New Concept Expert Module Gives VLA Robots 3D Spatial Awareness

Current Vision-Language-Action (VLA) models rely heavily on 2D inputs, limiting their spatial awareness for complex, high-precision tasks. Researchers Mingyang Sun, Jiude Wei, Xiujian Liang, Qichen He, Donglin Wang, Cewu Lu, and Jianhua Sun propose a novel 'Concept Expert' module that equips VLA models with 3D structural and commonsense knowledge. This module constructs executable Analytic Concepts—programmatic blueprints representing objects with explicit kinematic parameters—bridging the gap between 2D vision and physical 3D manipulation.

Operating in two phases, the Concept Expert first uses vision foundation models to estimate initial kinematic and structural parameters from 3D data. During manipulation, the VLA model dynamically tracks these concept parameters against real-time observations for persistent accuracy. These Analytic Concepts provide dense, programmatic rewards and precise spatial guidance for fine-tuning VLA models, enabling physically grounded interaction behaviors without sacrificing end-to-end learning flexibility. Experimental results show consistent improvements in success rate and learning efficiency across supervised and reinforcement learning settings.