Robot complex task autonomous planning method based on multi-role large model adaptive learning

Abstract

The invention belongs to the technical field of robot task planning, and discloses a robot complex task autonomous planning method based on multi-role large model adaptive learning. Constructing a task analysis large model, a task execution large model and an effect evaluation large model for assisting robot decision making; and 2, on the basis of the three large models constructed in the step 1, training the three models at the same time by utilizing a meta-learning technology, and dynamically adjusting model parameters by utilizing perception feedback information in a decision making process to realize the self-adaption of the robot to a new scene when the robot executes a task. According to the method, a plurality of robot multi-mode large models with different labor division roles are constructed, and the adaptive learning method based on meta-learning is utilized, so that the adaptation efficiency of the planning model of the robot language specified task based on the large model in an unknown environment is improved, and the planning model has the ability of autonomous learning; and continuous reflection in the execution process of the task is carried out to quickly adapt to a new environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot task planning, 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 Art

[0002] The existing robot large model can perform complex tasks after building a world model for the training scene, but the deviation of rule knowledge when migrating to a new environment will reduce the success rate of the task. In the process of knowledge transfer from the training environment to the unknown environment, the continuous increase of knowledge learned will gradually cause the model to forget the previously learned knowledge. The large model uses the interactive feedback with the environment in the new scene to learn more essential world rules, retaining old knowledge while learning new knowledge, improving the robot's adaptability when facing zero-sample scenes, especially when facing new abnormal situations, it can gradually learn and improve its decision-making performance for these abnormalities, so as to achieve efficient planning of diversified tasks for robots in unknown environments. Therefore, studying the problem of efficient adaptive learning of robots in zero-sample scenes in unknown environments is one of the keys to exploring robot planning efficiency enhancement technology based on large models. Summary of the invention

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

[0004] In order to solve the above technical problems, the specific technical solution of the method for autonomous planning of complex tasks of robots based on multi-role large model adaptive learning of the present invention is as follows: The method for autonomous planning of complex tasks of robots based on adaptive learning of multi-role large models includes the following steps: Step 1: Build a robot multi-role big model: Use an open source multi-modal big model with prior knowledge to build a variety of multi-role big models with prior knowledge to assist robot decision-making; Step 2. Use meta-learning to enhance the robot's ability to adapt to unknown environments: Based on the multi-role large model constructed in step 1, use meta-learning technology to adaptively learn methods, train the three models at the same time, and use perceptual feedback information to dynamically adjust model parameters during the decision-making process to enable the robot to adapt to new scenarios when performing tasks.

[0005] Furthermore, the step 1 utilizes existing pre-trained open source multimodal large models including LLaVA and Yi-VL-6B.

[0006] Furthermore, the step 1 of constructing the big models of three different roles includes: task analysis big model , Task Execution Model and large-scale model for effect evaluation , the task analysis model Responsible for visual analysis of the robot's current environment and planning the tasks specified by the language, breaking them down into multiple subtasks and calling different , various tasks execute large models Responsible for reasoning about actual visual perception and performing different subtasks, the effect evaluation model During the task execution, analyze the current visual perception, judge the effect of the current operation, provide feedback and make fine adjustments and . Then, a lightweight fine-tuning method is adopted to adjust the above three large models through step 2. , , Conduct training.

[0007] Furthermore, the three large models constructed in step 1 in step 2 , , On this basis, the three models are trained simultaneously using an adaptive learning method based on meta-learning.

[0008] Furthermore, the training process of step 2 is as follows: the process that the robot has experienced is regarded as the support set in meta-learning, and the process that has not been experienced is regarded as the query set, and the interactive exploratory loss function is designed using the MAML algorithm. , divided into two parts: task result reward function And the loss function has been experienced , respectively as the loss function of meta-learning on the process that has not been experienced and the loss function of the process that has been experienced. For each robot task, a large task analysis model is designed to judge whether the task is successful. , Task Execution Model Output a series of actions Good or bad .

[0009] Furthermore, the step 2 includes the following specific steps: First, let's evaluate the big model Learning to analyze the robot's current visual perception combined with the task objectives specified in the text, and generate a text-based result feedback description of the task execution results , using it as a prompt, together with the original task text and robot visual perception Simultaneous input of the task analysis large model To fine-tune it: , At the same time, using the effect evaluation large model Come and learn , denoted as ,in for The parameters of the , including all The observed video frames and the actions performed before the step are output, and the quality of these actions is output. When the robot navigates at each step, the effect evaluation model is used What I Learned According to the process before this step and the current visual perception result, with a step length Optimizing the previous task execution large model Model parameters , thus obtaining new model parameters , for a common Step by step task execution process, while optimizing and The loss function is calculated as follows: , In the above training process, the three large models , , It is optimized simultaneously in an adversarial and cooperative manner. After the training is completed, in the prediction process, Through multimodal perception feedback interaction with the environment , The model parameters are continuously and dynamically adjusted, and fine-tuned interactively to prevent similar misjudgments from happening again.

[0010] The method for autonomous planning of complex tasks of robots based on adaptive learning of multi-role large models of the present invention has the following advantages: the present invention constructs a multi-modal large model of robots with multiple different roles (including a large model for task analysis) , Task Execution Model , large model for effect evaluation ), through mutual cooperation and supervision, and using adaptive learning methods based on meta-learning, the adaptation efficiency of the planning model of robot language-specified tasks based on large models in unknown environments is improved, especially enabling it to have the ability of autonomous learning, and constantly reflect on the execution of tasks to quickly adapt to new environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 The present invention is a flow chart of a method for autonomous planning of complex robot tasks based on adaptive learning of a multi-role large model. DETAILED DESCRIPTION

[0012] In order to better understand the purpose, structure and function of the present invention, the method of autonomous planning of complex tasks of robots based on adaptive learning of multi-role large models of the present invention is further described in detail below in conjunction with the accompanying drawings.

[0013] like Figure 1 As shown, the method for autonomous planning of complex tasks of robots based on adaptive learning of multi-role large models of the present invention comprises the following steps: Step 1: Build a large robot model with multiple roles Use existing pre-trained open source multimodal large models (such as LLaVA, Yi-VL-6B, etc.) to initialize and build large models of three different roles: Task analysis large model , Task Execution Model , large model for effect evaluation , where the task analysis model Responsible for visual analysis of the current robot environment and planning the tasks specified by the language, breaking them down into multiple subtasks (such as navigation, grasping, cleaning, pouring water, etc.), calling different , various tasks execute large models Responsible for reasoning about actual visual perception and performing different subtasks. At the same time, the effect evaluation model During the task execution, analyze the current visual perception, judge the effect of the current operation, provide feedback and make fine adjustments and . Then, a lightweight fine-tuning method (such as Adapter, LoRA, Prompt Tuning, etc.) is used to tune the above three large models through the following step 2. , , Conduct training.

[0014] Step 2: Use meta-learning to enhance the robot's ability to adapt to unknown environments The three large models constructed in step 1 , , Based on this, we use an adaptive learning method based on meta-learning to train the three models at the same time. In the decision-making process, we use the perception feedback information to dynamically adjust the model parameters to achieve the robot's adaptation to new scenarios when performing tasks. The main process is: consider the processes that the robot has experienced as the support set in meta-learning, and the processes that have not yet been experienced as the query set, and use the MAML algorithm to design an interactive exploratory loss function. , mainly divided into two parts: task result reward function And the loss function has been experienced , respectively as the loss function of meta-learning on the process that has not been experienced and the loss function of the process that has been experienced. For each robot task, a large task analysis model is designed to judge whether the task is successful. , Task Execution Model Output a series of actions Good or bad .

[0015] First, let's evaluate the big model Learning to analyze the robot's current visual perception combined with the task objectives specified in the text, and generate a text-based result feedback description of the task execution results , using it as a prompt, together with the original task text and robot visual perception Simultaneous input of the task analysis model To fine-tune it:

[0016] At the same time, using the effect evaluation large model Come and learn , denoted as ,in for The parameters of the , including all The robot uses the effect evaluation model to evaluate the performance of each step. What I Learned According to the process before this step and the current visual perception result, with a step length Optimizing the previous task execution large model Model parameters , thus obtaining new model parameters Therefore, for a common Step by step task execution process, while optimizing and The loss function is calculated as follows:

[0017] In the above training process, the three large models , , It is optimized simultaneously in a confrontational and cooperative manner. After the training is completed, in the prediction process, It can be realized through multimodal perception feedback such as vision that interacts with the environment. , The model parameters are continuously and dynamically adjusted, and fine-tuned interactively to prevent similar misjudgments from happening again.

[0018] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.

Claims

1. A method for autonomous planning of complex robot tasks based on adaptive learning of multi-role large models, characterized in that: The steps include: Step 1: Build a robot multi-role big model: Use an open source multi-modal big model with prior knowledge to build a variety of multi-role big models with prior knowledge to assist robot decision-making; Step 2. Use meta-learning to enhance the robot's ability to adapt to unknown environments: Based on the multi-role large model constructed in step 1, use meta-learning technology to adaptively learn methods, and simultaneously train the multi-role large model constructed in step 1. During the decision-making process, use perceptual feedback information to dynamically adjust model parameters to enable the robot to adapt to new scenarios when performing tasks.

2. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 1 is characterized in that: The step 1 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 adaptive learning of multi-role large models according to claim 1 is characterized in that: The step 1 of constructing the big models of three different roles includes: the big model of task analysis , Task Execution Model and large-scale model for effect evaluation , the task analysis model Responsible for visual analysis of the robot's current environment and planning the tasks specified by the language, breaking them down into multiple subtasks and calling different , various tasks execute large models Responsible for reasoning about actual visual perception and performing different subtasks, the effect evaluation model During the task execution, analyze the current visual perception, judge the effect of the current operation, provide feedback and make fine adjustments and ; The lightweight fine-tuning method is then used to adjust the above three large models through step 2 , , Conduct training.

4. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 1 is characterized in that: The three large models constructed in step 1 in step 2 , , On this basis, the three models are trained simultaneously using an adaptive learning method based on meta-learning.

5. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 4 is characterized in that: The training process of step 2 is as follows: the process that the robot has experienced is regarded as the support set in meta-learning, and the process that has not been experienced is regarded as the query set, and the interactive exploratory loss function is designed using the MAML algorithm. , divided into two parts: task result reward function And the loss function has been experienced , respectively as the loss function of meta-learning on the process that has not been experienced and the loss function of the process that has been experienced. For each robot task, a large task analysis model is designed to judge whether the task is successful. , Task Execution Model Output a series of actions Good or bad .

6. The method for autonomous planning of complex robot tasks based on multi-role large model adaptive learning according to claim 1 is characterized in that: The step 2 includes the following specific steps: First, let's evaluate the big model Learning to analyze the robot's current visual perception combined with the task objectives specified in the text, and generate a text-based result feedback description of the task execution results , using it as a prompt, together with the original task text and robot visual perception Simultaneous input of the task analysis model To fine-tune it: ， At the same time, using the effect evaluation large model Come and learn , denoted as ,in for The parameters of the , including all The observed video frames and the actions performed before the step are output, and the quality of these actions is output. When the robot navigates at each step, the effect evaluation model is used What I Learned According to the process before this step and the current visual perception result, with a step length Optimizing the previous task execution large model Model parameters , thus obtaining new model parameters , for a common Step by step task execution process, while optimizing and The loss function is calculated as follows: ， In the above training process, the three large models , , It is optimized simultaneously in an adversarial and cooperative manner. After the training is completed, in the prediction process, Through multimodal perception feedback interaction with the environment , The model parameters are continuously and dynamically adjusted, and fine-tuned interactively to prevent similar misjudgments from happening again.

Images