General-purpose multimodal models like GPT-6 Astra and Claude Opus 5.5 show broad skills in embodied manipulation, according to a new arXiv preprint. The paper's analysis suggests that this generality comes with strong spatial perception, a crucial trait for physical AI.
The preprint introduces ACE-Brain-0.6, a contribution aimed at scaling spatial grounding for physical foundation models. Spatial grounding connects language and visual inputs to physical locations and actions, which is essential for robots and other embodied systems.
The authors see a link between general model capabilities and spatial reasoning. If their analysis is correct, it may guide future work on training and evaluating physical foundation models. The paper is a preprint and has not yet undergone peer review.