FocusVLA routes multi-view inputs using predicted future intent
Original title: FocusVLA: Future-Intent Guided Dynamic View Routing for Vision-Language-Action Models
On RoboTwin 2.0 it raises average success 10.10 points over π0.5, and real-world success from 71.7% to 85.0%.
The paper says existing vision-language-action policies keep relying on the wrist view and do not dynamically shift global-local view use across manipulation stages. FocusVLA predicts future interaction intent from historical actions and uses that signal to route views during action generation. Its intent and routing modules can be added to existing VLA frameworks without changing their core architectures. Reported gains are 10.10 percentage points on RoboTwin 2.0 and a rise from 71.7% to 85.0% in real-world evaluation versus π0.5.
- FocusVLA predicts future interaction intent from historical actions.
- RoboTwin 2.0 average success rose 10.10 points over the strongest baseline, π0.5.
- Real-world success rose from 71.7% to 85.0%, a 13.3% gain over π0.5.
- Routing modules can plug into existing VLA frameworks without core architecture changes.
- Products
- RoboTwin 2.0
From the original article
1 min read at huggingface.co
Vision-language-action (VLA) models have become an important paradigm for robotic manipulation by unifying visual perception, language understanding, and action generation. Since reliable manipulation requires sufficient observation of objects, end-effectors, and their spatial relationships, multi-view inputs are widely used to provide complementary spatial information.
By Huixin Xu, Ziya Yu, Peng Li · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.