Training Strategy Outperforms Model Complexity in Autonomous Driving
ORIGINAL / DriveMLP: When Training Strategy Outperforms Model Complexity in Autonomous Driving
The study finds that in closed-loop autonomous driving planning, training strategy rather than model complexity accounts for substantial performance gains. The authors propose the EGRL strategy and a lightweight MLP model DriveMLP, which outperforms the previous best IL planner on the nuPlan benchmark with faster inference.
01 ABSTRACT
Closed-loop planning in autonomous driving faces challenges such as distribution shift and compounding errors. The authors propose Expert-Guided RL (EGRL), which uses a trajectory-distance reward derived from expert data and RL purely as an optimization tool. Experiments show training procedure contributes significantly to performance, and the lightweight MLP model DriveMLP trained with EGRL achieves 86.51 on the nuPlan benchmark, surpassing the previous best IL planner.
02 KEY FINDINGS
- EGRL uses a trajectory-distance reward derived from expert data and treats RL purely as an optimization tool.
- The lightweight MLP model DriveMLP scores 86.51 on the nuPlan Val14 Reactive benchmark, a +3.2 point improvement over the previous best IL planner.
- DriveMLP achieves over 15x faster inference than the previous best IL planner.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG