Multimodal Token Networks for Autonomous Driving Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional systems for vehicle data processing and prediction rely on separate neural networks for different tasks, increasing system complexity and training costs, and often lack the ability to efficiently utilize complementary information across data modalities.

Innovation Solution

A multi-modal token processing neural network, such as a pre-trained multi-modal language model, is adapted to perform various vehicle data processing and prediction tasks by fine-tuning it to include tokens representing input and output data for vehicle systems, enabling efficient training and more accurate predictions across multiple data modalities.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If separate neural networks are used for different vehicle data processing tasks, then task-specific performance is improved, but system complexity and training costs increase

Engineering Contradiction:
Improvetask-specific performanceVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges multiple separate neural networks into a single unified neural network that processes various vehicle data types (images, point clouds, audio, text) through shared processing layers. This consolidation maintains task-specific performance while reducing system complexity by eliminating redundant components and streamlining the architecture.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified neural network is designed with multi-functionality to handle diverse vehicle data processing tasks including object detection, scene understanding, and navigation decisions. The network uses a universal token-based representation that can encode multiple data modalities, allowing a single system to perform previously separate functions.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Reliability

If separate neural networks are used for different vehicle data processing tasks, then task specialization is improved, but training costs increase

Engineering Contradiction:
Improvetask specializationVSAvoidtraining costs
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

By combining multiple specialized networks into one unified network, the patent enables joint training on multiple data modalities simultaneously. This approach reduces total training costs by sharing computational resources and avoiding redundant training processes, while maintaining specialization through the network's ability to process different token types for different tasks.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent employs parameter-efficient fine-tuning techniques where only specific portions of the unified network are retrained for particular tasks rather than training entire separate networks. This selective parameter updating significantly reduces training computational costs while preserving task-specific performance through targeted adaptation.

Inventive Principle:
Principle #35Parameter changes

3Ease of operation

If conventional systems process multiple data modalities separately, then processing simplicity is maintained, but ability to utilize complementary information decreases

Engineering Contradiction:
Improveprocessing simplicityVSAvoidcomplementary information utilization
Core Design Contradiction:
Ease of operationVSLoss of information

Solution Approach 1:

The patent introduces tokens as an intermediary representation that bridges multiple data modalities. Different data types (images, point clouds, audio, text) are converted into a unified token sequence that the neural network can process seamlessly. This intermediary approach maintains processing simplicity by providing a consistent interface while enabling full utilization of complementary information across modalities through the shared processing architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4650818A1Adapting foundation models for autonomous driving
Publication Date: 2025.11.19 WAYMO LLC
  • EP4650818A1 patent drawingFigure 1A
  • EP4650818A1 patent drawingFigure 1B
  • EP4650818A1 patent drawingFigure 2

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for processing and predicting multi-modal data characterizing a driving environment. In one aspect, a method comprises: receiving input data that characterizes a driving environment, wherein the input data comprises a respective input for each of a plurality of data modalities characterizing the driving environment; generating an input multimodal token sequence of input tokens that represents the inputs for each of the plurality of data modalities; and processing the input multimodal token sequence using a token processing neural network to generate an output token sequence representing a prediction about the driving environment.