S²-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation

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 S²-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, S²-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.

Key Contributions

• Diagnoses the specific mechanism behind long-horizon VLA degradation: static (fixed-weight) fusion of visual/language/action features, which cannot adapt as a task's requirements shift across its execution phases.
• Introduces State-Space Guided Adaptive Attention (SSGAA): a belief-state tracker over task progression that generates dynamic gating weights, letting the model re-weight visual perception, task-level planning intent, and temporal action consistency differently at different points in a long task.
• Demonstrates that a compact 2B-parameter model with this adaptive-fusion architecture beats 7B-scale baseline models on long-horizon benchmarks — an architecture-over-scale result directly relevant to the "does VLA capability require ever-larger models" question.

Methodology

SSGAA maintains an internal belief state tracking how far a task has progressed. From this belief state it generates dynamic gating weights that adaptively combine three information streams: visual features (spatial perception), task intents (high-level planning), and temporal action sequences (execution consistency). Unlike static fusion (fixed weighting regardless of task phase), this lets the model's attention shift — e.g., weighting planning-level task intent more heavily early in a task and execution-consistency signals more heavily as errors could compound later. Evaluated on the LIBERO and SimplerEnv long-horizon manipulation benchmarks against larger (7B-parameter) baseline VLA models.

Results

• State-of-the-art performance on LIBERO and SimplerEnv long-horizon manipulation benchmarks.
• A 2B-parameter S²-VLA model outperforms baseline models at 7B scale — a 3.5x parameter-efficiency advantage attributed specifically to the adaptive-fusion mechanism rather than any change in underlying model capacity.

Limitations

Not explicitly detailed in the fetched abstract; as with most benchmark-only VLA papers, real-world (non-simulated) deployment results are not reported here — LIBERO and SimplerEnv are simulation benchmarks, so physical-robot generalization is untested in what was fetched.

Full Content

S²-VLA is a direct, mechanism-level answer to an open problem already logged in this KB's frontier tracker under "Zero-Shot Loco-Manipulation via Foundation Models": "Long-horizon combined tasks still 56-63% success." Rather than attacking this with more data or a bigger model (the Xiaomi-Robotics-1 approach ingested in this same sweep), S²-VLA attacks it architecturally — arguing that the failure mode is a fusion-mechanism problem (static weighting can't track task phase) rather than a capacity or data problem. The parameter-efficiency result (2B beating 7B) is the more strategically interesting finding for anyone tracking embodied-AI compute economics: it's evidence that architectural innovation can substitute for scale in at least some VLA failure modes, which cuts against a pure "more compute wins" narrative for robotics foundation models.

Source: S²-VLA: State-Space Guided Vision-Language-Action Models for Long-Horizon Manipulation by Zhipeng Xie et al.