination with specific task requirements.

[0078] Task evaluation information is determined based on future observation information and embodied intelligent task information. For example, future observation information is first used to characterize the possible state of the environment at subsequent times. Then, embodied intelligent task information is used to provide the target conditions, object conditions, or spatial conditions that need to be met to complete the task. Finally, the two are matched, compared, classified, regressed, or jointly judged to obtain task evaluation information. For example, in a transportation task, future observation information can be used to determine whether the target object appears in the target placement area. Then, combined with embodied intelligent task information, it can be determined whether the area is the preset task area, thereby obtaining the task success probability or task completion label. In a splicing task, future observation information can be used to determine whether the splice is aligned with the target hole. Then, combined with embodied intelligent task information, it can be determined whether the splicing depth requirement has been met, thereby obtaining the task completion degree or task score. Using this processing logic helps to directly transform environmental prediction results into task-level evaluation results, thereby improving the problem of the separation between environmental modeling and task judgment.

[0079] Reward values ​​are quantified numerical values ​​used to characterize the quality of an embodied agent's current action or future task outcome. Reward values ​​can be discrete or continuous; positive, negative, or range-bound; and can be single-step or cumulative rewards. For example, in a grasping task, a higher reward value corresponds to a stable hold on the target object, while a lower or negative reward value corresponds to a drop. In a navigation task, a positive reward value corresponds to correct movement towards the target area, while a lower reward value corresponds to deviation from the target direction. Reward values, in numerical form, reflect the quality of task execution and the degree of proximity to the target, thus providing a quantitative basis for strategy selection, action evaluation, model training, or task prioritization.

[0080] Task success probability refers to probabilistic information characterizing the likelihood of an embodied agent completing a target task. It can be a probability value between zero and one, or a percentage, representing the degree to which the current task trajectory, current action plan, or future evolutionary outcome is likely to achieve the task objective. For example, in a placement task, when the target object has entered the target area but has not yet been fully released, the task success probability can be at a moderate or high level; in an assembly task, when parts are aligned but not yet inserted, the task success probability can be dynamically adjusted according to changes in alignment accuracy. Task success probability provides a measure of uncertainty for task decision-making, enabling embodied intelligent models to output not only hard-to-determine "whether it succeeded" but also soft-to-determine "how likely it is to succeed," thereby improving the situation where task evaluation results are too rigid.

[0081] Task scoring refers to rating-based information used to comprehensively characterize the quality of task execution. Task scores can use preset score ranges, such as normalized scores between 0 and 100, 0 and 10, or 0 and 1. They can also employ multi-dimensional scoring methods, such as simultaneously outputting stability, efficiency, and accuracy scores. Task scoring can comprehensively consider multiple evaluation factors, such as whether the task was completed, completion speed, path deviation, target object posture deviation, interaction stability, whether collisions occurred, and whether timing requirements were met. By outputting task scores, the task results of embodied agents can be expanded from a binary judgment of "whether it was completed" to a fine-grained judgment of "how well it was completed," thus providing richer evaluation criteria for action plan comparison, strategy selection, and task quality analysis.

[0082] Task completion rate refers to information characterizing the extent to which a target task has been completed. It can be expressed as a proportion, percentage, stage percentage, or continuous progress value. For example, in a drawer-opening task, completion rate can be determined by the ratio between the distance the drawer opens and the target opening distance; in a grasping and moving task, completion rate can be determined by the proportion of the object's movement from its original position to its target position; and in an assembly task, completion rate can be determined by the proportion of insertion depth to the target depth. Task completion rate can reflect the progress of a task in a continuous manner, enabling embodied intelligence models to more finely distinguish between different states such as "not yet started," "partially completed," "nearly completed," and "fully completed," thereby improving the problem of insufficient information when judging tasks solely based on binary success or failure results.

[0083] Task completion labels refer to label-type information used to characterize the category of task outcome. Task completion labels can be binary or multi-class labels. For example, task completion labels can include outcome categories such as "success," "failure," "partial success," "interrupted," "not met," and "pending confirmation"; or they can include more granular task outcome categories such as "successful capture," "failed capture," "successful placement," "placement offset," and "incomplete insertion." Task completion labels clearly identify task outcomes through discrete categories, facilitating subsequent execution result recording, strategy selection, anomaly sample identification, and training supervision. Compared to continuous results such as reward values ​​and task scores, task completion labels are more suitable for directly providing a clear judgment of task outcome.

[0084] Task progress information refers to information used to characterize the current stage or sequential position of a task during execution. Task progress information can include one or more of the following: stage number, stage name, stage probability, stage transition state, stage start marker, and stage end marker. For example, in a grasping and handling task, task progress information may include "approaching the target stage," "establishing contact stage," "completing the grasping stage," "performing the handling stage," and "completing the placement stage." In a door operation task, task progress information may include "approaching the handle stage," "establishing contact stage," "applying driving force stage," "door rotation stage," and "opening completion stage." Task progress information can describe the task execution status from a task flow perspective, enabling embodied intelligence models to identify the current stage of the task and generate task evaluation information accordingly. This improves the lack of stage interpretation capabilities when conducting overall evaluations of long-sequence tasks.

[0085] Task evaluation information may include at least one of the following: reward value, task success probability, task score, task completion degree, task completion tag, and task progress information. This means that task evaluation information is not limited to a fixed single result, but can flexibly output one result or a combination of multiple results depending on the task type, model capabilities, and deployment requirements. For example, in reinforcement learning scenarios, a reward value can be output; in online policy selection scenarios, the task success probability and task score can be output; in industrial quality inspection or execution record scenarios, the task completion tag and task progress information can be output; and in complex long-term tasks, task completion degree, task progress information, and task score can be output simultaneously. This embodiment helps improve adaptability to different embodied intelligence tasks, different evaluation criteria, and different output requirements, and reduces the problem of limited protection scope caused by overly narrow definition of task evaluation result formats.

[0086] See Figure 2 , Figure 2 This is a flowchart illustrating an embodied intelligence model training method provided in an embodiment of this application.

[0087] This application embodiment also provides an embodied intelligence model training method, which includes steps S201 to S204.

[0088] Step S201: Receive observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the embodied intelligent agent.

[0089] Step S202: Input the observation data, motion control input data, motion state data and embodied intelligence task information corresponding to the embodied intelligent agent into the embodied intelligent model to be trained, so as to perform the following processing based on the embodied intelligent model to be trained: perform joint modeling based on the observation data and motion control input data corresponding to the embodied intelligent agent to obtain the joint feature information corresponding to the embodied intelligent agent; Furthermore, based on the embodied intelligence model to be trained, perform at least one of the following processes: determine the future observation information corresponding to the embodied intelligence agent based on the joint feature information corresponding to the embodied intelligence agent; determine the future motion state information corresponding to the embodied intelligence agent based on the motion state data and joint feature information corresponding to the embodied intelligence agent; determine the task evaluation information corresponding to the embodied intelligence agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the embodied intelligence agent.

[0090] Step S203: Based on at least two of the following: future observation information, future motion state information, joint feature information, and task evaluation information corresponding to the embodied intelligent agent, and the corresponding target information, determine the training loss.

[0091] Step S204: Based on the training loss, update at least one model parameter of the embodied intelligence model to be trained.

[0092] In this embodiment, observation data, motion control input data, motion state data, and embodied intelligence task information from the embodied agent during task execution are jointly input into the embodied intelligence model to be trained. This allows the model to learn the correlation between environmental evolution, ontological motion evolution, and task result evaluation within a unified training framework. Specifically, the embodied intelligence model first performs joint modeling based on the observation data and motion control input data, extracting joint feature information that characterizes the environmental state and action-driven relationships. Based on this, the model further performs at least one prediction or evaluation process, including determining future observation information based on the joint feature information, determining future motion state information based on motion state data and joint feature information, and determining task evaluation information based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligence task information. Subsequently, at least two of the following—future observation information, future motion state information, joint feature information, and task evaluation information—are compared with the corresponding target information to determine the training loss. Then, based on the training loss, at least one model parameter of the embodied intelligence model to be trained is updated, so that the embodied intelligence model to be trained can gradually learn the environmental change law, the self state change law and the task completion judgment law driven by the embodied intelligence agent's actions, and improve the accuracy of future observation prediction, future motion state prediction and task evaluation.

[0093] This embodiment receives observation data, motion control input data, motion state data, and embodied intelligence task information, and inputs multiple types of heterogeneous data into the embodied intelligence model to be trained. This enables the model to simultaneously learn the correlation between environmental information, action-driven information, and ontology state information during the training phase, improving the completeness of model training and task adaptability. By jointly modeling based on observation data and motion control input data to obtain joint feature information, a correspondence between environmental changes and action-driven behavior can be established, enhancing the embodied intelligence model's understanding of the task process context. By performing at least one of the following processes based on the embodied intelligence model to determine future observation information, future motion state information, and task evaluation information, the training process can be compatible with one or more learning objectives in environmental prediction, ontology state prediction, and task result evaluation, improving the flexibility and applicability of the training method. By determining the training loss based on at least two of the following—future observation information, future motion state information, joint feature information, and task evaluation information—and the corresponding target information, the parameter optimization direction of the embodied intelligence model to be trained can be constrained from multiple dimensions, improving the consistency between the model's learning results and the real task process, and reducing training bias caused by a single supervision signal. By updating at least one model parameter of the embodied intelligence model to be trained based on training loss, the joint modeling ability of the embodied intelligence model to be trained for environmental evolution, ontology evolution and task achievement in embodied intelligence tasks can be continuously improved, thereby enhancing the accuracy, stability and generalization ability of the trained model in future observation prediction, future motion state prediction and task evaluation.

[0094] This embodiment addresses the problem in related technologies where the training process of embodied intelligence models tends to focus on a single prediction target and struggles to simultaneously consider environmental evolution learning, ontology motion learning, and task result evaluation learning. It adopts a concept that involves jointly inputting observation data, motion control input data, motion state data, and embodied intelligence task information, and combining at least one prediction result from future observation information, future motion state information, and task evaluation information for multidimensional loss constraints and model parameter updates. This improves the ability of the embodied intelligence model to jointly model the embodied intelligence task process, enhances the consistency between training results and the actual task execution process, and improves the predictive and task evaluation capabilities of the trained model.

[0095] In some embodiments, the embodied intelligence model training method may further include: determining the current embodied intelligence model to be trained as an embodied intelligence model when a preset end-training condition is met.

[0096] A trainable embodied intelligence model can refer to a model in the process of training, used to receive observation data, motion control input data, motion state data, and embodied intelligence task information, and continuously adjust model parameters based on training data. A trainable embodied intelligence model can be an initial model obtained just after initialization, or an intermediate model that has undergone several rounds of training but has not yet completed final parameter fixing. A trainable embodied intelligence model can include one or more of the following branches: joint modeling branch, state prediction branch, and task evaluation branch. It can also adopt a multi-module collaborative approach to achieve environmental evolution prediction, ontology motion state prediction, and task evaluation. A trainable embodied intelligence model is not a fixed model, but a dynamic training object that continuously optimizes under loss constraints, gradually approaching the target capability requirements.

[0097] The current trainable embodied intelligence model can refer to the trainable embodied intelligence model at a specific training moment, training epoch, training stage, or after a parameter update. The current trainable embodied intelligence model can be understood as an on-the-fly version of the model in the training process, whose parameter state reflects the training results completed at the current moment. For example, the current trainable embodied intelligence model could correspond to the model parameters after a batch of training or the model parameters after a complete training epoch.

[0098] In this embodiment, an embodied intelligence model refers to a model that has completed training or reached a predetermined training objective and can be used to formally execute embodied intelligence data processing tasks. An embodied intelligence model can be used to output one or more of the following based on observation data, motion control input data, motion state data, and embodied intelligence task information: future observation information, future motion state information, and task evaluation information. Compared to models that require training, embodied intelligence models place greater emphasis on their deployable, callable, and executable state after training. An embodied intelligence model can serve as one of the following: an online inference model, an offline evaluation model, a policy selection model, or a task analysis model. It can also serve as a foundational model for subsequent continuous learning or incremental training.

[0099] Preset training termination conditions refer to rules used to determine whether model training can be terminated. These conditions can be a single condition or a combination of multiple conditions. Preset training termination conditions may include at least one of the following: the number of training epochs reaches a preset number; the training loss decreases below a preset threshold; the validation set loss meets a stability condition; the future observation prediction accuracy reaches a preset requirement; the future motion state prediction accuracy reaches a preset requirement; the task evaluation accuracy reaches a preset requirement; the parameter update amplitude is less than a preset threshold for multiple consecutive epochs; the training time reaches a preset upper limit; or manual confirmation that the current model meets the application requirements. Preset training termination conditions can provide a clear basis for determining when model training should stop, thus enabling the training process to have executable termination logic and reducing the instability of training results caused by indefinitely extending or prematurely terminating the training process.

[0100] Whether a currently trainable embodied intelligence model has met one or more pre-defined criteria for stopping training during the training process can be determined by periodically checking training loss, validation metrics, prediction accuracy, evaluation accuracy, and parameter convergence. For example, after each round of training, one or more of the following can be evaluated: future observation prediction loss, future motion state prediction loss, and task evaluation loss. When a preset threshold is met or a stable condition is met for multiple consecutive rounds, the training termination condition can be determined to have been met.

[0101] Designating the current embodied intelligence model as an embodied intelligence model can mean retaining, registering, outputting, or deploying the current model state as the final training result after determining that training has reached the termination condition. This designation is a state switching and result solidification operation in the training process, and it is not limited to name changes. It can include saving the current model parameters, writing the current model to a model repository, registering the current model as a formal inference model, using the current model for subsequent embodied intelligence data processing tasks, or using the current model as a base model for subsequent incremental training.

[0102] This embodiment continuously performs data input, prediction output, loss calculation, and parameter updates during model training, and evaluates the current model training status at preset detection nodes. When the evaluation results indicate that the current model has met preset performance requirements, convergence requirements, epoch requirements, or time requirements, the training process triggers a termination condition, transforming the current model from a training object into an applicable model. This improves upon the problem of unclear end times in model training, providing a clear start point, iteration process, and end point for the model training process. This helps improve the standardization of the training process and provides a stable foundation for subsequent model deployment, evaluation, and reuse.

[0103] In some embodiments, the embodied intelligence model to be trained may include a joint modeling module, a state expert module, and an evaluation expert module.

[0104] The joint modeling module is used to perform joint modeling based on the observation data and motion control input data corresponding to the embodied intelligent agent, so as to obtain the joint feature information corresponding to the embodied intelligent agent, wherein the future observation information corresponding to the embodied intelligent agent is determined based on the joint feature information corresponding to the embodied intelligent agent.

[0105] The state expert module is used to determine the future motion state information of the embodied intelligent agent based on the motion state data and joint feature information corresponding to the embodied intelligent agent.

[0106] The evaluation expert module is used to determine the task evaluation information corresponding to the embodied intelligent agent based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information.

[0107] In this embodiment, the embodied intelligence model to be trained is divided into a joint modeling module, a state expert module, and an evaluation expert module. Each module undertakes different responsibilities: environmental evolution modeling, ontology motion evolution modeling, and task result evaluation, forming a collaborative processing mechanism within the same training and inference framework. Specifically, the joint modeling module performs joint modeling based on the observation data and motion control input data corresponding to the embodied intelligence agent, extracting joint feature information that characterizes the environmental state and action-driven relationships. This joint feature information is used to determine future observation information, representing the environmental changes the embodied intelligence agent may face in subsequent moments. The state expert module determines future motion state information based on motion state data and joint feature information, representing the motion state the embodied intelligence agent may reach in subsequent moments. The evaluation expert module comprehensively analyzes the achievement of the task objective based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information, determining task evaluation information. Through this modular division of labor, the embodied intelligence model to be trained can simultaneously complete environmental prediction, ontology state prediction, and task evaluation within a unified model structure, thereby achieving joint modeling of the embodied intelligence task process.

[0108] This embodiment establishes a joint modeling module, which performs joint modeling based on observation data and motion control input data to obtain joint feature information. This enables the establishment of a correlation between environmental state and action drive, enhancing the understanding of task context. By determining future observation information based on joint feature information, the joint modeling module can provide a forward-looking representation of the environmental evolution of the embodied agent in subsequent moments, improving the model's ability to predict environmental change trends. Furthermore, by setting up a state expert module, which determines future motion state information based on motion state data and joint feature information, the evolution trend of the embodied agent's actions can be simultaneously depicted, improving the modeling ability for the continuity of agent motion and changes in action structure. Finally, by setting up an evaluation expert module, which determines task evaluation information based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information, the task execution can be comprehensively judged from multiple dimensions, including environmental results, agent state, and task objectives, improving the accuracy and stability of task evaluation results. By adopting a modular structure in which joint modeling modules, state expert modules, and evaluation expert modules work together, environmental modeling, ontology modeling, and task evaluation functions can be integrated within the same trained embodied intelligent model. This reduces the process fragmentation caused by deploying multiple independent processing modules separately, and improves the consistency and processing efficiency of model training and application.

[0109] This embodiment addresses the problem in related technologies that embodied intelligent models struggle to simultaneously model environmental evolution, ontology motion state, and task result evaluation within a unified framework. It adopts the concept of dividing the embodied intelligent model to be trained into a joint modeling module, a state expert module, and an evaluation expert module, and then coordinating their processing. This enhances the embodied intelligent model's ability to jointly model the task process, improves the accuracy and stability of task evaluation results, and increases the overall consistency of model training and application.

[0110] In some embodiments, the training loss may include at least two of the following: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss.

[0111] In some embodiments, the future observation prediction loss is determined based on the future observation information corresponding to the embodied agent and its corresponding target future observation information.

[0112] In some embodiments, the future motion state prediction loss is determined based on the future motion state information corresponding to the embodied agent and the corresponding target future motion state information.

[0113] In some embodiments, the joint feature loss is determined based on the joint feature information corresponding to the embodied agent and its corresponding target joint feature information.

[0114] In some embodiments, the task evaluation loss is determined based on the task evaluation information corresponding to the embodied agent and its corresponding target task evaluation information.

[0115] Training loss refers to the quantitative information used to characterize the degree of difference between the model's output and the corresponding target information during the training process of an embodied intelligence model. Training loss can be a single loss term or a joint loss obtained by combining multiple (two or more) loss terms according to preset weights. The role of training loss is to transform different training objectives, such as future observation prediction, future motion state prediction, joint feature learning, and task evaluation learning, into a calculable and optimizable basis for parameter updates, enabling the embodied intelligence model to continuously adjust its parameters in the direction of reducing prediction and evaluation errors. Training loss not only reflects whether the model's current output is consistent with the target result, but also reflects the model's learning quality in areas such as environment modeling, ontology state modeling, intermediate representation learning, and task understanding.

[0116] Training loss can include at least two of the following: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss. This means that training loss is not limited to a fixed, single loss form, but can flexibly select two, three, or all four types of losses to participate in model optimization, depending on the task scenario, data annotation, model structure design, or training phase requirements. For example, in scenarios emphasizing environment modeling, future observation prediction loss and joint feature loss can be used together; in scenarios emphasizing ontology action understanding and task completion judgment, future motion state prediction loss and task evaluation loss can be used together; and in scenarios requiring comprehensive modeling of environment evolution, ontology evolution, and task results, all four types of losses can be used simultaneously for joint training. This embodiment helps improve adaptability to different tasks and training resource conditions, and reduces the implementation limitations caused by overly narrow training objectives.

[0117] Future observation prediction loss can be defined as a loss term characterizing the degree of difference between the future observation information output by the embodied intelligence model and the corresponding future observation information of the target. This loss can supervise the environmental evolution prediction ability of the embodied intelligence model, enabling it to learn the possible changes in the future scene under current observation and action-driven conditions. Future observation prediction loss can be determined through methods such as pixel-level difference calculation, feature-level difference calculation, structural similarity calculation, semantic consistency calculation, temporal consistency calculation, or multi-scale comparison calculation. For example, when the future observation information is a future image result, the future observation prediction loss can be calculated based on the pixel or feature differences between the predicted image and the target image; when the future observation information is a future point cloud result or a future scene layout result, it can also be calculated based on geometric or spatial relationship differences.

[0118] Target future observation information can refer to reference information corresponding to future observation information, used as a supervisory target for future observation prediction. Target future observation information can originate from subsequent moment-in-time perception results during real task execution, labeled results from offline collected data, reference results generated by the simulation environment, or future state results that have been manually verified. Target future observation information can include one or more of the following: target future image, target future video clip, target future depth result, target future point cloud result, and target future scene semantic result. By comparing future observation information with target future observation information, the prediction error of the embodied intelligent model on the environmental evolution trend can be quantified, further providing a supervisory basis for model parameter updates.

[0119] The future motion state prediction loss can be defined as a loss term characterizing the degree of difference between the future motion state information output by the embodied intelligence model and the corresponding target future motion state information. This loss can supervise the embodied intelligence model's ability to model ontological state evolution, enabling it to learn how the agent's actions continuously change under control-driven and current state conditions. The future motion state prediction loss can be determined based on differences in position, posture, joint angles, velocity, acceleration, temporal trajectory, or state consistency. For example, when the future motion state information is the future end-effector pose, the prediction loss can be determined based on the deviation between the predicted and target end-effector positions, as well as the deviation between the predicted and target postures; when the future motion state information is the future joint state, the prediction loss can be determined based on the deviations in joint angles or velocities.

[0120] The target's future motion state information can refer to reference information corresponding to the future motion state information, used as a supervisory target for predicting the future motion state. This information can originate from subsequent ontology state records during the actual execution of the embodied agent, reference state sequences in the simulated execution trajectory, state annotation results in offline data samples, or manually verified target state results. The target's future motion state information may include one or more of the following: the target's future end-effector position, the target's future end-effector posture, the target's future joint angle, the target's future joint velocity, the target's future clamping state, and the target's future base posture. By comparing the future motion state information with the target's future motion state information, the prediction error of the embodied intelligent model on the evolution of the ontology's actions can be quantified, thus providing a supervisory basis at the ontology state level for updating model parameters.

[0121] Joint feature loss refers to a loss term used to characterize the degree of difference between the joint feature information output by the embodied intelligence model and the corresponding target joint feature information. Joint feature loss focuses on whether the intermediate representation space has learned effective features relevant to task and environmental changes, thus constraining the internal representational capabilities of the embodied intelligence model. Joint feature loss can be determined through methods such as vector distance calculation, feature map difference calculation, distribution consistency calculation, contrastive learning loss calculation, semantic feature matching calculation, or cross-modal feature alignment calculation. For example, joint feature loss can be determined by comparing the distance between the predicted joint feature vector and the target joint feature vector, or by comparing the structural and semantic differences between the predicted and target feature maps. Introducing joint feature loss helps improve the stability, discriminativeness, and task relevance of the intermediate representations in the embodied intelligence model.

[0122] Target joint feature information can refer to reference information corresponding to the joint feature information, used as a supervisory target in the joint feature learning process. Target joint feature information can originate from pre-built teacher model outputs, reference features extracted through annotation constraints, intermediate representations generated by historically optimized models, target feature representations corresponding to simulation results, or manually selected feature benchmarks. Target joint feature information can include one or more of the following: semantic feature vectors, temporal feature sequences, spatial structure feature maps, and cross-modal fusion representations. By comparing the joint feature information with the target joint feature information, the embodied intelligence model can be constrained to learn intermediate representations that better conform to environmental semantics, action semantics, and task semantics, thereby improving the problem that although the model can output partially correct results, the intermediate feature representations are unstable.

[0123] Task evaluation loss can be defined as a loss term that characterizes the degree of difference between the task evaluation information output by the embodied intelligence model and the corresponding target task evaluation information. Task evaluation loss can supervise the embodied intelligence model's ability to understand tasks and judge task results, enabling the model to learn which environmental changes, ontology state changes, and interaction results correspond to task success, task failure, or different levels of task completion under different task objectives. Task evaluation loss can be determined based on classification error, regression error, ranking error, probability distribution differences, label matching differences, or stage judgment differences. For example, when the task evaluation information is a reward value, the task evaluation loss can be determined based on the difference between the predicted reward value and the target reward value; when the task evaluation information is the task success probability, it can be determined based on the difference between the predicted probability and the target probability distribution; when the task evaluation information is a task completion label, it can be determined based on the difference between the label classification result and the target label.

[0124] Target task evaluation information refers to reference information corresponding to task evaluation information, used as a supervised objective in the task evaluation learning process. Target task evaluation information may include at least one of the following: target reward value, target task success probability, target task score, target task completion degree, target task completion label, and target task progress information. Target task evaluation information can originate from manually labeled results, task execution truth values, rule calculation results, offline statistical results, or simulation reference results. For example, in a grasping task, target task evaluation information may be a task completion label or reward value generated based on whether the target object is grasped and held stably; in an assembly task, target task evaluation information may be a task completion degree result or task score result generated based on whether the component is inserted completely and whether the insertion depth meets the standard. By comparing task evaluation information with target task evaluation information, embodied intelligence models can gradually learn evaluation rules oriented towards task objectives.

[0125] As an example, the embodied intelligence model first outputs one or more results from future observation information, future motion state information, joint feature information, and task evaluation information; then, it acquires the target information corresponding to these results; next, it quantifies the difference between the predicted results and the target information to form corresponding loss terms; finally, it updates the model parameters based on one or more loss terms. Thus, this embodiment can simultaneously incorporate environmental prediction errors, ontology state prediction errors, intermediate representation errors, and task evaluation errors into the training loop, thereby improving the embodied intelligence model's ability to jointly model environmental evolution, ontology evolution, and task results.

[0126] In this embodiment, during the training of the embodied intelligence model, instead of relying solely on a single prediction target to update model parameters, at least two of the following are introduced: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss. This constructs a multi-dimensional training constraint mechanism oriented towards environmental evolution, ontology motion evolution, intermediate representation learning, and task result evaluation. Specifically, the future observation prediction loss constrains the embodied intelligence model's environmental prediction capability by comparing the differences between future observation information and the corresponding target future observation information. The future motion state prediction loss constrains the embodied intelligence model's ontology state evolution modeling capability by comparing the differences between future motion state information and the corresponding target future motion state information. The joint feature loss constrains the embodied intelligence model's representation quality in the intermediate feature space by comparing the differences between joint feature information and the corresponding target joint feature information. The task evaluation loss constrains the embodied intelligence model's ability to judge task completion by comparing the differences between task evaluation information and the corresponding target task evaluation information. Based on the joint constraints of at least two types of losses mentioned above, the embodied intelligence model can learn the laws of environmental change, the laws of action state change, the laws of feature expression, and the laws of task achievement during the training process, thereby improving the model's comprehensive performance in future observation prediction, future motion state prediction, and task evaluation.

[0127] This embodiment sets the training loss as at least two of the following: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss. This allows for joint constraints on the embodied intelligence model from multiple dimensions: environment prediction, ontology state prediction, intermediate feature learning, and task result judgment, improving the completeness of model training and its multi-objective adaptability. By determining the future observation prediction loss based on future observation information and its corresponding target future observation information, the environmental evolution results output by the embodied intelligence model can be supervised, improving the accuracy of future observation predictions and reducing the problem of the model focusing only on local visual changes while ignoring the overall task scene evolution trend. By determining the future motion state prediction loss based on future motion state information and its corresponding target future motion state information, the ontology action evolution results of the embodied intelligent agent can be supervised, improving the realism and continuity of future motion state predictions and addressing the disconnect between environment prediction and ontology state prediction. By determining the joint feature loss based on joint feature information and its corresponding target joint feature information, the intermediate representation space can be constrained, improving the stability and discriminativeness of feature representations and reducing the risk of joint features deviating from task-related semantics. By determining the task evaluation loss based on task evaluation information and its corresponding target task evaluation information, the embodied intelligence model can directly learn the judgment rules of task results during training, improving the accuracy and task adaptability of task evaluation results. By using multiple training losses to jointly determine the training objective, the embodied intelligence model can simultaneously take into account environment modeling, ontology modeling, feature modeling, and task evaluation modeling, improving the generalization ability and application effect of the trained model.

[0128] This embodiment addresses the problem in related technologies where the training process of embodied intelligent models tends to focus on a single supervisory signal and struggles to simultaneously consider environmental evolution learning, ontology motion learning, feature representation learning, and task result evaluation learning. It adopts a joint constraint concept where the training loss consists of at least two of the following: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss. This enhances the multi-dimensional supervision capability of the embodied intelligent model training process, improves the model's joint modeling effect across the entire embodied intelligent task process, and increases the prediction accuracy, evaluation accuracy, and generalization ability of the trained model.

[0129] See Figure 3 , Figure 3 This is a flowchart of an embodied intelligent data processing embodiment provided in this application.

[0130] In a specific application scenario, embodiments of this application provide an embodied intelligence data processing flow, which may include the following steps: First, the observation data corresponding to the embodied intelligent agent is encoded using an encoder to obtain an observation encoding result; the motion control input data corresponding to the embodied intelligent agent is encoded using an encoder to obtain a motion control input encoding result, wherein the motion control input data may be, for example, motion trajectory control input data. Subsequently, the observation encoding result and the motion control input encoding result are jointly modeled using a joint modeling module to obtain joint feature information; the joint feature information is decoded using a decoder to obtain future observation information.

[0131] Next, the state expert module performs state prediction based on motion state data and joint feature information to obtain future motion state information. Motion state data can include current motion state data and historical motion state data. The current motion state data can include current pose data corresponding to at least one of the distal end and joints, and the historical motion state data can include historical pose data corresponding to at least one of the distal end and joints. Future motion state information can include future pose information corresponding to at least one of the distal end and joints.

[0132] Based on this, the evaluation expert module performs task evaluation based on future observation information and embodied intelligent task information to obtain task evaluation information, such as future reward prediction information.

[0133] In some embodiments, the embodied intelligence model may include a joint modeling module, a state expert module, and an evaluation expert module. The embodied intelligence model may also include one or more encoders and one or more decoders, wherein the one or more encoders are used to encode observation data and motion control input data, and the one or more decoders are used to generate future observation information based on joint feature information.

[0134] See Figure 4 , Figure 4 This is a structural block diagram of an embodied intelligent data processing system provided in an embodiment of this application.

[0135] This application also provides an embodied intelligent data processing system, which includes a first receiving module and a first processing module.

[0136] The first receiving module is used to receive observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the target embodied intelligent agent.

[0137] The first processing module is used to input the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligent agent into the trained embodied intelligent model, so as to perform the following processing based on the embodied intelligent model: perform joint modeling based on the observation data and motion control input data corresponding to the target embodied intelligent agent to obtain joint feature information corresponding to the target embodied intelligent agent; determine the future observation information corresponding to the target embodied intelligent agent based on the joint feature information corresponding to the target embodied intelligent agent; determine the future motion state information corresponding to the target embodied intelligent agent based on the motion state data and joint feature information corresponding to the target embodied intelligent agent; and determine the task evaluation information corresponding to the target embodied intelligent agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the target embodied intelligent agent.

[0138] See Figure 5 , Figure 5 This is a structural block diagram of an embodied intelligent model training system provided in an embodiment of this application.

[0139] This application also provides an embodied intelligent model training system, which includes a second receiving module, a second processing module, a loss determination module, and a parameter update module.

[0140] The second receiving module is used to receive observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the embodied intelligent agent.

[0141] The second processing module is used to input the observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the embodied intelligent agent into the embodied intelligent model to be trained, so as to perform the following processing based on the embodied intelligent model to be trained: perform joint modeling based on the observation data and motion control input data corresponding to the embodied intelligent agent to obtain joint feature information corresponding to the embodied intelligent agent; and perform at least one of the following processing based on the embodied intelligent model to be trained: determine the future observation information corresponding to the embodied intelligent agent based on the joint feature information corresponding to the embodied intelligent agent; determine the future motion state information corresponding to the embodied intelligent agent based on the motion state data and joint feature information corresponding to the embodied intelligent agent; and determine the task evaluation information corresponding to the embodied intelligent agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligent task information corresponding to the embodied intelligent agent.

[0142] The loss determination module is used to determine the training loss based on at least two of the following: future observation information, future motion state information, joint feature information, and task evaluation information corresponding to the embodied agent, and the corresponding target information.

[0143] The parameter update module is used to update at least one model parameter of the embodied intelligence model to be trained based on the training loss.

[0144] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0145] This application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0146] The computer program product may be a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the computer program product of the embodiments of this application is not limited thereto, and the computer program product may be any combination of one or more computer-readable media.

[0147] See Figure 6 , Figure 6 This is a structural block diagram of a computer device provided in an embodiment of this application.

[0148] This application also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0149] The embodiments of this application do not limit the computer device, which may be, for example, a local computer device, a cloud computer device, a distributed computer device, etc.

[0150] The computer device may include: a memory 110, a processor 120, and a communication interface 130. The memory 110, the processor 120, and the communication interface 130 are connected through internal connection paths.

[0151] The memory 110 is used to store computer programs, which in some implementations may include code for implementing the methods of the embodiments of this application.

[0152] The processor 120 executes the computer program stored in the memory 110 to control the communication interface 130 to receive input data and information, and output operation results and other data. In some implementations, when the solutions of the embodiments of this application are implemented by software or firmware, the computer program used to implement the solutions of the embodiments of this application can be stored in the processor 120 and executed by the processor 120.

[0153] The memory 110 may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM). It should be noted that the memory 110 described herein is intended to include, but is not limited to, any memory of these and other suitable types. As an example, the memory 110 includes random access memory (RAM), cache memory, and read-only memory (ROM). The memory 110 stores a computer program that can be executed by processor 120, causing processor 120 to implement the steps of any of the methods described above.

[0154] The processor 120 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 120 can be any conventional processor.

[0155] In implementation, each step of the above method can be completed by the integrated logic circuitry of the hardware in the processor 120 or by instructions in software form. The method disclosed in the embodiments of this application can be directly implemented by the hardware processor, or by a combination of hardware and software modules in the processor 120. The software modules can be located in mature storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in the memory 110, and the processor 120 reads the information in the memory 110 and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are not provided here.

[0156] In some implementations, in addition to the hardware units described above, computer devices may also include software modules, such as operating systems, basic input / output systems (BIOS), and application software.

[0157] An operating system is used to manage the hardware and / or software resources of a computer device; it is the kernel and foundation of the computer. The operating system handles fundamental tasks such as managing and configuring memory, determining the priority of system resource allocation, controlling input and output devices, operating the network, and managing the file system. To facilitate user operation, most operating systems provide a user interface for interaction with the system.

[0158] The BIOS is used to perform hardware initialization during the power-on boot phase and to provide runtime services for the operating system and applications. In some implementations, the BIOS can also monitor and display processor temperature and execute temperature protection strategies.

[0159] Application software, also known as an application program, can be understood as software written for a specific user application purpose, and is one of the main categories of computer software. For example, application software can be a program used to achieve purposes such as power control and temperature management.

[0160] It is understood that the specific examples in this application are only intended to help those skilled in the art better understand the implementation of this application, and are not intended to limit the scope of protection of this application.

[0161] It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application.

[0162] It is understood that the various implementation methods described in this application can be implemented individually or in combination, and this application does not limit them.

[0163] Unless otherwise stated, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0164] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0165] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the embodiments described above can be referred to the corresponding processes in other embodiments, and will not be repeated here.

[0166] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0167] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the technical solution in this application, depending on actual needs.

[0168] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0169] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, essentially, or the part that contributes to related technologies, or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0170] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for embodied intelligence data processing, characterized in that, The embodied intelligence data processing method includes: Receive observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the target embodied intelligent agent; The observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligent agent are input into the trained embodied intelligent model, and the following processing is performed based on the embodied intelligent model: Joint modeling is performed based on the observation data and motion control input data corresponding to the target embodied intelligent agent to obtain the joint feature information corresponding to the target embodied intelligent agent; Based on the joint feature information corresponding to the target embodied intelligent agent, the future observation information corresponding to the target embodied intelligent agent is determined; Based on the motion state data and joint feature information corresponding to the target embodied intelligent agent, the future motion state information corresponding to the target embodied intelligent agent is determined. Based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information corresponding to the target embodied intelligent agent, the task evaluation information corresponding to the target embodied intelligent agent is determined.

2. The embodied intelligence data processing method according to claim 1, characterized in that, The observation data includes multi-view observation data; and / or, the observation data includes current observation data.

3. The embodied intelligence data processing method according to claim 1, characterized in that, The motion control input data includes at least one of end-effector control input information and joint control input information; and / or, The motion state data includes at least one of end-effector state information and joint state information; and / or, The motion status data includes current motion status data and historical motion status data.

4. The embodied intelligence data processing method according to claim 1, characterized in that, The task evaluation information is determined based on the future observation information and the embodied intelligence task information; and / or, The task evaluation information includes at least one of the following: reward value, task success probability, task score, task completion rate, task completion tag, and task progress information.

5. A method for training an embodied intelligence model, characterized in that, The embodied intelligence model training method includes: Receive observation data, motion control input data, motion state data, and embodied intelligent task information corresponding to the embodied intelligent agent; The observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the embodied intelligent agent are input into the embodied intelligent model to be trained, so as to perform the following processing based on the embodied intelligent model to be trained: joint modeling is performed based on the observation data and motion control input data corresponding to the embodied intelligent agent to obtain the joint feature information corresponding to the embodied intelligent agent; Furthermore, based on the embodied intelligence model to be trained, perform at least one of the following processes: determine future observation information corresponding to the embodied intelligence agent based on the joint feature information corresponding to the embodied intelligence agent; determine future motion state information corresponding to the embodied intelligence agent based on the motion state data and joint feature information corresponding to the embodied intelligence agent; determine task evaluation information corresponding to the embodied intelligence agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the embodied intelligence agent. The training loss is determined based on at least two of the future observation information, future motion state information, joint feature information and task evaluation information corresponding to the embodied intelligent agent and the corresponding target information. Based on the training loss, update at least one model parameter of the embodied intelligence model to be trained.

6. The embodied intelligent model training method according to claim 5, characterized in that, The embodied intelligent model to be trained includes a joint modeling module, a state expert module, and an evaluation expert module. The joint modeling module is used to perform joint modeling based on the observation data and motion control input data corresponding to the embodied intelligent agent to obtain joint feature information corresponding to the embodied intelligent agent, wherein the future observation information corresponding to the embodied intelligent agent is determined based on the joint feature information corresponding to the embodied intelligent agent. The state expert module is used to determine the future motion state information of the embodied intelligent agent based on the motion state data and joint feature information corresponding to the embodied intelligent agent; The evaluation expert module is used to determine the task evaluation information corresponding to the embodied intelligent agent based on at least two of the following: future observation information, future motion state information, joint feature information, and embodied intelligent task information.

7. The embodied intelligent model training method according to claim 5, characterized in that, The training loss includes at least one of the following: future observation prediction loss, future motion state prediction loss, joint feature loss, and task evaluation loss. The future observation prediction loss is determined based on the future observation information corresponding to the embodied agent and its corresponding target future observation information; and / or, The future motion state prediction loss is determined based on the future motion state information corresponding to the embodied agent and the corresponding target future motion state information; and / or, The joint feature loss is determined based on the joint feature information corresponding to the embodied agent and its corresponding target joint feature information; and / or, The task evaluation loss is determined based on the task evaluation information corresponding to the embodied intelligent agent and its corresponding target task evaluation information.

8. An embodied intelligent data processing system, characterized in that, The embodied intelligent data processing system includes: The first receiving module is used to receive observation data, motion control input data, motion state data and embodied intelligent task information corresponding to the target embodied intelligent agent; The first processing module is used to input the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the target embodied intelligent agent into the trained embodied intelligent model, and to perform the following processing based on the embodied intelligent model: performing joint modeling based on the observation data and motion control input data corresponding to the target embodied intelligent agent to obtain joint feature information corresponding to the target embodied intelligent agent; determining future observation information corresponding to the target embodied intelligent agent based on the joint feature information corresponding to the target embodied intelligent agent; determining future motion state information corresponding to the target embodied intelligent agent based on the motion state data and joint feature information corresponding to the target embodied intelligent agent; and determining task evaluation information corresponding to the target embodied intelligent agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information.

9. An embodied intelligent model training system, characterized in that, The embodied intelligence model training system includes: The second receiving module is used to receive observation data, motion control input data, motion state data and embodied intelligent task information corresponding to the embodied intelligent agent; The second processing module is used to input the observation data, motion control input data, motion state data, and embodied intelligence task information corresponding to the embodied intelligent agent into the embodied intelligent model to be trained, so as to perform the following processing based on the embodied intelligent model to be trained: performing joint modeling based on the observation data and motion control input data corresponding to the embodied intelligent agent to obtain joint feature information corresponding to the embodied intelligent agent; and performing at least one of the following processing based on the embodied intelligent model to be trained: determining future observation information corresponding to the embodied intelligent agent based on the joint feature information corresponding to the embodied intelligent agent; determining future motion state information corresponding to the embodied intelligent agent based on the motion state data and joint feature information corresponding to the embodied intelligent agent; and determining task evaluation information corresponding to the embodied intelligent agent based on at least two of the future observation information, future motion state information, joint feature information, and embodied intelligence task information corresponding to the embodied intelligent agent. The loss determination module is used to determine the training loss based on at least two of the following: future observation information, future motion state information, joint feature information, and task evaluation information corresponding to the embodied intelligent agent, and the corresponding target information; The parameter update module is used to update at least one model parameter of the embodied intelligence model to be trained based on the training loss.

10. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method according to any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

12. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 7.

Images