S$^2$-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation
For robotic manipulation, this work provides a method to improve long-horizon task performance by adaptive feature fusion, showing strong results against larger models.
S^2-VLA introduces a State-Space Guided Adaptive Attention mechanism to address cumulative error in long-horizon robotic manipulation, achieving state-of-the-art performance on LIBERO and SimplerEnv benchmarks with a 2B parameter model, outperforming larger 7B-scale models.
Vision-Language-Action (VLA) models have demonstrated strong capabilities in robotic manipulation, but their performance degrades significantly in long-horizon tasks due to cumulative error propagation. This limitation largely arises from static feature fusion mechanisms that rely on fixed weights to combine visual, language, and action representations, preventing the model from adapting to different phases of task execution. To address this limitation, we propose S$^2$-VLA, a framework that introduces a State-Space Guided Adaptive Attention (SSGAA) mechanism. SSGAA maintains a belief state that tracks task progression and generates dynamic gating weights to adaptively fuse information from three complementary sources visual features for spatial perception, task intents for high-level task planning, and temporal action sequences for execution consistency. This adaptive fusion allows the model to shift its focus throughout task execution, aligning with the evolving requirements of different task stages. Despite its compact 2B parameter size, S$^2$-VLA consistently outperforms larger 7B-scale models and achieves state-of-the-art performance on long-horizon manipulation benchmarks, including LIBERO and SimplerEnv. highlighting the importance of adaptive feature fusion for long-horizon robotic manipulation.