Autonomous Vehicle Runtime Monitoring With Vision-Language Risk Scoring

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional runtime monitoring systems for autonomous vehicles fail to consider the broader context of the environment and often make erroneous decisions due to misinterpretation of obstacles or traffic elements, posing safety risks.

Innovation Solution

A vision-language model (VLM) powered runtime monitoring system that generates a risk score by analyzing sensor data, obstacle detections, and future vehicle behavior plans, using embeddings and natural language prompts to evaluate safety risks and adjust vehicle control accordingly.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If heuristic rules are used for runtime monitoring, then the system can check consistency of sensor information, but the system fails to understand broader environmental context and may make erroneous decisions

Engineering Contradiction:
Improveaccuracy of safety assessmentVSAvoidunderstanding of environmental context
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent introduces a large language model (LLM) as an intermediary component between the heuristic monitoring system and the safety assessment process. The LLM receives sensor data, obstacle detections, and vehicle behavior plans, then generates natural language explanations that capture broader environmental context. This intermediary enables the system to maintain the computational efficiency of heuristics while gaining contextual understanding through the LLM's ability to reason about environmental semantics and generate human-like safety assessments.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If conventional runtime monitoring is used, then the system can detect obstacles and check safety constraints, but the system misinterprets obstacles or traffic elements leading to erroneous control decisions

Engineering Contradiction:
Improvesafety of vehicle controlVSAvoidaccuracy of obstacle identification
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent implements a feedback mechanism where the LLM continuously monitors and evaluates the vehicle's understanding of the environment by generating natural language explanations of detected obstacles and traffic elements. The system compares these explanations against the original sensor data and heuristic assessments, providing feedback that identifies misinterpretations. This feedback loop enables the system to correct erroneous obstacle identification and improve the reliability of vehicle control decisions through iterative refinement of environmental understanding.

Inventive Principle:
Principle #23Feedback

Data Source

PatentUS20250333079A1Techniques for controlling autonomous vehicles using vision-language models
Publication Date: 2025.10.30 NVIDIA CORP
  • US20250333079A1 patent drawing
  • US20250333079A1 patent drawing
  • US20250333079A1 patent drawing

AI summary

One embodiment of a method for controlling vehicles includes generating, based on sensor data, a first plan for controlling a vehicle, generating, using a trained visual language model (VLM), a final plan for controlling the vehicle based on the first plan and a second plan, and controlling the vehicle based on the final plan.