PanoFuse: Panorama-Enhanced Vision-Language-Action Learning with Decoupled Semantic-Geometric Routing

Vision-Language-Action (VLA) policies have shown promising performance in language-conditioned robotic manipulation, but conventional perspective cameras provide only limited fields of view and may miss global scene context under occlusion, distractors, and unseen environments. We propose PanoFuse, a panorama-enhanced VLA framework that complements local wrist-camera observations with global panoramic perception. A dedicated panoramic branch leverages a pretrained panoramic foundation model to extract complementary semantic and geometric representations. Rather than directly mixing these heterogeneous features, Decoupled Semantic-Geometric Routing (DSGR) maintains them as separate context streams and selectively routes them to downstream state and action representations through structured block-wise attention. We further develop a synchronized teleoperation pipeline and collect a real-world manipulation dataset containing panoramic RGB observations, wrist-view images, language instructions, robot states, and actions. Across seven evaluation settings, PanoFuse reaches an average success rate of 52.9% and shows consistent improvements under novel-object, unseen-background, and distractor-rich settings.