SpatialActor: Exploring Disentangled Spatial Representations for Robust Robotic Manipulation

The introduction of SpatialActor marks a significant advancement in robotic manipulation, focusing on the critical need for precise spatial understanding. Traditional point-based and image-based methods often struggle with depth noise and entangled semantics, which can hinder performance. By explicitly decoupling semantics and geometry, SpatialActor enhances the robustness of robotic systems, achieving state-of-the-art performance with an 87.4% success rate on RLBench. This framework not only improves performance under noisy conditions by 13.9% to 19.4% but also enhances few-shot generalization to new tasks. Evaluated across more than 50 tasks, SpatialActor demonstrates its potential to revolutionize how robots interact with their environments, paving the way for more reliable and effective robotic applications.