Agent Response Time Modeling Using Surprise in Traffic Scenarios

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Autonomous vehicles face challenges in predicting the response times of other agents in traffic scenarios due to the ambiguity of stimulus onset and end times, especially in naturally occurring conflicts where human behavior is unpredictable and situation-dependent.

Innovation Solution

A system using surprise-based evidence accumulation models to predict response times by measuring the accumulation of surprising evidence over time, incorporating generative models and heuristic approaches to handle various traffic scenarios, and leveraging machine learning to account for factors like visibility, weather, and driver state.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional detection methods are used to identify traffic conflicts, then the system can detect other agents and basic traffic situations, but it cannot accurately determine the onset and end times of stimuli for response time measurement

Engineering Contradiction:
Improveresponse time measurement accuracyVSAvoidstimulus onset and end time identification
Core Design Contradiction:
Measurement precisionVSDifficulty of detecting and measuring

Solution Approach 1:

The patent applies dynamics by transitioning from static detection thresholds to dynamic, adaptive stimulus identification. The system continuously updates the expected behavior model based on accumulated evidence, allowing the stimulus onset and end times to be dynamically determined rather than fixed by predetermined thresholds. This enables accurate response time measurement by adapting to the evolving nature of traffic conflicts.

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The patent implements preliminary action by pre-establishing a generative model of expected agent behavior before traffic conflicts occur. This model accumulates evidence during normal operation and is ready to immediately identify stimulus onset when deviations occur, rather than waiting for conflict detection algorithms to reactively analyze the situation after it has developed.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If the system uses detailed generative models to predict agent behavior, then prediction accuracy improves, but computational complexity increases

Engineering Contradiction:
Improveagent behavior prediction accuracyVSAvoidcomputational model complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies partial action by implementing a two-stage evidence accumulation process that only fully activates detailed generative modeling when necessary. During normal operation, the system uses lighter-weight expected behavior modeling, and only transitions to full generative model computation when stimulus onset is detected or suspicion of conflict arises, thus balancing accuracy with computational efficiency.

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent implements nesting by embedding multiple levels of modeling complexity within a unified framework. The generative model of expected behavior contains nested sub-models for different aspects of agent behavior (trajectory prediction, intent recognition, response time estimation), allowing the system to activate only the necessary levels of complexity for each specific situation rather than always using the full computational burden.

Inventive Principle:
Principle #7Nested doll (Nesting)

3Measurement precision

If the system accumulates surprise evidence continuously, then response time prediction becomes more accurate, but the processing time and computational load increase

Engineering Contradiction:
Improveresponse time prediction accuracyVSAvoidcomputational processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies skipping by implementing selective evidence accumulation that bypasses continuous full-model computation. The system skips detailed generative model evaluation during periods when behavior conforms to expectations, and only rushes through the complete computational process when surprise evidence thresholds are approached or exceeded, thus reducing average processing time while maintaining prediction accuracy.

Inventive Principle:
Principle #21Skipping (Rushing through)

Solution Approach 2:

The patent implements periodic action by evaluating surprise evidence at discrete time intervals rather than continuously. The system accumulates evidence periodically and compares it against thresholds at these intervals, allowing computational resources to be dedicated to other tasks between evaluation points while still maintaining accurate response time predictions through the accumulated evidence framework.

Inventive Principle:
Principle #19Periodic action

Data Source

PatentEP4322126B1Computing agent response times in traffic scenarios
Publication Date: 2025.10.01 WAYMO LLC
  • EP4322126B1 patent drawingFigure 1
  • EP4322126B1 patent drawingFigure 2
  • EP4322126B1 patent drawingFigure 3A

AI summary

This specification describes techniques for a system configured to measure and model the response times of agents in a driving environment. There is provided a computer-implemented method comprising: receiving a request to compute a predicted response time to the onset of a stimulus for an agent in a scenario in a traffic environment; continually updating, at each time step of a plurality of time steps, an accumulated measure of surprise for the agent due to the movements of another entity in the traffic environment, including performing, at each time step, operations comprising: obtaining, using a generative model, a distribution of previously predicted trajectories at a previous time step for the other entity in the environment, computing, using an updated state of the other entity, a measure of surprise for the time step from the perspective of the agent including comparing the updated state of the other entity to the distribution of previously predicted trajectories at the previous time step for the other entity, and updating the accumulated measure of surprise for the time step using the computed measure of surprise for the agent; determining that an accumulated measure of surprise crosses a threshold at a particular point in time; and in response, generating a predicted response time for the agent based on the particular point in time that the accumulated measure of surprise crosses the threshold.