The authors propose LIME, a vision-language model that generates language-conditioned camera motion by predicting relative target poses from RGB observations and natural-language intents.
- The system mines multi-intention supervision from egocentric video, pairing plausible intents with observation-gain descriptions and SE(3) target poses.
- LIME combines an auto-regressive observation-gain output with a continuous flow-matching pose head to jointly predict what the next view should reveal.
- This design allows the model to represent multi-hypothesis target views for tasks such as inspecting objects or revealing occluded regions.
The approach demonstrates that LIME can learn to actively choose camera poses from passive human video, effectively turning ordinary egocentric recordings into supervision for intent-aware active perception.