Robots need spatial understanding to act, and vision-language-action (VLA) models are often given 3D awareness by aligning their representations with pre-trained 3D foundation models. The authors of this paper observe that such direct geometric alignment effectively acts as low-rank regularization, leading to severe rank collapse in the VLA representations and reducing their expressivity.
Rather than enforcing point-wise equality between representations, the researchers propose aligning pair-wise relationships. This relational approach injects 3D spatial awareness without compromising the model's original capabilities. The method was tested on public benchmarks, a newly introduced spatial reasoning benchmark, and real-world manipulation scenarios.
The results indicate that the approach successfully incorporates 3D spatial representations into VLA reasoning, achieving robust performance and enhanced spatial awareness across diverse manipulation tasks. The paper is available on arXiv.