DECODE: Domain-Aware Continual Domain Expansion for Motion Prediction.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Autonomous vehicles require robust motion prediction for safe navigation.
• Continual learning is crucial for adapting models to new driving scenarios without full retraining.

Purpose of the Study:

• To develop a continual learning framework (DECODE) for adaptive motion prediction in autonomous driving.
• To balance model specialization and generalization for improved performance and robustness.

Main Methods:

• DECODE incrementally expands specialized models from a pre-trained generalized model.
• Utilizes a hypernetwork for parameter generation and normalizing flow for real-time domain inference.
• Fuses outputs from specialized and generalized models using Bayesian uncertainty estimation.

Main Results:

• Achieved a low forgetting rate of 0.044 and an average minADE of 0.584 m.
• Outperformed prior methods in motion prediction across diverse driving domains.
• Demonstrated broad applicability by extending to image classification tasks.

Conclusions:

• DECODE effectively balances specialization and generalization for continual learning in motion prediction.
• The framework offers improved performance and robustness in autonomous driving systems.
• DECODE shows potential for various continual learning applications beyond motion prediction.