| --- |
| title: OraRL |
| emoji: 🎬 |
| colorFrom: blue |
| colorTo: purple |
| sdk: static |
| pinned: false |
| --- |
| |
| # OraRL |
|
|
| [[🏠 Homepage](https://orarl.github.io/)] [[📖 Arxiv Paper](https://arxiv.org/abs/2608.20492)] [[🤗 Models & Data](https://huggingface.co/OraRL)] [[💻 Code](https://github.com/HVision-NKU/OraRL)] |
|
|
| We build **OraRL**, a reinforcement-learning framework for **unified video multimodal large language models**. |
|
|
| Our goal is to make broad video RL both **reliable** and **efficient** by treating task annotations as positive rollouts while retaining policy samples for the on-policy baseline. The resulting **Video-ORA** family handles seven task families without chain-of-thought decoding. |
|
|
| <video controls playsinline preload="metadata" width="100%" src="https://orarl.github.io/assets/orarl-teaser.mp4"></video> |
|
|
| <p align="center"><strong>▶ OraRL project overview (1:38)</strong></p> |
|
|
| ## What we release |
|
|
| - **Video-ORA-9B**: a unified video MLLM for temporal and spatial grounding, segmentation, tracking, video QA, and spatial intelligence. |
| - **OraRL-Data**: canonical evaluation annotations and raw media for reproducible comparison across the seven task families. |
| - **OraRL resources**: project documentation, model cards, inference examples, and evaluation guidance. |
|
|
| ## Research interests |
|
|
| - Unified video understanding |
| - Multimodal reinforcement learning |
| - Temporal & spatial grounding |
| - Video segmentation, tracking & spatial intelligence |
|
|
