Event
IG-VLA: A VLA Framework That Internalizes Future Scene Imagination into Compact Tokens
1 reports · 1 sources
Shenglan Li and co-authors propose IG-VLA, a vision-language-action (VLA) reasoning framework that enables the model to imagine task-relevant future scene evolution in visual representation space, and uses Scene Gist Memory to internalize the reasoning results into compact Scene Gist Tokens, so that the inference stage can skip explicit future imagination. The authors state that its latent spatiotemporal reasoning directly predicts future scene evolution in visual representation space to guide action prediction, without generating pixel-level video. This is currently a method proposed at the paper level; the report does not provide performance data for this framework on specific tasks.
Report timeline
- IG-VLA: Efficient VLA Reasoning via Internalized Spatiotemporal Imagination
The authors propose the IG-VLA framework, which lets a vision-language-action model imagine task-relevant future scene evolution in visual representation space and internalize the reasoning results into a compact Scene Gist Token via Scene Gist Memory, thereby skipping explicit future imagination at inference time.
arXiv cs.RO ·