This research introduces SmartPretrain, a model-agnostic and dataset-agnostic approach to improve motion prediction for autonomous vehicles despite limited training data.

• Uses self-supervised learning techniques inspired by NLP and computer vision breakthroughs
• Develops a generalizable framework that works across different models and datasets
• Enhances prediction of complex interactions and road geometries with limited training data
• Directly addresses critical safety challenges in autonomous vehicle deployment

Improving motion prediction capabilities is essential for autonomous vehicle security, allowing systems to better anticipate potential hazards and reduce collision risks in dynamic, mixed human-robot environments.

SmartPretrain: Model-Agnostic and Dataset-Agnostic Representation Learning for Motion Prediction