LEXIS: LatEnt ProXimal Interaction Signatures for 3D HOI from an Image

The key technical bet LEXIS makes is that interaction geometry is better captured by a dense, continuous field across surfaces than by discrete contact points, which tend to be sparse and brittle when inferred from a single 2D image. This framing shifts the problem from 'where do surfaces touch' to 'how close are all surface regions to each other,' which is a more tractable signal when depth information is absent.

Recent coverage on Modelwire has concentrated heavily on LLM capability races and agentic coding tools, so LEXIS sits largely disconnected from that activity. It belongs instead to a quieter but consequential thread in computer vision: giving machines a structured understanding of how bodies relate to objects in physical space. That capability feeds directly into robotics, AR/VR scene reconstruction, and any system that needs to reason about physical manipulation. The closest thematic neighbor in the archive is the shortest-path generalization paper from arXiv (April 16), which also probed how well learned representations hold up when the problem geometry changes, though the domains are quite different.

Watch whether LEXIS benchmarks extend to video input or multi-view settings within the next year. If the proximity field representation generalizes beyond single-image input, that would signal the approach is robust rather than a solution fitted tightly to a constrained problem setup.