Direct Action-Head Injection of A Grounded 3D Point Unlocks Spatial and Task Generalization
The work proposes injecting a 3D spatial embedding directly into a robot's action head to improve generalization in Vision-Language-Action (VLA) models. By lifting a 2D grounding signal into 3D space and feeding this geometric relationship directly into the policy, the model significantly enhances its ability to handle variations in object positions and language instructions during testing.