Modular Trajectory Generation for Interpretable Autonomous Driving

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing autonomous vehicle systems face challenges in making safe and effective decisions due to the complexity of understanding the environment and predicting future scenarios, with conventional programmed and learned approaches lacking safety assurances, interpretability, and requiring extensive data for training.

Innovation Solution

A data-driven, modular decision-making system that combines programmed and learned processes, utilizing a hybrid architecture with localized learning modules tuned to specific contexts and actions, ensuring safety and naturalistic trajectories through hypertuning loss functions and contextual awareness.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional programmed motion planners are used, then decision-making rules can be hand-tuned and optimized by experts, but the system produces unnatural decisions and motions and requires an exhaustive list of scenarios to program

Engineering Contradiction:
Improvesafety assurancesVSAvoidnaturalistic decisions
Core Design Contradiction:
ReliabilityVSEase of operation

Solution Approach 1:

The motion planning problem is segmented into multiple sub-problems handled by separate learning modules, each responsible for specific aspects of decision-making. This allows the system to produce naturalistic decisions for different driving scenarios while maintaining overall safety through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Different learning modules are trained on specific local scenarios and conditions, allowing each module to develop specialized knowledge for particular driving situations. This local specialization enables naturalistic decisions while the collective system provides comprehensive safety coverage.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If holistic end-to-end learned motion planners are used, then the system can analyze large amounts of human driving data, but safety assurances are lost and sample sparsity remains an issue

Engineering Contradiction:
Improvelearning capabilityVSAvoidsafety assurances
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The end-to-end learning system is divided into multiple specialized learning modules, each trained on specific aspects of driving behavior. This segmentation allows the system to learn from diverse data while maintaining interpretability and safety through modular architecture, where each module's decisions can be individually validated.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Intermediate representations and safety validation layers are introduced between the learned modules and final execution. These intermediaries provide safety assurances by validating decisions before execution, while still allowing the system to leverage large amounts of training data for adaptability.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Adaptability or versatility

If single monolithic networks are used to address entire driving policy, then the system can learn from data, but interpretability and explainability are lost

Engineering Contradiction:
Improvelearning from dataVSAvoidinterpretability
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The monolithic network is segmented into multiple specialized learning modules, each responsible for specific driving decisions. This segmentation enables interpretability by allowing individual module decisions to be analyzed and explained, while the system as a whole retains the ability to learn from extensive driving data.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Each learning module develops specialized local knowledge for specific driving scenarios, making the system's decisions more interpretable. The local expertise of each module can be independently analyzed and explained, unlike a monolithic network where interpretability is lost.

Inventive Principle:
Principle #3Local quality

4Productivity

If conventional learning algorithms treat motion planning in end-to-end fashion, then the system can process data efficiently, but safety assurances regarding trajectories cannot be provided

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidtrajectory safety
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

Safety validation intermediaries are introduced into the end-to-end processing pipeline. These intermediaries validate trajectory safety without significantly impacting processing efficiency, as they operate on the outputs of existing learning modules rather than requiring complete system redesign.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The processing pipeline is segmented into learning modules and validation layers, allowing efficient data processing through specialized modules while ensuring trajectory safety through dedicated validation components. This segmentation maintains productivity while adding safety assurances.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4162339B1Method and system for data-driven and modular decision making and trajectory generation of an autonomous agent
Publication Date: 2025.07.02 GATIK AI INC
  • EP4162339B1 patent drawingFigure 1
  • EP4162339B1 patent drawingFigure 2
  • EP4162339B1 patent drawingFigure 3A

AI summary

A system for data-driven, modular decision making and trajectory generation includes a computing system. A method for data-driven, modular decision making and trajectory generation includes: receiving a set of inputs; selecting a learning module such as a deep decision network and/or a deep trajectory network from a set of learning modules; producing an output based on the learning module; repeating any or all of the above processes; and/ or any other suitable processes. Additionally or alternatively, the method can include training any or all of the learning modules; validating one or more outputs; and/or any other suitable processes and/or combination of processes.