MAVEN-T: A Reinforcement Learning-Based Knowledge Distillation Framework for Multi-Agent Trajectory Prediction

MAVEN-T is a reinforcement learning-based knowledge distillation framework for multi-agent trajectory prediction. It breaks through the imitation ceiling of traditional distillation through complementary architecture co-design, multi-granularity progressive distillation, and reinforcement learning enhancement. This framework achieves 6.2x parameter compression and 3.7x inference speedup while maintaining SOTA prediction accuracy, even surpassing the teacher model in robustness, providing a new path for efficient model deployment in autonomous driving scenarios.

Autonomous driving trajectory prediction faces three major challenges: complexity (multi-agent interaction, multi-level information understanding), real-time performance (millisecond-level inference), and uncertainty (randomness of human behavior). Traditional knowledge distillation is effective for simple tasks, but in multi-agent scenarios, it has limitations such as behavior cloning (only learning surface behaviors), distribution shift (differences between training and deployment environments), and insufficient interaction modeling, forming an 'imitation ceiling'.

MAVEN-T adopts a complementary architecture design: the teacher network uses a hybrid attention mechanism to maximize representation capability, while the student network is optimized for lightweight deployment. Knowledge transfer is achieved through multi-granularity progressive distillation: trajectory-level (output matching), intent-level (middle layer alignment), and interaction-level (attention weight transfer). Combined with adaptive curriculum learning to dynamically adjust training difficulty, it ensures that the student understands the decision-making logic rather than just imitating trajectories.

MAVEN-T introduces a reinforcement learning module that allows the student model to verify and optimize knowledge through simulated environment interaction: accurate predictions get positive rewards, collisions or violations get negative rewards, and conservative/aggressive predictions receive moderate penalties. This trial-and-error learning enables the student to discover robust strategies ignored by the teacher, moving beyond simple replication to break the imitation ceiling, even surpassing the teacher model in decision-making robustness.

Evaluated on NGSIM and highD datasets: MAVEN-T achieves 6.2x parameter compression (the student only needs 16% of the teacher's parameters), 3.7x inference speedup; maintains SOTA prediction accuracy; RL enhancement allows the student to surpass the teacher model in robustness metrics (extreme scenarios, out-of-distribution tests), verifying the effectiveness of the framework's efficiency-accuracy trade-off.

Theoretical contribution: First to prove that RL enhancement can break the distillation imitation ceiling for complex decision-making tasks; Methodological contribution: Complementary architecture, multi-granularity distillation, and adaptive curriculum learning provide a paradigm for efficient model development; Practical contribution: 6.2x compression and 3.7x speedup enable complex reasoning models to be deployed on autonomous driving edge devices, promoting the implementation of technology in safety-critical fields.

Limitations: Simulation-reality gap affects RL strategy transfer, reward function design relies on manual work, and RL training has high computational costs. Future directions: Introduce world models to reduce environmental interaction, explore offline RL to lower training costs, and extend the framework to other complex decision-making tasks.