One-Shot Trajectory Prediction via Distillation for Low-Latency Driving

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing autonomous vehicle systems face challenges in accurately and efficiently predicting the behavior of objects in driving environments, particularly due to high computational latency and resource requirements of autoregressive models.

Innovation Solution

The implementation of a distillation-trained one-shot model that leverages an autoregressive model as a teacher to predict trajectories, reducing latency and resource needs by learning to replicate the teacher's outputs in a single pass.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If autoregressive models are used for trajectory prediction, then prediction accuracy is improved, but computational latency and resource requirements increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational latency
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent creates a simplified copy (student model) of the complex autoregressive model (teacher model). The student model is trained to replicate the teacher model's predictions through knowledge distillation, achieving comparable accuracy with significantly reduced computational latency and resource requirements for real-time autonomous vehicle applications

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent divides the trajectory prediction task into multiple sequential steps that the student model can process efficiently. By breaking down the complex prediction process into manageable segments, the system achieves real-time performance while maintaining accuracy through the distillation framework

Inventive Principle:
Principle #1Segmentation

2Measurement precision

If autoregressive models are used for trajectory prediction, then prediction accuracy is improved, but computational resource requirements increase

Engineering Contradiction:
Improveprediction accuracyVSAvoidcomputational resource requirements
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent creates a simplified copy (student model) of the complex autoregressive model (teacher model). The student model is trained to replicate the teacher model's predictions through knowledge distillation, achieving comparable accuracy with significantly reduced computational latency and resource requirements for real-time autonomous vehicle applications

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent modifies the architectural parameters of the model by creating a student model with fewer layers and parameters compared to the teacher model. This parameter reduction maintains prediction accuracy through knowledge distillation while decreasing computational resource requirements for deployment in autonomous vehicles

Inventive Principle:
Principle #35Parameter changes

3Loss of time

If distillation-trained one-shot models are used, then computational latency is reduced, but model training complexity increases

Engineering Contradiction:
Improvecomputational latencyVSAvoidmodel training complexity
Core Design Contradiction:
Loss of timeVSDevice complexity

Solution Approach 1:

The patent performs knowledge distillation in advance during the training phase, where the student model learns from the teacher model's predictions. This preliminary action consolidates the complex learning process into a one-shot model that can then execute predictions with minimal latency during real-time autonomous vehicle operation

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4571588A1Distillation-trained machine learning models for efficient trajectory prediction
Publication Date: 2025.06.18 WAYMO LLC
  • EP4571588A1 patent drawingFigure 1
  • EP4571588A1 patent drawingFigure 2
  • EP4571588A1 patent drawingFigure 3

AI summary

The described aspects and implementations enable training and deploying of accurate one-shot models capable of predicting trajectories of vehicles and other objects in driving environments. The disclosed techniques include, in one implementation, obtaining training data that includes a training input representative of a driving environment of a vehicle and one or more ground truth trajectories associated with a forecasted motion of the vehicle within the driving environment. The one or more ground truth trajectories are generated by a teacher model using the training input. The techniques further include training, using the training data, a student model to predict one or more trajectories of the vehicle and/or objects in the driving environment of the vehicle.