A batch execution method and system based on expected-observation consistency verification
By adopting a batch execution method based on expectation-observation consistency verification, the problems of low efficiency in iterative execution and batch execution errors are solved, achieving efficient and reliable agent task execution and improving the execution efficiency and performance of the agent.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN INST OF TECH
- Filing Date
- 2026-02-11
- Publication Date
- 2026-06-09
AI Technical Summary
In existing technologies, iterative execution methods have low inference efficiency in long-view tasks, while batch execution methods are prone to cascading errors, leading to a decline in agent performance and making it difficult to balance high performance and high efficiency.
A batch execution method based on expectation-observation consistency verification is proposed. By generating an action-expected observation sequence, the method executes and verifies the consistency of feedback from the real environment. If there is a discrepancy, the method interrupts and re-plans the action sequence. By combining batch execution with timely replanning, the method ensures that the strategy conforms to the reality of the environment.
It achieves high efficiency and high performance in long-view tasks, reduces the number of model calls, improves execution efficiency and responds to environmental feedback in a timely manner, while maintaining compatibility with existing planning methods.
Smart Images
Figure CN122174869A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a batch execution method and system based on expectation-observation consistency verification, belonging to the field of artificial intelligence technology. Background Technology
[0002] Task planning requires agents to explore the environment autonomously and execute human instructions. Recent research has proposed numerous task planning methods based on LLM (Limited Learning Model), most of which consist of peripheral enhancement modules (such as task decomposition, reflective replanning, and memory systems) and a core execution module. Existing methods typically focus on the design of the enhancement modules and share similar core execution module designs. These execution modules can be categorized into two paradigms: iterative execution and batch execution.
[0003] The iterative execution method executes each action immediately in the environment and updates the agent's context based on environmental feedback, then repeats the planning process based on the updated context. Its advantages include timely monitoring of real-time feedback during planning, resulting in strong robustness and performance. However, the iterative execution method requires reprocessing the ever-growing context at each step, leading to lower inference efficiency and higher inference costs in long-view tasks.
[0004] Batch execution generates a sequence of actions at a time, executes them sequentially in the environment, and updates the agent's context in batches based on environmental feedback. The planning process is then repeated based on the updated context. Since each planning iteration can execute multiple environmental actions, batch execution offers higher inference efficiency. However, because the strategy cannot be adjusted based on environmental feedback during batch execution, it is prone to cascading errors, leading to a degraded agent performance. Summary of the Invention
[0005] The technical problem addressed by this invention is how to construct an execution mechanism that possesses both high performance in iterative execution and high efficiency in batch execution, while ensuring its compatibility with existing planning methods. Furthermore, this invention proposes a batch execution method and system based on expectation-observation consistency verification.
[0006] The technical solution adopted by this invention to solve the above problems is: a batch execution method based on expectation-observation consistency verification, comprising:
[0007] Step 1: Based on historical observation information, the agent generates an action-expected observation sequence that includes actions and corresponding expected observations; Step 2: The agent executes the current action in the sequence in order, obtains feedback from the real environment, and updates the agent context; Step 3: Calculate the similarity between the actual environmental feedback and the expected observation, and determine their consistency. If they are consistent, continue to execute the next action in the sequence. If they are inconsistent, interrupt the action execution and abandon the unexecuted actions, and return to regenerate the action-expected observation sequence.
[0008] Furthermore, the action-expected observation sequence in step 1 takes the form of: ,in, The continuous actions generated for the intelligent agent. The expected observations generated for the agent. n This represents the number of actions.
[0009] Furthermore, in step 2, the context update is based on the execution result of the current action.
[0010] Furthermore, step 3 specifically includes: The similarity between the expected observation and the actual observation is calculated using a non-exact matching similarity calculation method. If the expected observation and the actual environmental feedback are consistent, the next action in the sequence is executed sequentially. If they are inconsistent, the action execution is interrupted and the unexecuted actions are abandoned. Then, the process returns to step 1 to re-plan the action-expected observation sequence.
[0011] Furthermore, this invention proposes a batch execution system based on expectation-observation consistency verification. The batch execution system adopts a batch execution method based on expectation-observation consistency verification. Through batch execution and timely replanning, and by verifying whether the agent's strategy conforms to the environmental reality based on the expected environmental feedback of the generated actions, it achieves high inference efficiency and iterative execution of task planning.
[0012] The beneficial effects of this invention are: 1. This invention verifies whether an agent's policy conforms to reality by comparing the consistency between the expected environmental feedback and the actual environmental feedback in an LLM (Low-Level Model). When the expected environmental feedback matches the actual environmental state, EOCV (Electronic Operational Continuous Vehicle) executes actions sequentially; once the expected environmental feedback does not match the actual environmental state, execution stops and a replanning is triggered. By combining batch execution with timely replanning, EOCV reduces the number of model calls and saves costs, while also responding promptly to environmental feedback, possessing both the high inference efficiency of batch execution and the high performance of iterative execution.
[0013] 2. The batch execution system proposed in this invention maintains an interface consistent with existing execution modules, thus easily replacing execution modules in existing planning methods, thereby improving the execution efficiency or performance of state-of-the-art planning methods. Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating a batch execution method based on expectation-observation consistency verification. Detailed Implementation
[0015] like Figure 1 As shown, the steps of a batch execution method based on expectation-observation consistency verification described in this embodiment include: S1: Based on historical observation information, generate an action-expected observation sequence that includes actions and corresponding expected observations; The agent generates actions based on previous observations—the expected sequence of observations. , The continuous actions generated for the intelligent agent. The continuous actions generated for the intelligent agent. n This represents the number of actions.
[0016] S2: The agent executes the current action in the sequence in order, obtains the environmental observation action, updates the agent context, and obtains the expected observation action; The body executes the actions in the sequence sequentially. Obtain environmental observation The context update is based on the execution result of the current action and the acquired environmental observation sequence.
[0017] S3: Calculate the similarity between the predicted observation action and the environmental observation action, determine consistency, and determine the next step of execution.
[0018] The similarity between the expected observation and the actual observation is calculated using a non-exact matching similarity calculation method. If the expected observation and the actual environmental feedback are consistent, the next action in the sequence is executed sequentially. If they are inconsistent, the action execution is interrupted and the unexecuted actions are abandoned. Then, the process returns to S1 to re-plan the action-expected observation sequence.
[0019] Furthermore, this invention proposes a batch execution system based on expectation-observation consistency verification. This system is a model-independent execution module (EOCV) that, based on the aforementioned expectation-observation consistency verification mechanism, achieves both high performance and high inference efficiency through batch execution and timely replanning. Moreover, the system maintains an interface consistent with existing execution modules, thus easily replacing execution modules in existing planning methods, thereby improving the execution efficiency or performance of state-of-the-art planning methods.
[0020] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.
Claims
1. A batch execution method based on expectation-observation consistency verification, characterized in that, include: Step 1: Based on historical observation information, the agent generates an action-expected observation sequence that includes actions and corresponding expected observations; Step 2: The agent executes the current action in the sequence in order, obtains feedback from the real environment, and updates the agent context; Step 3: Calculate the similarity between the actual environmental feedback and the expected observation, and determine their consistency. If they are consistent, continue to execute the next action in the sequence. If they are inconsistent, interrupt the action execution and abandon the unexecuted actions, and return to regenerate the action-expected observation sequence.
2. The batch execution method based on expectation-observation consistency verification according to claim 1, characterized in that, The action-expected observation sequence in step 1 is in the form of ,in, The continuous actions generated for the intelligent agent. The expected observations generated for the agent. n This represents the number of actions.
3. The batch execution method based on expectation-observation consistency verification according to claim 1, characterized in that, In step 2, the context update is based on the execution result of the current action.
4. The batch execution method based on expectation-observation consistency verification according to claim 1, characterized in that, Step 3 specifically includes: The similarity between the expected observation and the actual observation is calculated using a non-exact matching similarity calculation method. If the expected observation and the actual environmental feedback are consistent, the next action in the sequence is executed sequentially. If they are inconsistent, the action execution is interrupted and the unexecuted actions are abandoned. Then, the process returns to step 1 to re-plan the action-expected observation sequence.
5. A batch execution system based on expectation-observation consistency verification, applied to the batch execution method based on expectation-observation consistency verification as described in any one of claims 1-4, characterized in that, include: The batch execution system adopts a batch execution method based on expectation-observation consistency verification. Through batch execution and timely replanning, and by verifying whether the agent's strategy conforms to the environmental reality based on the expected environmental feedback of the generated actions, it achieves high inference efficiency and iterative execution of task planning.