SpatialHarness: Test-Time Spatial Scaffolding for Fine Robotic Manipulation

This research introduces SpatialHarness, a test-time embodied harness that improves fine robotic manipulation by enhancing spatial observability without policy fine-tuning. The method synchronizes a simulated scene with real-world execution, identifies critical spatial relationships, and exposes them to a frozen multimodal policy. Evaluated on tasks like plug insertion and Tower of Hanoi, SpatialHarness significantly improves task success rates, highlighting the potential of existing foundation models with enhanced spatial scaffolding.