Robot action reasoning method and system based on Gaussian action field

Abstract

The invention relates to a robot motion reasoning method and system based on a Gaussian motion field, and the method comprises the steps: inputting a sparse and uncalibrated multi-view RGB image, extracting mixed scene features through a visual Transform backbone network, constructing a dynamic Gaussian motion field, endowing each Gaussian unit with a learnable motion attribute, and carrying out the reasoning of the motion of a robot. And realizing synchronous modeling of scene geometry and motion evolution. The system further comprises a multi-modal query module, a point cloud registration module, a diffusion model optimization module and a closed-loop control module which are respectively used for current and future scene reconstruction, mechanical arm tail end clamping jaw action estimation, action sequence optimization and real-time feedback and model updating in the execution process. According to the method, through a unified space-time modeling and closed-loop control mechanism, the robot operation problem in dynamic, shielding and uncalibrated environments is effectively solved, and the method is suitable for various application scenes such as industrial automation and service robots.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of three-dimensional vision and embodied intelligence in computer vision, and particularly relates to a robot action reasoning method and system based on Gaussian action field GAF. BACKGROUND

[0002] Multi-view generation of three-dimensional vision is a revolutionary technology for the future, and its core is to reconstruct a complete three-dimensional scene through multiple view image data. This technology usually relies on the combination of deep learning and computer vision, and can extract rich spatial information from images from different angles, thereby achieving high-precision three-dimensional reconstruction. With the emergence of emerging deep neural network models such as NeRF (Neural Radiance Field), researchers can now generate high-quality three-dimensional visual effects from a small number of views. NeRF uses deep learning algorithms to analyze and synthesize images, and can infer complete three-dimensional structures from sparse image data, significantly reducing the complexity of acquisition and calculation. Compared with traditional methods, these new technologies not only improve the efficiency of reconstruction, but also enhance the details and realism of the generated model.

[0003] Embodied intelligence refers to the ability to acquire knowledge and skills through interaction with the environment, and has received widespread attention in the field of robotics and artificial intelligence in recent years. This form of intelligence emphasizes the importance of the body in perception and cognition, especially in complex tasks and dynamic environments. With the help of advanced sensors and deep learning technology, embodied intelligent systems can analyze and understand their surroundings in real time, and thus make flexible responses. This ability not only improves the adaptability of robots, but also opens up new possibilities for human-robot collaboration, enabling robots to interact more naturally with humans and the environment. SUMMARY

[0004] The present application aims to at least partially solve one of the technical problems in the related art.

[0005] The present application proposes a robot action reasoning method based on Gaussian action field GAF, aiming to improve the understanding of dynamic scenes and the accuracy of action generation through four-dimensional spatiotemporal modeling. On the one hand, the present application uses dynamic Gaussian point cloud to perform high-fidelity reconstruction of the scene, and on the other hand, combines motion attribute prediction and diffusion optimization to significantly improve the adaptability and success rate of robot operation.

[0006] Another object of the present application is to propose a robot action reasoning system based on Gaussian action field GAF.

[0007] To achieve the above-mentioned purpose, the present application proposes a robot action reasoning method based on Gaussian action field GAF, which comprises: inputting sparse, uncalibrated multi-view RGB images; extracting a mixed scene feature of the multi-view RGB image using a visual Transformer backbone network; A dynamic Gaussian action field GAF is constructed, and a learnable motion attribute is given to each Gaussian unit by extending a three-dimensional Gaussian point cloud to model the evolution of scene geometry and motion synchronously. In response to a current scene reconstruction request, a Gaussian point cloud representation at a current time step is generated based on the GAF, and multi-view scene images are rendered; in response to a future state prediction request, a Gaussian point cloud representation at a future time step is predicted based on the GAF, and future view images are generated; and in response to a motion estimation request, Gaussian point clouds related to the gripper at the end of the robot arm in the current frame and the future frame are extracted, and an initial motion sequence is calculated by a point cloud registration module. The initial motion sequence, the current multi-view scene image, and the robot state information are input into a diffusion model optimization module to generate a high-precision executable motion sequence after denoising. Through a closed-loop control module, new multi-view RGB images are periodically collected according to the changes in the environment during execution, the GAF is updated, and the motion estimation and optimization steps are iteratively executed until the task is completed.

[0008] The robot motion reasoning method based on the Gaussian action field GAF of the embodiment of the application can further have the following additional technical features: In an embodiment of the application, the dynamic Gaussian action field GAF includes a Gaussian center head, a Gaussian parameter head, and a motion prediction head, which are used to predict the position, feature parameters, and displacement attributes of the Gaussian point cloud, respectively.

[0009] In an embodiment of the application, the point cloud registration module uses an iterative closest point ICP algorithm to calculate the rigid transformation between the Gaussian point clouds of the current frame and the future frame to generate the motion trajectory of the gripper at the end of the robot arm.

[0010] In an embodiment of the application, the diffusion model optimization module combines the RGB guide image related to the motion and the robot state information to optimize the motion sequence through a constraint loss function, and the constraint loss function includes a denoising direction constraint, a noise constraint, and a robot switch action constraint.

[0011] In an embodiment of the application, the closed-loop control module re-collects multi-view RGB images of the current scene after executing the motion sequence every several time steps, inputs the dynamic Gaussian action field modeling module to update the Gaussian point cloud representation, and re-executes the motion estimation and optimization steps.

[0012] To achieve the above purpose, another aspect of the application provides a robot motion reasoning system based on a Gaussian action field GAF, comprising: An image acquisition module is configured to acquire sparse, uncalibrated multi-view RGB images; A feature extraction module is configured to extract mixed scene features of the multi-view RGB images through a visual Transformer backbone network. A dynamic Gaussian action field modeling module is configured to expand a three-dimensional Gaussian point cloud into a dynamic representation, and endow each Gaussian unit with a motion attribute to model the scene geometry and motion evolution synchronously. A multi-modal query module includes a current query sub-module, a future query sub-module, and an action query sub-module, which are respectively configured to generate a Gaussian point cloud representation of a current time step, predict a Gaussian point cloud representation of a future time step, and calculate an initial action sequence of a gripper at the end of a robot arm through a point cloud registration module. An action optimization module is configured to combine the initial action sequence, current multi-view scene images, and robot state information to generate a high-precision executable action sequence after denoising through a diffusion model. A closed-loop control module is configured to periodically acquire new multi-view RGB images during execution of the action sequence, update the dynamic Gaussian action field, and iteratively perform action estimation and optimization steps until the task is completed.

[0013] The robot action reasoning method and system based on the Gaussian action field GAF according to the embodiments of the present application provide a robot action reasoning technology capable of uniformly modeling scene geometry and motion evolution, supporting multi-modal query, and combining diffusion model optimization, thereby realizing high-precision, high-robustness, and high-real-time robot operation.

[0014] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0015] The above and / or additional aspects and advantages of the present application will become apparent and be readily understood from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 is a flowchart of a robot action reasoning method based on a Gaussian action field GAF according to an embodiment of the present application; Figure 2 is a GAF technical architecture diagram according to an embodiment of the present application; Figure 3 is an operation flowchart according to an embodiment of the present application; Figure 4 is a robot action reasoning system structure diagram based on a Gaussian action field GAF according to an embodiment of the present application. DETAILED DESCRIPTION

[0016] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0017] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0018] The robot action reasoning method and system based on Gaussian action field GAF proposed according to the embodiments of the present application will be described below with reference to the drawings.

[0019] Figure 1 is a flowchart of the robot action reasoning method based on Gaussian action field GAF according to the embodiments of the present application, as Figure 1 shown, comprising: S1, input sparse, uncalibrated multi-view RGB images; S2, use a visual Transformer backbone network to extract mixed scene features of the multi-view RGB images; S3, construct a dynamic Gaussian action field GAF, assign learnable motion attributes to each Gaussian cell by extending a three-dimensional Gaussian point cloud, and model the scene geometry and motion evolution synchronously; S4, in response to a current scene reconstruction request, generate a Gaussian point cloud representation of the current time step based on the GAF, and render multi-view scene images; in response to a future state prediction request, predict a Gaussian point cloud representation of a future time step based on the GAF, and generate future view images; in response to an action estimation request, extract Gaussian point clouds related to the end gripper of the robot arm in the current frame and the future frame, and calculate an initial action sequence through a point cloud registration module; S5, input the initial action sequence, the current multi-view scene image, and the robot arm state information into a diffusion model optimization module to generate a high-precision executable action sequence after denoising; S6, through a closed-loop control module, periodically collect new multi-view RGB images according to the changes in the environment during execution, update the GAF, and iteratively perform the action estimation and optimization steps until the task is completed.

[0020] Specifically, the application adopts 3D Gaussian point cloud (3DGS) to extend to dynamic Gaussian action field (GAF), and gives each Gaussian unit a learnable motion attribute, so that it can not only represent the static scene geometry, but also predict the future state change. This method only needs to input any two uncalibrated RGB images, and can reconstruct the current scene and predict the future frame in real time, while generating the initial action sequence. Unlike existing robot control methods, the application realizes spatiotemporal consistent motion reasoning through the three types of query functions (current reconstruction, future prediction, and action estimation) of GAF, and further improves the robustness and accuracy of the action by combining diffusion model optimization.

[0021] When the method is applied to robot control, the following steps are specifically included: Multi-view RGB image input: sparse, uncalibrated multi-view RGB images are collected as initial observation data; Dynamic Gaussian action field (GAF) modeling: a visual Transformer (ViT) backbone is used to extract mixed scene features; the position, feature parameters and displacement vector of the Gaussian point cloud are predicted through the Gaussian center head, Gaussian parameter head and motion prediction head; Multi-modal query and action generation: current query: render high-fidelity new view images of the current scene to support accurate geometric understanding; future query: predict the Gaussian point cloud of the future frame through the motion attribute, which is used for time series dynamic supervision; action query: based on the Gaussian point cloud matching (such as ICP algorithm) of the current and future frames, the initial action sequence of the robot arm is calculated; Diffusion model optimization and closed-loop execution: combine the multi-view observations and initial actions generated by GAF, and use the diffusion model for denoising optimization to generate high-precision executable actions; through the closed-loop framework, the dynamic scene changes are continuously adapted to ensure the task success rate.

[0022] In an embodiment of the application, the Gaussian action field (GAF) is defined as a unified spatiotemporal representation, which associates the Gaussian cell g(x) at each time step t with geometric properties and motion dynamics. Specifically, GAF is parameterized by a continuous function, where each Gaussian contains 3D position, predicted displacement vector, and feature attributes such as color, opacity, rotation, and scale.

[0023]

[0024] The rendering process follows the 3D Gaussian splatting (3DGS) method, and the pixel color is rendered by alpha according to formula 2.

[0025]

[0026] To support current scene reconstruction, future state prediction, and action estimation, the present invention defines three types of queries in GAF as shown in Equation 3: current query obtains the Gaussian position and feature parameters at the current time step, supporting rendering the scene from multiple new viewpoints; future query applies the predicted displacement to the current position to generate the future position, forming the future Gaussian field to render the future view; action query extracts the Gaussian motion attributes related to the operation, calculating the initial action through the point cloud matching of the current frame and the future frame.

[0027]

[0028] The GAF architecture unifies scene representation, dynamic motion prediction, and action reasoning. Its goal is to directly reconstruct a motion-enhanced Gaussian model from sparse, pose-free RGB inputs, supporting downstream temporal queries and operation control. The overall design is shown in Figure 2 Sparse multi-view images are extracted, mixed scene features are extracted, and then decoded through three heads to predict Gaussian point cloud positions, feature parameters, and displacement attributes.

[0029] GAF adopts a geometry-agnostic, pose-free dynamic scene reconstruction method to directly reconstruct a high-fidelity motion-enhanced Gaussian model of the input view through a feedforward network (including a visual Transformer backbone and three dedicated heads). The specific process includes: unregistered images are labeled and spliced, and features are extracted using a shared weight ViT; a decoupled double-head design (Gaussian center head and Gaussian parameter head) is used to predict the Gaussian center and other parameters; the motion prediction head is used to predict the displacement of each Gaussian, representing the motion in the future time interval; the displacement and the current Gaussian center are fused to form the future Gaussian field, and the multi-view images of the current and future states are rendered, and the reconstruction of the motion-enhanced Gaussian is optimized through the loss function; based on the Gaussian model of the current and future frames, the Gaussian point cloud related to the end gripper of the robot arm is extracted, the rigid transformation is estimated through the ICP algorithm, the motion of the end gripper of the robot arm is captured, and the corresponding initial action sequence is generated.

[0030] In one embodiment of the present invention, GAF supports three types of queries, and the initial action of the action query may not be accurate due to interaction noise. For this purpose, the present invention introduces an optimization module based on a diffusion model, which jointly uses the multi-view observations rendered by GAF and the initial action prediction to generate higher-quality denoised actions.

[0031] Figure 3 The operation process: the current and action queries of GAF provide the current multi-view observations and the initial action estimation (left side). These information are then input as conditions into an optimization network to generate executable actions (right side). The process is iterated until the task is completed.

[0032] As shown in Figure 3As shown, the optimization module projects the initial action corresponding robot position to pixel coordinates and renders into the current multi-view RGB image, forming "action-related RGB guidance". These visual cues, together with the initial action and robot state, guide the diffusion model to minimize the constraint loss (including denoising direction, noise, and robot switch action constraints). The denoised action sequence can be directly executed and continuously adapted to dynamic scene changes through a closed-loop framework.

[0033] Further, one embodiment of the robot action inference technology based on the Gaussian world model is described in combination with the drawings in the embodiments of the application as follows: Step S101, data acquisition and training. Multi-view training data is rendered using tasks in RLBench, covering 10 typical robot operation tasks (such as "put down the toilet cover" and "close the notebook", etc.). Each frame renders 20-view RGB images (resolution 256x256), without depth information or accurate camera calibration, relying only on sparse view input. Train the neural network of dynamic Gaussian action field (GAF), including: ViT backbone network: extract mixed features of multi-view images; Gaussian center head: predict the three-dimensional position of Gaussian point cloud; Gaussian parameter head: predict attributes such as color, opacity, rotation and scale; motion prediction head: learn the displacement vector of each Gaussian for future state modeling.

[0034] Step S102, dynamic Gaussian action field (GAF) generation. Input any two uncalibrated RGB images (256x256), extract features through the ViT encoder. Use the Gaussian center head and the Gaussian parameter head to reconstruct the three-dimensional Gaussian point cloud (3DGS) of the current scene. Estimate the motion attributes of each Gaussian through the motion prediction head to form a four-dimensional dynamic Gaussian field (GAF) that supports current reconstruction, future prediction and action query.

[0035] Step S103, multi-view rendering and future prediction: current query: render the RGB image (128x128x3) of the specified view for real-time scene understanding. Future query: apply the motion attributes to the current Gaussian point cloud to generate the Gaussian distribution of the future frame and render the future view image (128x128x3). The output image is used for visual supervision and action optimization to ensure spatiotemporal consistency. Action query: extract the Gaussian point cloud related to the end gripper of the robot arm, calculate the point cloud registration of the current frame and the future frame through the ICP algorithm, and obtain the initial action sequence (including position and attitude changes).

[0036] Step S104, input the initial action, RGB guidance related to multi-view action and mechanical arm state (gripper opening and closing) into the diffusion model to perform action denoising optimization: introduce diffusion time step (Diffusion Time Step) and gripper state (Gripper State); generate smooth and accurate action sequence through the Transformer-based Diffusion Model.

[0037] Step S105, real machine deployment and closed loop execution: convert the optimized action sequence into robot control instructions. Deploy on a real mechanical arm and adjust the action in real time through closed loop feedback: reacquire the RGB image of the current scene every time sequence (8 time steps in the example of the present application) is executed, input the GAF to update the dynamic Gaussian field; iteratively execute steps S102-S104 until the task is completed (such as successful grasping, placing or assembling.

[0038] The robot action reasoning method based on the Gaussian action field GAF according to the embodiment of the present application uses the dynamic Gaussian point cloud to perform high-fidelity reconstruction on the scene, and on the other hand, combines motion attribute prediction and diffusion optimization to significantly improve the adaptability and success rate of robot operation. Without depth information or accurate camera calibration, only two RGB images are needed to run in real time; predict the future state through motion attribute prediction to enhance the adaptability of the robot to complex interactive scenes; combine the geometric reasoning and diffusion optimization of GAF to significantly improve the success rate of operation tasks (experimental verification shows that the success rate is improved by an average of 10.33%).

[0039] The method of the present application can be widely applied to the fields of industrial automation, service robots, augmented reality (AR) and the like, and is especially suitable for precision control, occluded environment interaction and dynamic object operation tasks.

[0040] To realize the above-mentioned embodiments, as Figure 4 shown, the embodiment further provides a robot action reasoning system 10 based on the Gaussian action field GAF, which comprises: An image acquisition module 100 is configured to acquire sparse, uncalibrated multi-view RGB images; A feature extraction module 200 is configured to extract mixed scene features of the multi-view RGB images through a visual Transformer backbone network; A dynamic Gaussian action field modeling module 300 is configured to extend the three-dimensional Gaussian point cloud to a dynamic representation, assign motion attributes to each Gaussian unit, and model the geometric and motion evolution of the scene synchronously; The multi-modal query module 400 includes a current query submodule, a future query submodule and an action query submodule, which are respectively used to generate a Gaussian point cloud representation of a current time step, predict a Gaussian point cloud representation of a future time step, and calculate an initial action sequence of a gripper at the end of a robot arm through a point cloud registration module; The action optimization module 500 is used to combine the initial action sequence, a current multi-view scene image and robot state information, and generate a high-precision executable action sequence after denoising through a diffusion model; The closed-loop control module 600 is used to periodically collect new multi-view RGB images during execution of the action sequence, update the dynamic Gaussian action field, and iteratively perform action estimation and optimization steps until the task is completed.

[0041] Further, the dynamic Gaussian action field modeling module 300 includes a Gaussian center head, a Gaussian parameter head and a motion prediction head, which are respectively used to predict the position, feature parameters and displacement attributes of the Gaussian point cloud.

[0042] Further, an iterative closest point (ICP) algorithm is used to calculate the rigid transformation between the Gaussian point clouds of the current frame and the future frame, and generate a motion trajectory of the gripper at the end of the robot arm.

[0043] Further, the action optimization module 500 combines the RGB guide image related to the action and the robot state information, and optimizes the action sequence through a constraint loss function, which includes a denoising direction constraint, a noise constraint and a robot switch action constraint.

[0044] Further, the closed-loop control module 600 re-collects multi-view RGB images of the current scene after each certain time step of executing the action sequence, inputs the dynamic Gaussian action field modeling module to update the Gaussian point cloud representation, and re-executes the action estimation and optimization steps.

[0045] The robot action inference system based on a Gaussian action field (GAF) according to the embodiment of the application uses a dynamic Gaussian point cloud to perform high-fidelity reconstruction on a scene, and on the other hand, combines motion attribute prediction and diffusion optimization to significantly improve the adaptability and success rate of robot operation.

[0046] In the description of the specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Also, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments or examples. Furthermore, the person skilled in the art can combine and combine the different embodiments or examples described in the specification and the features of the different embodiments or examples, without contradiction.

[0047] In addition, the terms "first", "second", are used only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Thus, the features defined with "first", "second" can include at least one of the features, explicitly or implicitly. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise explicitly specified.

Claims

1. A robot motion reasoning method based on Gaussian Action Field (GAF), characterized in that, Includes the following steps: Input a sparse, unlabeled multi-view RGB image; The mixed scene features of the multi-view RGB images are extracted using a visual Transformer backbone network; A dynamic Gaussian motion field (GAF) is constructed, and each Gaussian unit is given learnable motion properties by expanding the 3D Gaussian point cloud, simultaneously modeling scene geometry and motion evolution; In response to the current scene reconstruction request, a Gaussian point cloud representation for the current time step is generated based on GAF, and multi-view scene images are rendered; in response to the future state prediction request, a Gaussian point cloud representation for the future time step is predicted based on the GAF, and a future view image is generated; in response to the motion estimation request, Gaussian point clouds related to the end gripper of the robotic arm are extracted from the current frame and future frames, and the initial motion sequence is calculated through the point cloud registration module. The initial action sequence, the current multi-view scene image, and the robotic arm state information are input into the diffusion model optimization module to generate a denoised high-precision executable action sequence. The closed-loop control module periodically acquires new multi-view RGB images based on environmental changes during execution, updates the GAF, and iteratively executes motion estimation and optimization steps until the task is completed.

2. The method according to claim 1, characterized in that, The Dynamic Gaussian Action Field (GAF) includes a Gaussian center head, a Gaussian parameter head, and a motion prediction head, which are used to predict the position, feature parameters, and displacement attributes of the Gaussian point cloud, respectively.

3. The method according to claim 2, characterized in that, The point cloud registration module uses the Iterative Closest Point (ICP) algorithm to calculate the rigid transformation between the Gaussian point clouds of the current frame and future frames, and generates the motion trajectory of the end effector gripper of the robotic arm.

4. The method according to claim 3, characterized in that, The diffusion model optimization module combines the motion-related RGB guidance image with the robotic arm state information and optimizes the motion sequence through a constraint loss function, which includes denoising direction constraint, noise constraint, and robotic arm switching motion constraint.

5. The method according to claim 4, characterized in that, After executing the action sequence at several time steps, the closed-loop control module re-acquires multi-view RGB images of the current scene, inputs them into the dynamic Gaussian action field modeling module to update the Gaussian point cloud representation, and re-executes the action estimation and optimization steps.

6. A robot motion reasoning system based on Gaussian Action Field (GAF), characterized in that, include: The image acquisition module is used to acquire sparse, unlabeled multi-view RGB images; The feature extraction module is used to extract the mixed scene features of the multi-view RGB images through the visual Transformer backbone network; The Dynamic Gaussian Action Field Modeling Module is used to extend the 3D Gaussian point cloud into a dynamic representation, assign motion attributes to each Gaussian unit, and simultaneously model the scene geometry and motion evolution. The multimodal query module includes a current query submodule, a future query submodule, and an action query submodule, which are used to generate the Gaussian point cloud representation of the current time step, predict the Gaussian point cloud representation of the future time step, and calculate the initial action sequence of the robotic arm end gripper through the point cloud registration module, respectively. The motion optimization module is used to combine the initial motion sequence, the current multi-view scene image, and the robotic arm state information to generate a denoised high-precision executable motion sequence through a diffusion model. The closed-loop control module is used to periodically acquire new multi-view RGB images, update the dynamic Gaussian action field, and iteratively execute action estimation and optimization steps during the execution of the action sequence until the task is completed.

7. The system according to claim 6, characterized in that, The dynamic Gaussian motion field modeling module includes a Gaussian center head, a Gaussian parameter head, and a motion prediction head, which are used to predict the position, feature parameters, and displacement attributes of the Gaussian point cloud, respectively.

8. The system according to claim 7, characterized in that, The Iterative Closest Point (ICP) algorithm is used to calculate the rigid transformation between the Gaussian point clouds of the current frame and future frames, and to generate the motion trajectory of the end effector gripper of the robotic arm.

9. The system according to claim 8, characterized in that, The motion optimization module combines motion-related RGB guidance images with robotic arm state information and optimizes the motion sequence through a constraint loss function, which includes denoising direction constraints, noise constraints, and robotic arm switching motion constraints.

10. The system according to claim 9, characterized in that, After executing the action sequence at several time steps, the closed-loop control module re-acquires multi-view RGB images of the current scene, inputs them into the dynamic Gaussian action field modeling module to update the Gaussian point cloud representation, and re-executes the action estimation and optimization steps.

Images