|
Full title—Eco-VLA: An Eco-driving Framework for Autonomous Driving Based on Vision-Language-Action Model with Reinforcement Fine-Tuning
The flourishing development in Vision-Language Action (VLA) models has showcased significant potential for end-to-end autonomous driving through robust data-driven and reasoning capabilities. Current autonomous driving studies lever aging VLA primarily emphasize safety, regulatory compliance, efficiency, and comfort across diverse driving scenarios.
In this study, we introduce Eco-VLA, an innovative eco-driving frame work based on the VLA model with reinforcement fine-tuning to enhance energy efficiency in autonomous vehicles. Eco-VLA conducts semantic reasoning and energy-efficient trajectory planning directly from raw visual inputs and language instructions
To enhance eco-driving performance and efficiency, we develop a refined reinforcement learning fine-tuning technique, termed Discounted Group Relative Policy Optimization (D-GRPO), to generate better long-term benefits and improve the energy-saving effects. Furthermore, Eco-VLA incorporates an energy-tailored reward function in D-GRPO specifically designed for eco-driving, ensuring energy savings while maintaining safety, comfort, and driving efficiency.
Extensive experiments conducted on the CoVLA benchmark demonstrate that Eco-VLA achieves up to over 8% energy-saving effects compared with state-of-the-art VLA models, while maintaining competitive performance in safety, comfort, and efficiency metrics. Additional cross-dataset evaluations on nuPlan and Waymo further validate the generalization capability of the proposed framework.
The results confirm that Eco-VLA provides an effective and scalable solution for energy-efficient autonomous driving.
Full Article: IEEE Transactions on Vehicular Technology, Early Access
| 
