Dual-Language-Model Control for Error-Resilient Autonomous Decisions

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing artificial intelligence architectures for autonomous systems, such as robots and vehicles, suffer from poor decision-making precision due to decoupled task modules, where errors in upstream modules lead to incorrect downstream decisions.

Innovation Solution

An end-to-end data processing method using a first and second language model to integrate perception, prediction, and decision-making tasks, allowing for unified handling of various sensing requirements through interaction between the models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of manufacture

If task modules are decoupled with clear upstream-downstream dependencies, then interpretability and ease of development are improved, but decision-making precision deteriorates due to error propagation

Engineering Contradiction:
Improveease of developmentVSAvoiddecision-making precision
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The patent segments the decision-making process into multiple specialized modules (perception module, prediction module, decision module) that process different aspects of the control task. Each module focuses on specific functions while maintaining modular architecture for ease of development, yet they work together as an integrated system to improve overall decision-making precision.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent implements feedback mechanisms where downstream modules provide information back to upstream modules. The decision module feeds back to the prediction and perception modules, allowing error correction and iterative refinement of decisions, thus preventing error propagation while maintaining modular structure.

Inventive Principle:
Principle #23Feedback

2Ease of repair

If task modules are decoupled with clear upstream-downstream dependencies, then ease of maintenance is improved, but control precision deteriorates due to error propagation

Engineering Contradiction:
Improveease of maintenanceVSAvoidcontrol precision
Core Design Contradiction:
Ease of repairVSMeasurement precision

Solution Approach 1:

The patent segments the control system into independently maintainable modules (perception, prediction, decision) that can be updated and maintained separately, preserving ease of maintenance while working together as an integrated system to achieve high control precision.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Feedback loops allow downstream modules to correct errors from upstream modules, maintaining high control precision even as modules are independently maintained and updated over time.

Inventive Principle:
Principle #23Feedback

3Measurement precision

If language models are used for end-to-end processing, then decision-making precision is improved, but computational complexity increases

Engineering Contradiction:
Improvedecision-making precisionVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the language model processing into specialized modules (first language model for scenario understanding, second language model for decision-making) that handle specific tasks, reducing overall computational complexity while maintaining high decision-making precision through focused processing.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4708072A1Data processing method and related apparatus therefor
Publication Date: 2026.03.11 HUAWEI TECH CO LTD
  • EP4708072A1 patent drawingFigure 1a
  • EP4708072A1 patent drawingFigure 1b
  • EP4708072A1 patent drawingFigure 1c

AI summary

A data processing method is applied to the field of artificial intelligence, and includes: obtaining scenario information based on sensor data and a first prompt text by using a first language model, where a target object is a vehicle or a robot; generating a second prompt text based on the scenario information and a current execution objective of the target object by using a second language model, where the second prompt text indicates a query for information that needs to be known by the target object when the target object completes the execution objective and that is not included in the scenario information; obtaining, based on the sensor data and the second prompt text by using the first language model, a reply text corresponding to the second prompt text; and obtaining a control instruction of the target object based on the reply text, the scenario information, and the current execution objective by using the second language model.