S$^2$-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation

2026-06-26

AI Summary

Existing vision-language-action (VLA) models suffer significant performance degradation in long-horizon manipulation tasks due to error accumulation stemming from static feature fusion mechanisms. To address this limitation, this work proposes the S²-VLA framework, which introduces a novel state-space-guided adaptive attention mechanism that dynamically integrates visual observations, language intent, and action sequences while adaptively adjusting multimodal attention weights according to task phases. This approach overcomes the constraints of conventional static fusion strategies, enabling efficient belief-state tracking and cross-modal coordination. Evaluated on long-horizon manipulation benchmarks such as LIBERO and SimplerEnv, S²-VLA achieves state-of-the-art performance with only 2 billion parameters, outperforming existing models scaled up to 7 billion parameters.

Abstract

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.

Problem

• Vision-Language-Action models
• long-horizon manipulation
• feature fusion
• error propagation
• adaptive attention

Innovation

• State-Space Model
• Adaptive Attention
• Vision-Language-Action
• Long-Horizon Manipulation
• Dynamic Feature Fusion