README.md

metadata

license: apache-2.0
pipeline_tag: robotics

EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies

EventVLA is an end-to-end vision-language-action framework designed for long-horizon robotic manipulation tasks. It introduces an event-driven visual evidence memory mechanism to address memory bottlenecks when task-relevant cues become occluded or unobservable over time.

• Project Page: ganlin-yang.github.io/EventVLA.github.io
• Repository: GitHub - InternRobotics/EventVLA
• Paper: EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies
• Dataset: RoboTwin-MeM on Hugging Face

Model Description

EventVLA addresses the memory constraints of standard Vision-Language-Action (VLA) policies by employing a dynamic Keyframe Evidence Memory (KEM) module alongside foundational visual anchors. KEM predicts future keyframe probabilities from latent embeddings to store sparse, task-critical visual events, preserving visual evidence before it is lost or obscured.

This repository contains the trained weights evaluated on two benchmarks:

• RoboTwin-MeM: Checkpoints for evaluation on the eight RoboTwin-MeM tasks.
• RMBench: Checkpoints for evaluation on the RMBench benchmark.

Each release directory contains:

• pytorch_model.pt
• config.yaml
• dataset_statistics.json
• summary.jsonl

Please refer to the official GitHub repository for instructions on installation, training, and evaluation.

Citation

If you find this work useful, please cite:

@article{yang2026eventvla,
  title={EventVLA: Event-Driven Visual Evidence Memory for Long-Horizon Vision-Language-Action Policies},
  author={Yang, Ganlin and Tu, Zhangzheng and Yang, Yuqiang and Mao, Sitong and Dong, Junyi and Chen, Tianxing and Peng, Jiaqi and Xiong, Jing and Cao, Jiafei and Dai, Jifeng and others},
  journal={arXiv preprint arXiv:2606.20092},
  year={2026}
}