SpatialClaw: Rethinking the Action Interface for Agentic Spatial Reasoning

arXiv·medium signal

SpatialClaw targets a core weakness in vision-language and embodied agents — reliably determining where objects are, how they relate, and how they move in 3D — by redesigning the action interface agents use for spatial reasoning rather than just scaling the model. Relevant to practitioners building robotics or embodied/computer-use agents where spatial grounding is the bottleneck.