Geometry-Aware Vision-Language-Action Models for 3D Spatial Reasoning

Vision-Language-Action (VLA) models represent a new paradigm for multimodal robotic policies, integrating visual perception, language understanding, and action generation within a unified framework to control a wide range of autonomous systems, including collaborative industrial robots, autonomous vehicles, drones, and humanoid robots. Driven by the rising interest in embodied AI, researchers have developed a diverse set of architectures tailored to solve specific physical tasks, from dexterous pick-and-place manipulation and industrial assembly, to dynamic autonomous navigation, incorporating multiple sensor modalities and control strategies for vastly different hardware platforms. Consequently, robotics datasets have significantly expanded, aggregating thousands of recorded demonstrations, pairing visual observations with kinematic trajectories, to enable these models to better generalize across novel environments. Despite their versatility, current VLA policies struggle with tasks that demand precise spatial awareness and geometric reasoning. The primary bottleneck resides in the underlying Large Language Model (LLM): because it naturally operates on the syntax rules given by the words, it inherently lacks the 3D physical grounding required to accurately translate relational spatial concepts (e.g., "left of," "behind") into precise physical actions. Furthermore, while fusing additional sensor modalities could mitigate this issue, progress is severely hindered by the absence of standardized, spatially-annotated training datasets. Therefore, explicitly enhancing the geometric comprehension of these underlying architectures has become a critical challenge to develop truly context-aware robotic agents. The purpose of this research is twofold. First, we aim to systematically evaluate the capabilities of current state-of-the-art VLA systems in executing spatial reasoning tasks across both novel and familiar environmental setups. Second, we investigate novel methodologies to enhance the geometric awareness of these models without relying on expensive, third-party 3D hardware. Instead, we propose leveraging data the robot already accesses directly: its proprioceptive state alongside standard multi-view RGB observations. To address the aforementioned lack of training data, we developed a custom procedural dataset within the RLBench simulator, strictly focused on relational spatial commands. We then augmented the foundational OpenPI (π0.5)policy by explicitly injecting spatial features from multi-view geometric information from VGGT-Ω encoder and structural point cloud embeddings from a Utonia Point Transformer. Our empirical evaluation reveals that the strategy for multimodal integration is critical. While the naive concatenation of high-dimensional spatial features causes a catastrophic disruption of the model's pre-trained latent space, properly managing modality alignment through fully fused integration layers successfully stabilizes the architecture. Ultimately, the fused multimodal policy outperforms the unmodified baseline, demonstrating stability in resolving complex object centric, robot centric and camera centric reference frames.