MTR++: Multi-Agent Motion Prediction With Symmetric Scene Modeling and Guided Intention Querying

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving systems require accurate motion prediction for safe navigation in complex environments.
• Predicting diverse traffic participant behaviors and contextual interactions remains a significant challenge.

Purpose of the Study:

• To propose the Motion TRansformer (MTR) framework for efficient and accurate multimodal motion prediction.
• To develop MTR++ for simultaneous multi-agent motion prediction, enhancing scene understanding and interaction modeling.

Main Methods:

• Developed MTR using a transformer encoder-decoder with learnable intention queries for trajectory prediction.
• Implemented global intention localization and local movement refinement processes within the MTR framework.
• Introduced MTR++ with symmetric context modeling and mutually-guided intention querying for multi-agent prediction.

Main Results:

• MTR achieved state-of-the-art performance on competitive motion prediction benchmarks.
• MTR++ demonstrated superior performance and efficiency in predicting multimodal future trajectories for multiple agents.
• The proposed frameworks effectively reduce reliance on dense goal candidates and improve prediction accuracy.

Conclusions:

• The MTR and MTR++ frameworks offer significant advancements in autonomous driving motion prediction.
• These models provide accurate, efficient, and scene-compliant future trajectory predictions for single and multiple agents.
• The approach enhances the decision-making capabilities of autonomous driving systems.