E-VLA boosts robot manipulation in dark/blurred scenes from 0% to 90%

Robotic vision-language-action (VLA) models have excelled at open-ended manipulation, but they collapse under sensing degradations like extreme low light, motion blur, or black clipping. A new paper accepted to ECCV 2026 introduces E-VLA (Event-Augmented Vision-Language-Action Model), a framework that directly integrates event streams—asynchronous pixel-level brightness changes—into VLA pipelines without reconstructing images. Built by Jiajun Zhai, Hao Shi, and collaborators from Zhejiang University, the system uses a DAVIS346 event camera and a custom teleoperation platform to collect a synchronized RGB-event-action manipulation dataset across diverse tasks and illumination levels.

The key innovation is treating event data as an additional perceptual modality rather than a reconstruction source. E-VLA explores lightweight, pretrained-compatible fusion strategies—from parameter-free overlays to learned event adapters. Results are striking: at 20 lux (roughly dim indoor lighting), a standard image-only VLA achieves 0% success on a Pick-Place task. Simply overlaying accumulated event maps onto RGB frames boosts success to 60%, and a dedicated event adapter reaches 90%. Under severe motion blur simulated with 1000 ms exposure, the image-only baseline again scores 0%, while the overlay fusion reaches 20–25%, and a more complex sorting task improves from 5% to 32.5%. These results provide systematic evidence that event-driven perception can be effectively fused into VLA models, pointing toward robust embodied intelligence beyond conventional frame-based imaging. Code and the new dataset will be released upon publication.