A robot control method and system based on a streaming flow matching world model

Abstract

This invention discloses a robot control method and system based on a streaming flow matching world model. The method maps the robot's current state observation information and target state observation information to a latent space using a characterization encoder. Based on the streaming flow matching world model, the latent state is continuously advanced under the motion condition velocity field, avoiding direct modeling in the high-dimensional observation space and reducing the computational complexity of prediction. Simultaneously, each intermediate latent state obtained from integration serves as a continuously inheritable latent world state in subsequent predictions, reducing error accumulation in long-term autoregressive predictions. Furthermore, by combining the evaluation of multiple candidate motion flow sequences with the current control segmented rolling execution, control commands can be corrected in real time based on the latest observations, improving the robot's real-time control performance, trajectory smoothness, and closed-loop robustness in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of robot intelligent control, motion planning and world model learning technology, specifically involving a robot control method and system based on a streaming flow matching world model. Background Technology

[0002] With the widespread application of robotics in industrial operations, home services, warehousing and logistics, medical assistance, and embodied intelligence, robot control systems need to achieve stable, accurate, and efficient motion decisions under high-dimensional perceptual inputs, complex dynamic environments, and long-term time-domain task objective constraints. Traditional robot control methods typically rely on precisely modeled dynamic equations, explicit state estimation, or manually designed task cost functions. In complex environments, contact tasks, and vision-driven tasks, these methods are prone to problems such as high modeling costs, weak generalization ability, and insufficient online planning efficiency.

[0003] In recent years, learning-based world model methods have provided a new technical approach for robot planning and control by learning the state evolution patterns of the environment and the robot itself from historical interaction data. However, existing world model methods generally have the following shortcomings: First, directly modeling in pixel space or high-dimensional observation space results in high computational costs and large prediction errors; second, when using autoregressive prediction methods, long-term rolling prediction can easily lead to error accumulation, affecting control stability; third, although some methods can be used for offline trajectory generation, they lack a streaming decision-making mechanism suitable for online closed-loop control, making it difficult to balance real-time performance and robustness.

[0004] Furthermore, traditional generative flow matching or diffusion-based generation paradigms typically transform noise or intermediate distributions stepwise to the target sample. The intermediate denoised states or intermediate flow states generated during the control process are often not directly used as the robot's world state. In online robot control scenarios, a world model that starts from current real-world observations and continuously advances state evolution under action condition constraints is needed, so that intermediate potential states can be continuously inherited and used for closed-loop decision-making.

[0005] Therefore, how to efficiently model the robot's state evolution in the potential space, predict the future state through conditional velocity field integration, and complete real-time control by combining candidate action flow evaluation and closed-loop rolling execution mechanism has become a technical problem that urgently needs to be solved in the field of robot intelligent control. Summary of the Invention

[0006] This invention provides a robot control method and system based on a streaming flow matching world model, aiming to solve the technical problems of low efficiency in high-dimensional observation modeling, significant long-term error accumulation in autoregressive world models, and insufficient robustness of online rolling control in the prior art, while taking into account real-time performance, stability, and generalization ability under complex task conditions.

[0007] This invention applies a flow matching mechanism to the continuous dynamic prediction of a latent world model. It uses the latent state encoded from current real observations as the initial value for integration, and progressively advances the latent state through a conditional velocity field under action flow conditions, rather than generating the final sample solely from random noise. Therefore, intermediate latent states during the integration process can be continuously inherited as effective world states and used for candidate action evaluation and piecewise closed-loop control.

[0008] In a first aspect, the present invention provides a robot control method based on a streaming flow matching world model, the method comprising:

[0009] The robot's current state observation information and target state observation information are acquired, and the current state observation information and target state observation information are encoded by a representation encoder to obtain the current potential state and the target potential state.

[0010] Multiple sets of candidate action flow sequences are generated within the execution segment, and each set of candidate action flow sequences includes multiple control actions arranged in chronological order;

[0011] The current potential state is used as the initial value for integration, and each group of candidate action flow sequences is used as a condition. The input is given to the streaming flow matching world model, and the potential state is continuously integrated based on the action condition velocity field output by the streaming flow matching world model to obtain the future potential state corresponding to each group of candidate action flow sequences. In the process of continuous integration, the intermediate potential state obtained in each integration step is used as the current potential state for the next integration step and is continued to be input into the streaming flow matching world model.

[0012] Based on the difference between the future potential state and the target potential state corresponding to each group of candidate action flow sequences, the candidate action flow sequences are evaluated and the optimal candidate action flow sequence is selected.

[0013] Execute the current control segment in the optimal candidate action flow sequence and repeat the above process until the preset task completion conditions are met.

[0014] Preferably, the training process of the streaming world matching model is as follows:

[0015] Historical interaction data of the robot is acquired, and a training dataset is constructed from the historical interaction data. The training dataset includes observed trajectory segments and corresponding motion flow sequences. The observed trajectory segments are input into a representation encoder to obtain the latent state. Time variables are sampled within the normalized trajectory segments, and a training flow state and a reference velocity field corresponding to the time variables are constructed based on the latent state.

[0016] The parameterized conditional velocity field is trained by using the training flow state, time variables, and corresponding action flow as inputs, and the reference velocity field as the supervision target, thus completing the training of the flow-matching world model.

[0017] Preferably, the training stream state is formed by superimposing a time-decaying random perturbation onto the latent state.

[0018] Preferably, the continuous integral propagation process is as follows:

[0019] Within the normalized trajectory segment, the time variable is updated according to a preset integration step size. In each integration step, the potential state, time variable, and conditional action input stream corresponding to the time variable in the candidate action stream sequence of the current integration step are matched with the world model to obtain the action conditional velocity. The potential state is updated according to the action conditional velocity and the preset integration step size until the ratio of the execution segment time length to the trajectory segment time length is reached.

[0020] Preferably, a comprehensive cost function is constructed to evaluate each group of candidate action flow sequences; the comprehensive cost function includes a target proximity term between the future potential state and the target potential state, and is superimposed with one or more of the following: obstacle distance penalty term, altitude range penalty term, action smoothing term, yaw change penalty term, energy consumption term, joint limitation term, or task semantic constraint term.

[0021] Preferably, the candidate action stream sequence is obtained through one or more of the following methods: random sampling, generation based on action prior policy, sampling from a historical action library, generation based on a reinforcement learning policy network, generation based on a diffusion policy, or generation based on a sampling optimizer.

[0022] Preferably, the current state observation information includes one or more of the following: robot visual image, depth information, force information, joint state, body motion state, and pose state; the target state observation information includes one or more of the following: target pose, target image, target region information, target depth information, target scene configuration, and task semantic information.

[0023] Preferably, the target state observation information is further integrated with text task instructions, language embeddings, or skill tags to construct a streaming matching world model under the joint conditions of vision-action-semantics.

[0024] Preferably, the execution segment time length is shorter than the trajectory segment time length.

[0025] Secondly, the present invention provides a robot control system based on a streaming flow matching world model, which is used to execute the above-mentioned robot control method; the robot control system includes an observation acquisition module, a latent representation module, a candidate action generation module, a streaming flow matching prediction module, a candidate action evaluation module, and a closed-loop execution module;

[0026] The observation acquisition module is used to acquire the robot's current state observation information and target state observation information; the latent representation module is used to encode the current state observation information and target state observation information respectively through a representation encoder to obtain the current latent state and the target latent state; the candidate action generation module is used to generate multiple sets of candidate action flow sequences within the execution segment, each set of candidate action flow sequences including multiple control actions arranged in chronological order; the streaming flow matching prediction module is used to use the current latent state as the initial value of the integral, and each set of candidate action flow sequences as the condition input to the streaming flow matching world model, and outputs based on the streaming flow matching world model. The velocity field of the action condition is continuously integrated to advance the potential state, resulting in the future potential state corresponding to each group of candidate action flow sequences. The candidate action evaluation module is used to evaluate each group of candidate action flow sequences based on the difference between the future potential state corresponding to each group of candidate action flow sequences and the target potential state, and select the optimal candidate action flow sequence. The closed-loop execution module is used to execute the current control segment in the optimal candidate action flow sequence, and after execution, it triggers the observation acquisition module, potential characterization module, candidate action generation module, streaming flow matching prediction module and candidate action evaluation module to repeatedly execute the corresponding processing until the preset task completion conditions are met.

[0027] The beneficial effects of this invention are:

[0028] 1. This invention significantly reduces modeling complexity and improves the efficiency of online prediction and control decisions by learning the conditional velocity field in the latent space, rather than directly modeling in the high-dimensional original observation space. Simultaneously, this invention models the continuous evolution of the latent state over time using a streaming flow matching method, which, compared to traditional autoregressive prediction methods, reduces the accumulation of long-term temporal errors caused by discrete frame-by-frame prediction and improves trajectory smoothness. Furthermore, compared to traditional generative flow matching or diffusion-based generative paradigms, intermediate latent states during the integration process do not need to be discarded but are continuously adopted as usable latent world states, thereby improving online prediction efficiency and closed-loop control adaptability.

[0029] 2. This invention combines candidate action flow evaluation with segmented closed-loop execution, possessing both the online optimization capability of model predictive control and the generalization capability of learning world models. This can improve the robustness of robots in complex environments. Furthermore, this invention is applicable to robotic arm grasping and assembly, mobile robot navigation, dual-arm collaborative operation, and multimodal embodied intelligence tasks, and has strong versatility and engineering application value. Attached Figure Description

[0030] Figure 1 This is the overall flowchart of Embodiment 2 of the present invention.

[0031] Figure 2 This is a schematic diagram of the training process of the streaming flow matching world model in Embodiment 2 of the present invention.

[0032] Figure 3 This is a schematic diagram of the closed-loop control process in Embodiment 2 of the present invention. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings.

[0034] Example 1

[0035] A robot control system based on a streaming flow matching world model can be deployed on standalone controllers, edge computing platforms, cloud-edge collaborative control platforms, robotic arm platforms, mobile robot platforms, dual-arm collaborative platforms, or other embodied intelligent systems. For different robot platforms, only the observation interface, motion interface, and representation encoder input format need to be replaced to adapt it to various robot task scenarios. The robot control system includes an observation acquisition module, a latent representation module, a candidate motion generation module, a streaming flow matching prediction module, a candidate motion evaluation module, and a closed-loop execution module. The observation acquisition module is used to collect the robot's current observation information and target observation information. The latent representation module is used to encode the current observation information (current image, depth information, or flight state vector) into a current latent state and to encode the target observation information (target waypoint, target image, or target region information) into a target latent state.

[0036] The candidate action generation module is used to generate multiple sets of candidate action flow sequences within the execution segment. Each set of candidate action flow sequences consists of multiple control actions arranged in chronological order.

[0037] The streaming flow matching prediction module is used to predict the conditional velocity field integral in the potential space based on the candidate action flow sequence; the candidate action evaluation module is used to evaluate the merits of the candidate actions based on the difference between the future potential state and the target potential state; the closed-loop execution module is used to execute the current optimal control segment and re-trigger observation, prediction and planning after execution.

[0038] In some embodiments, the observation acquisition module includes one or more of a forward-looking camera, a depth camera, an inertial measurement unit, an altimeter, and a flight control status interface.

[0039] Example 2

[0040] like Figure 1 As shown, a robot control method based on a flow-matching world model is used to drive a robot to perform visual navigation, target approach, or fixed-point landing tasks in indoor or low-altitude environments. The robot control method includes the following steps:

[0041] Step 1: Construct the training dataset

[0042] like Figure 2 As shown, historical interaction data of the robot in the environment is obtained to construct a training dataset. ;in, Indicates a segment of the observed trajectory; Representation of observation trajectory segments The corresponding action flow sequence.

[0043] In some embodiments, the robot may be a drone or other flying robot.

[0044] Step 2: Data Preprocessing

[0045] A set of historical trajectory segments (h, ξ) are sampled from the training dataset, and the time length of each trajectory segment is... Normalize the time length to Then, randomly sample the time variable t from the interval [0,1]. Analyze the observed trajectory segments from the historical trajectory. Input the characterization encoder to obtain the latent state. To enhance the model's robustness to noise and perturbations, in the latent state... Based on this, a random perturbation that decays over time is superimposed to construct the training stream state. It is represented as:

[0046]

[0047] in, This represents the encoding mapping that characterizes the encoder; Indicates gradient freezing; The initial disturbance coefficient; The attenuation coefficient; It is random noise that follows a standard distribution.

[0048] In this embodiment, the characterization encoder is a variational autoencoder, a contrastive learning encoder, or other neural network encoder suitable for high-dimensional observation compression, thereby mapping high-dimensional observations to a unified latent space for subsequent state evolution modeling and target evaluation.

[0049] Step 3: Train the streaming world model

[0050] Streaming-matched world models are used to learn the conditional velocity field of latent states in the time dimension. According to the potential state Construct a reference velocity field It is represented as:

[0051]

[0052] Constructing a parameterized conditional velocity field The parameterized velocity field is trained using a conditional flow matching loss function to approximate the reference velocity field; the conditional flow matching loss function It is expressed as follows:

[0053]

[0054] in, This represents the conditional action flow corresponding to time.

[0055] The model parameters θ of the streaming flow matching world model are minimized by the conditional flow matching loss function. Training is performed, and after training is complete, a streaming flow-matched world model that can be used to predict the potential state evolution in the control phase is obtained. Conditional flow-matching loss function. It is expressed as follows:

[0056]

[0057] Where λ is the learning rate.

[0058] Step 4: Online Robot Control

[0059] Step 4-1. Latent State Encoding

[0060] like Figure 3 As shown, current state observation information is collected through cameras, depth sensors, force sensors, encoders, or the robot's body state interface. And determine the target state observation information according to the mission requirements. The current state observation information includes airborne camera images, depth maps, aircraft attitude angles, angular velocities, linear velocities, altitude information, and optional position estimation information. The target state observation information includes the target's spatial location, target pose, target region image, target depth map, target scene configuration, or target region semantic information provided by the upper-level task system. The current state observation information and the target state observation information are encoded using a pre-built representation encoder to obtain the current potential state. and target potential state .

[0061] In some embodiments, to enhance the model's ability to express the semantics of complex tasks, target state observation information is further integrated with textual task instructions, language embeddings, or skill tags to construct a streaming matching world model under the joint conditions of vision, action, and semantics.

[0062] Step 4-2. Generate candidate action stream sequence

[0063] In time length Generate N candidate action flow sequences within the execution segment. , Must meet Each group of candidate action stream sequences It consists of multiple consecutive control actions arranged in chronological order.

[0064] In some embodiments, candidate action flow sequences are generated through random initialization, generation based on prior flight action strategies, sampling from a historical action library, or a combination of these methods. For example, random initialization can sample multiple sets of control command sequences within safety constraints of speed, acceleration, yaw rate, and altitude changes; prior strategy-based generation can construct candidate sequences based on flight action templates such as hovering, forward movement, turning, ascent, descent, deceleration, and obstacle avoidance; and historical action library sampling can extract action segments from successful navigation, target approach, obstacle avoidance, or landing trajectories already executed by the UAV, and match, extract, or perturb and expand them based on current observations and target states to form multiple sets of candidate control schemes.

[0065] In some embodiments, the generation process of the candidate action stream sequence may combine a diffusion strategy, a reinforcement learning policy network, or a sampling optimizer to improve the quality of candidate actions and reduce search overhead.

[0066] In some embodiments, the control action includes one or more of the following: system velocity command, vertical velocity command, yaw rate command, attitude rate command, and acceleration command. Taking system velocity and yaw rate control as an example, the control action at each moment can be expressed as:

[0067]

[0068] in, These respectively represent the forward / backward, left / right, and vertical speed commands of the drone within the system. This indicates the yaw rate command.

[0069] Step 4-3. Streaming flow matching world model prediction

[0070] The current latent state is used as the initial state, and the candidate action flow sequence is used as the conditional input to the trained streaming matching world model. In the streaming matching world model, for each set of candidate action flow sequences, the current latent state is used as the initial state. As the initial state, streaming integral prediction is performed starting from normalized time t=0. Within one execution segment, iterative calculations are performed according to the discrete integral step size Δt:

[0071]

[0072] The above update expression represents closed-loop flow state progression: current potential state First, the initial value for integration is the state obtained after one integration. Immediately use the current latent state as the input to the streaming flow matching world model for the next integration step. Therefore, each intermediate potential state is an effective world state evolution result driven by the velocity field under action conditions, that is, the future flight state or visual state that the robot may reach under the action of corresponding control commands, and can continue to participate in subsequent integration and cost evaluation. Thus, it is possible to learn the evolution law of the UAV's observed state with the change of control actions based on historical flight data without explicitly building a complex aerodynamic model; at the same time, unlike traditional generative flow matching algorithms or diffusion models, it is not necessary to treat intermediate flow states or intermediate denoised states as merely generation transition quantities and discard them after obtaining the final sample. Instead, each intermediate potential state can be used as an effective world state that can be continuously inherited and reused for subsequent prediction and control evaluation.

[0073] Then update the time variable t←t+Δt until t reaches its maximum value. ;in, The duration of the segmentation; denoted as the time length of the trajectory segment. After integration, the future potential state corresponding to the candidate action flow sequence at the end of the current execution segment is obtained.

[0074] Step 4-4. Select the optimal candidate action flow sequence

[0075] A comprehensive cost function is constructed based on each future potential state and the target potential state. The comprehensive cost function includes a target approach term, an obstacle distance penalty term, an altitude range penalty term, a velocity smoothing term, a yaw change penalty term, and an energy consumption term: Comprehensive Cost Function The expression is:

[0076]

[0077] in, Candidate action stream sequence The corresponding potential future state; Indicates obstacle distance or collision risk penalty; This indicates a high degree of constraint and punishment; Indicates a penalty for smoothing velocity and attitude changes; Indicates energy consumption penalty; These are the corresponding weighting coefficients.

[0078] In some embodiments, the integrated cost function is composed not only of the potential distance to the endpoint, but also introduces intermediate critical state alignment constraints to adapt to multi-stage assembly, obstacle avoidance, or long sequence operation tasks.

[0079] The candidate action flow sequence with the minimum objective cost is selected as the optimal candidate action flow sequence. The objective cost includes the distance term between the future potential state and the target potential state, and may be supplemented with action smoothing terms, energy consumption terms, collision constraint terms, joint constraint terms, or task semantic constraint terms depending on the actual application.

[0080] Steps 4-5. Closed-loop execution and rolling update control

[0081] The current control segment in the optimal candidate action flow sequence is executed, and steps 4-1 to 4-4 are repeated after execution to reacquire the latest state observation information of the robot. The potential encoding, prediction, evaluation, and execution processes are repeated until the preset task completion conditions are met. The task completion conditions include any one of the following: the robot's end effector reaches the target pose, the manipulated object reaches the target state, the mobile robot reaches the target area, or the preset task success criteria are met. Compared with the method of generating a complete long-term control sequence at once, this implementation adopts a streaming closed-loop control strategy of "long prediction, short execution, and real-time replanning". Through this closed-loop rolling mechanism, when obstacle position changes, visual observation disturbances, airflow disturbances, or positioning errors occur during the control process, subsequent control commands can be corrected in a timely manner based on the latest observations, improving the robustness of navigation, obstacle avoidance, target approach, and fixed-point landing processes, thereby reducing the adverse effects of long-term prediction errors and environmental changes on the control effect.

[0082] The above-described UAV visual navigation application is merely an exemplary embodiment of the present invention and does not constitute a limitation on robot type, sensor type, task form, or control interface. The present invention is also applicable to robotic arm grasping, mobile robot navigation, dual-arm collaborative operation, wheeled robot inspection, and other embodied intelligent control tasks.

[0083] In summary, this invention proposes a robot control method and system based on a streaming flow matching world model. By learning the action condition velocity field in the potential space, the world state evolution is continuously advanced using the potential state encoded by the current real observation as the initial value of the integral. Furthermore, by employing a streaming candidate action evaluation and closed-loop segmented execution mechanism, intelligent robot control with real-time performance, stability, and generalization ability is achieved.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above-described embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A robot control method based on a flow-matching world model, characterized in that: The method includes: The robot's current state observation information and target state observation information are acquired, and the current state observation information and target state observation information are encoded by a representation encoder to obtain the current potential state and the target potential state. Multiple sets of candidate action flow sequences are generated within the execution segment, and each set of candidate action flow sequences includes multiple control actions arranged in chronological order; The current potential state is used as the initial value for integration, and each group of candidate action flow sequences is used as a condition. The input is given to the streaming flow matching world model, and the potential state is continuously integrated based on the action condition velocity field output by the streaming flow matching world model to obtain the future potential state corresponding to each group of candidate action flow sequences. In the process of continuous integration, the intermediate potential state obtained in each integration step is used as the current potential state for the next integration step and is continued to be input into the streaming flow matching world model. Based on the difference between the future potential state and the target potential state corresponding to each group of candidate action flow sequences, the candidate action flow sequences are evaluated and the optimal candidate action flow sequence is selected. Execute the current control segment in the optimal candidate action flow sequence and repeat the above process until the preset task completion conditions are met.

2. The robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The training process of the streaming world matching model is as follows: Historical interaction data of the robot is acquired, and a training dataset is constructed from the historical interaction data; the training dataset includes observation trajectory segments and motion flow sequences corresponding to the observation trajectory segments; The observed trajectory segments are input into the characterization encoder to obtain the latent state; Time variables are sampled within the normalized trajectory segment, and training flow states and reference velocity fields corresponding to the time variables are constructed based on the latent states. The parameterized conditional velocity field is trained by using the training flow state, time variables, and corresponding action flow as inputs, and the reference velocity field as the supervision target, thus completing the training of the flow-matching world model.

3. The robot control method based on a streaming flow matching world model according to claim 2, characterized in that: The training stream state is formed by superimposing time-decaying random perturbations onto the latent state.

4. The robot control method based on a streaming flow matching world model according to claim 2, characterized in that: The execution segment's time length is less than the trajectory segment's time length.

5. The robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The continuous integral propagation process is as follows: In each integration step, the potential state, time variable, and conditional action input stream corresponding to the time variable in the candidate action stream sequence of the current integration step are matched with the world model to obtain the action conditional velocity; the potential state and time variable are updated according to the action conditional velocity and the preset integration step size until the time variable reaches the preset time length threshold.

6. The robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The candidate action flow sequences are evaluated by constructing a comprehensive cost function. The comprehensive cost function includes a target proximity term between the future potential state and the target potential state, and is superimposed with one or more of the following: obstacle distance penalty term, altitude range penalty term, action smoothing term, yaw change penalty term, energy consumption term, joint limitation term, or task semantic constraint term.

7. The robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The candidate action stream sequence is obtained through one or more of the following methods: random sampling, generation based on action prior policy, sampling from historical action library, generation based on reinforcement learning policy network, generation based on diffusion policy, or generation based on sampling optimizer.

8. The robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The current state observation information includes one or more of the following: robot visual image, depth information, force information, joint state, body motion state, and pose state. The target state observation information includes one or more of the following: target pose, target image, target region information, target depth information, target scene configuration, and task semantic information.

9. A robot control method based on a streaming flow matching world model according to claim 1, characterized in that: The target state observation information is also integrated with text task instructions, language embeddings, or skill tags to construct a streaming matching world model under the joint conditions of vision, action, and semantics.

10. A robot control system based on a flow-matching world model, characterized in that: This is used to execute a robot control method based on a streaming flow matching world model as described in claim 1; the robot control system includes an observation acquisition module, a latent representation module, a candidate action generation module, a streaming flow matching prediction module, a candidate action evaluation module, and a closed-loop execution module; The observation acquisition module is used to acquire the robot's current state observation information and the target state observation information; The latent representation module encodes the current state observation information and the target state observation information respectively through the representation encoder to obtain the current latent state and the target latent state. The candidate action generation module generates multiple sets of candidate action flow sequences within the execution segment. Each set of candidate action flow sequences includes multiple control actions arranged in chronological order. The streaming flow matching prediction module uses the current latent state as the initial value of the integral, inputs each set of candidate action flow sequences as conditions into the streaming flow matching world model, and continuously integrates and advances the latent state based on the action condition velocity field output by the streaming flow matching world model to obtain the future latent state corresponding to each set of candidate action flow sequences. The candidate action evaluation module evaluates each set of candidate action flow sequences based on the difference between the future latent state corresponding to each set of candidate action flow sequences and the target latent state, and selects the optimal candidate action flow sequence. The closed-loop execution module executes the current control segment in the optimal candidate action flow sequence, and after execution, triggers the observation acquisition module, latent representation module, candidate action generation module, streaming flow matching prediction module, and candidate action evaluation module to repeatedly execute the corresponding processing until the preset task completion conditions are met.

Images