ETA: Efficiency through Thinking Ahead, A Dual Approach to Self-Driving with Large Models
Published in ICCV 2025, 2025
Recommended citation: S. Hamdan et al., "ETA: Efficiency through Thinking Ahead, A Dual Approach to Self-Driving with Large Models", ICCV 2025 https://arxiv.org/abs/2506.07725
This paper proposes a dual-system architecture that enables the use of slow, heavy vision-language model planners in tandem with lightweight planners for autonomous driving. ETA achieves state-of-the-art performance in the Bench2Drive benchmark with a 6x speedup compared to previous approaches.
The work was conducted during my research internship at OpenDriveLab, Shanghai AI Lab, under the supervision of Asst. Prof. Hongyang Li.
Access
Recommended citation:
- S. Hamdan et al., “ETA: Efficiency through Thinking Ahead, A Dual Approach to Self-Driving with Large Models”, ICCV 2025