A Method for Autonomous Planning of Robots in Complex Tasks Based on Adaptive Learning of Multi-Role Large Models

Abstract

This invention belongs to the field of robot task planning technology and discloses a method for autonomous planning of complex robot tasks based on adaptive learning of a multi-role large model. The method includes: Step 1, using an open-source multimodal large model with prior knowledge to construct a task analysis model, a task execution model, and an effect evaluation model to assist robot decision-making; Step 2, based on the three large models constructed in Step 1, using meta-learning technology to train these three models simultaneously, dynamically adjusting model parameters during the decision-making process using perceptual feedback information, thereby enabling the robot to adapt to new scenarios when performing tasks. This invention constructs multiple multimodal large models of robots with different roles and uses a meta-learning-based adaptive learning method to improve the adaptability of the robot language-specified task planning model based on the large model in unknown environments, enabling it to have autonomous learning capabilities and continuously reflect during task execution to quickly adapt to new environments.

Description

Technical Field

[0001] This invention belongs to the field of robot task planning technology, and in particular relates to a method for autonomous planning of complex robot tasks based on adaptive learning of a multi-role large model. Background Technology

[0002] Existing large-scale robot models can perform complex tasks after building world models for training scenarios, but biases in rule knowledge reduce task success rates when transferred to new environments. During knowledge transfer from the training environment to the unknown environment, the increasing amount of knowledge learned by the large-scale robot model gradually leads to the model forgetting previously learned knowledge. In new scenarios, the large-scale model learns more fundamental world rules through interaction and feedback with the environment, retaining old knowledge while learning new knowledge. This improves the robot's adaptability to zero-shot scenarios, especially in the face of new anomalies, where it can gradually learn and improve its decision-making performance on these anomalies. This is essential for achieving efficient planning of diverse tasks in unknown environments. Therefore, researching efficient adaptive learning for robots in zero-shot scenarios in unknown environments is one of the keys to exploring techniques for enhancing robot planning efficiency based on large-scale models. Summary of the Invention

[0003] The purpose of this invention is to provide a method for autonomous planning of complex robot tasks based on adaptive learning of a multi-role large model, so as to solve the above-mentioned technical problems.

[0004] To address the aforementioned technical problems, the specific technical solution of the robot's autonomous planning method for complex tasks based on multi-role large model adaptive learning of the present invention is as follows:

[0005] A method for autonomous planning of complex tasks for robots based on multi-role large model adaptive learning includes the following steps:

[0006] Step 1: Construct a large multi-role robot model: Utilize an open-source multimodal model with prior knowledge to construct a large multi-role model with prior knowledge that assists the robot in decision-making.

[0007] Step 2: Enhance the robot's ability to adapt to unknown environments using meta-learning: Based on the multi-role large model built in Step 1, use meta-learning technology and adaptive learning methods to train the three models simultaneously. During the decision-making process, use perceptual feedback information to dynamically adjust the model parameters, so as to enable the robot to adapt to new scenarios when performing tasks.

[0008] Furthermore, step one utilizes existing pre-trained open-source multimodal large models, including LLaVA and Yi-VL-6B.

[0009] Furthermore, step one, which involves constructing a large model with three different roles, includes: a large model for task analysis. Large-scale task execution model and large model for effect evaluation The large-scale task analysis model Responsible for visual analysis of the robot's current environment and planning tasks specified by language, breaking them down into multiple sub-tasks, and calling different APIs. Large-scale models for various task execution The large-scale performance evaluation model is responsible for reasoning about actual visual perception and performing different sub-tasks. During task execution, the system analyzes the current visual perception, assesses the effectiveness of the current operation, provides feedback, and makes fine adjustments. and Subsequently, a lightweight fine-tuning approach was adopted, and the three large models mentioned above were adjusted in step two. , , Conduct training.

[0010] Furthermore, step two involves the three large models constructed in step one. , , Based on this, we use an adaptive learning method based on meta-learning to train these three models simultaneously.

[0011] Furthermore, the training process in step two is as follows: The processes already experienced by the robot are considered as the Support Set in meta-learning, and the processes not yet experienced are considered as the Query Set. An interactive exploratory loss function is designed using the MAML algorithm. It consists of two parts: the task result reward function. and the loss function of the process already completed These are respectively used as the loss function for meta-learning on the part where the process has not yet been experienced and the loss function for the part where the process has been experienced. For each robot task, a large task analysis model is designed to determine whether the task is successful. Large-scale task execution model A series of output actions Good or bad .

[0012] Furthermore, step two includes the following specific steps:

[0013] First, let the large model evaluate the effects. The robot analyzes its current visual perception combined with the text-defined task objective, and generates a text-based feedback description of the task execution result, indicating whether the task was performed well or poorly. Use it as a prompt, along with the original task text. and robot visual perception Simultaneously input the large task analysis model Make minor adjustments as prompted:

[0014] ,

[0015] At the same time, utilize large-scale performance evaluation models. Come to learn , recorded as ,in for The parameters are based on the processes that have already been completed. , including all the The robot observes video frames and executes actions before each step, outputting the quality of these actions. During each navigation step, the robot uses a large-scale performance evaluation model. What I learned Based on the process prior to this step And the current visual perception results, in step size Optimize the previous large task execution model Model parameters Thus, new model parameters are obtained. For a common The task execution process is optimized step by step. and The loss function is calculated as follows:

[0016] ,

[0017] During the training process described above, three large models , , It optimizes simultaneously through adversarial and cooperative methods, and after training, during the prediction process, Through multimodal sensing feedback that interacts with the environment , The model parameters are continuously and dynamically adjusted, with fine-tuning done interactively, to prevent similar misjudgments from occurring again.

[0018] The method for autonomous planning of complex robot tasks based on adaptive learning of a multi-role large model in this invention has the following advantages: This invention constructs a multimodal large model of a robot with multiple roles and responsibilities (including a task analysis large model). Large-scale task execution model Large-scale model for effect evaluation Through mutual cooperation and supervision, and by utilizing meta-learning-based adaptive learning methods, the efficiency of planning models for robot language-specified tasks based on large models in unknown environments can be improved, especially enabling them to have the ability to learn autonomously and to continuously reflect on the execution of tasks in order to quickly adapt to new environments. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to the present invention. Detailed Implementation

[0020] To better understand the purpose, structure, and function of this invention, the following detailed description, in conjunction with the accompanying drawings, provides a more comprehensive overview of the method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning.

[0021] like Figure 1 As shown, the method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning of the present invention includes the following steps:

[0022] Step 1: Construct a large-scale model of the robot with multiple roles

[0023] We will initialize and construct three different large models using existing pre-trained open-source multimodal large models (such as LLaVA, Yi-VL-6B, etc.): a task analysis large model. Large-scale task execution model Large-scale model for effect evaluation Among them, the large model of task analysis It is responsible for visually analyzing the robot's current environment and planning tasks specified by language, breaking them down into multiple sub-tasks (such as navigation, grasping, cleaning, and water pouring), and calling different APIs. Large-scale models for various task execution It is responsible for reasoning about actual visual perception and performing different sub-tasks, while simultaneously evaluating the performance of the large model. During task execution, the system analyzes the current visual perception, assesses the effectiveness of the current operation, provides feedback, and makes fine adjustments. and Subsequently, lightweight fine-tuning methods (such as Adapter, LoRA, Prompt Tuning, etc.) are used to fine-tune the three large models mentioned above through the following step two. , , Conduct training.

[0024] Step 2: Enhance the robot's adaptive ability in unknown environments using meta-learning

[0025] The three major models constructed in step one , , Building upon this foundation, an adaptive learning method based on meta-learning is used to train three models simultaneously. During the decision-making process, perceptual feedback information is used to dynamically adjust the model parameters, enabling the robot to adapt to new scenarios when performing tasks. The main process is as follows: the processes already experienced by the robot are considered as the support set in meta-learning, and the processes not yet experienced are considered as the query set. An interactive exploratory loss function is designed using the MAML algorithm. It mainly consists of two parts: the task result reward function. and the loss function of the process already completed , respectively, serve as the loss function for meta-learning on the part where the process has not yet been experienced and the part where the process has been experienced. For each robot task, a task analysis model is designed to determine whether the task is successful. Large-scale task execution model A series of output actions Good or bad .

[0026] First, let the large model evaluate the effects. The robot analyzes its current visual perception combined with the text-defined task objective, and generates a text-based feedback description of the task execution result, indicating whether the task was performed well or poorly. Use it as a prompt, along with the original task text. and robot visual perception Simultaneously input the large task analysis model Make minor adjustments as prompted:

[0027]

[0028] At the same time, utilize large-scale performance evaluation models. Come to learn , recorded as ,in for The parameters are based on the processes that have already been completed. , including all the The robot observes video frames and executes actions before each step, outputting the quality of these actions. During each navigation step, the robot uses a large-scale performance evaluation model. What I learned Based on the process prior to this step And the current visual perception results, in step size Optimize the previous large task execution model Model parameters Thus, new model parameters are obtained. Therefore, for a common The task execution process is optimized step by step. and The loss function is calculated as follows:

[0029]

[0030] During the training process described above, three large models , , It optimizes simultaneously through adversarial and cooperative methods. And after training, during the prediction process... This allows for multimodal perception feedback, such as visual feedback, through interaction with the environment. , The model parameters are continuously and dynamically adjusted, with fine-tuning done interactively, to prevent similar misjudgments from occurring again.

[0031] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for autonomous planning of complex tasks in robots based on adaptive learning of a multi-role large model, characterized in that, Includes the following steps: Step 1: Construct a large multi-role robot model: Utilize an open-source multimodal model with prior knowledge to construct a large multi-role model with prior knowledge that assists the robot in decision-making. Step one involves constructing three large models with different roles, including: a large model for task analysis M. a Task Execution Model M e And the large model M for effect evaluation v The task analysis big model M a Responsible for visual analysis of the robot's current environment and planning the language-specified tasks, breaking them down into multiple sub-tasks, and calling different APIs. e Various task execution models M e The large model M, responsible for reasoning about actual visual perception and performing different sub-tasks, is used for performance evaluation. v During task execution, the current visual perception is analyzed to determine the effectiveness of the current operation, and feedback and fine-tuning are provided. v and M e Subsequently, a lightweight fine-tuning approach was adopted, and the three large models M mentioned above were adjusted in step two. a M e M v Conduct training; Step 2: Enhance the robot's ability to adapt to unknown environments using meta-learning: Based on the multi-role large model built in Step 1, the multi-role large model built in Step 1 is trained simultaneously using meta-learning adaptive learning methods. During the decision-making process, the model parameters are dynamically adjusted using perceptual feedback information to enable the robot to adapt to new scenarios when performing tasks. During training, three large models M a Mx, M v It optimizes simultaneously using both adversarial and cooperative methods, and after training, during the prediction process, M... v M is monitored through multimodal sensing feedback that interacts with the environment. a M e The model parameters are continuously and dynamically adjusted, with fine-tuning done interactively, to prevent similar misjudgments from occurring again.

2. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 1, characterized in that, Step one utilizes existing pre-trained open-source multimodal large models, including LLaVA and Yi-VL-6B.

3. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 1, characterized in that, The second step involves the construction of three large models M in the first step. a M e M v Based on this, we use an adaptive learning method based on meta-learning to train these three models simultaneously.

4. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 3, characterized in that, The training process in step two is as follows: The processes already experienced by the robot are considered as the Support Set in meta-learning, and the processes not yet experienced are considered as the Query Set. An interactive exploratory loss function is designed using the MAML algorithm. It consists of two parts: the task result reward function. and the loss function of the process already completed These are respectively used as the loss function for meta-learning on the process before it has been experienced and the loss function for the process after it has been experienced. For each robot task, a task analysis large model M is designed to determine whether the task is successful. a Task Execution Model M e A series of output actions Good or bad 5. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 4, characterized in that, Step two includes the following specific steps: First, let the large model M evaluate the effect. v The robot analyzes its current visual perception combined with the text-defined task objective, and generates a text-based feedback description of the task execution result, indicating whether the task was performed well or poorly. Use it as a prompt, along with the original task text T o And robot visual perception V R Simultaneously input the large task analysis model M a Make minor adjustments as prompted: At the same time, the large model M is used for effect evaluation. v Come to learn Recorded as Where φ is M v The parameters are based on the processes that have already been completed. This includes all observed video frames and executed actions up to step i, outputting the quality of these actions. The robot uses the performance evaluation model M to assess the effectiveness of each navigation step. v What I learned Based on the process prior to this step Based on the current visual perception results, optimize the previous task execution large model M with a step size β. e Model parameters θ are used to obtain new model parameters. For a task execution process with a total of q steps, the loss function for simultaneously optimizing θ and φ is calculated as follows:

Images