Accepted to ECCV 2026. Paper
E-VLA augments vision-language-action models with event-camera signals to support robotic manipulation in dark and motion-blurred scenes. It uses synchronized RGB, events, and actions collected with a DAVIS346 camera and an SO-100 manipulator.
compressed_demo.mp4
E-VLA was evaluated under extremely low light and severe motion blur, where frame-only perception can become unreliable.
The implementation is organized into two branches:
| Branch | Contents | Guide |
|---|---|---|
event_adapter |
Synchronized DAVIS capture, event-aware datasets, event-adapter training, and deployment | Event-adapter guide |
image_only |
Image-only SmolVLA baseline, multitask training, and image video profiles | Image-only guide |
The synchronized RGB-event-action dataset is available at JJayzee/E-VLA-MultiLight. Dataset layout and feature contracts are documented in each branch's guide.
- Robust perception: Event streams retain motion and structural cues when RGB frames are degraded by low light or blur.
- Efficient integration: Event representations and lightweight fusion strategies extend a pretrained SmolVLA policy.
- Synchronized data: A DAVIS346 camera and SO-100 manipulator provide RGB, event, timestamp, action, and task data across illumination conditions.
- Experimental evidence: The paper reports manipulation results for image-only and event-augmented policies under adverse visual conditions.
@inproceedings{zhai2026evla,
author = {Zhai, Jiajun and Shi, Hao and Guo, Shangwei and Yang, Kailun and Wang, Kaiwei},
title = {E-VLA: Event-Augmented Vision-Language-Action Model for Dark and Blurred Scenes},
booktitle = {European Conference on Computer Vision (ECCV)},
year = {2026}
}E-VLA builds on SmolVLA and LeRobot. We thank their maintainers for the base model, robotics framework, and SO-10x ecosystem.