Libra-VLA: Achieving Learning Equilibrium
via Asynchronous Coarse-to-Fine Dual-System

Vision-Language-Action (VLA) models have emerged as a promising paradigm for generalist robotic manipulation by grounding high-level semantic instructions into executable physical actions. However, prevailing approaches typically adopt a monolithic generation paradigm, attempting to map high-level visual-linguistic features directly to high-frequency motor commands in a flat, non-hierarchical fashion. This strategy overlooks the inherent hierarchy of robotic manipulation, where complex actions can be naturally modeled in a Hybrid Action Space, decomposing into discrete macro-directional reaching and continuous micro-pose alignment. Bypassing this structural hierarchy widens the semantic-actuation gap, compelling models to rely on cross-modal alignment via massive-scale robotic pre-training to bridge the gap.

To address this, we introduce Libra-VLA, a novel Coarse-to-Fine Dual-System VLA architecture. Our core insight is twofold: we explicitly decouple the learning complexity into a coarse-to-fine hierarchy to strike a training equilibrium, while simultaneously leveraging this structural modularity to implement an asynchronous execution strategy. Specifically, our architecture integrates: (1) Semantic Planner, which focuses on macro-intent, predicting discrete action tokens that guide the robot's general direction, and (2) Action Refiner, which conditions on the coarse intent to generate high-frequency continuous actions for precise pose alignment.

Crucially, our empirical analysis reveals that performance follows an inverted-U curve relative to the granularity of action decomposition. We demonstrate that Libra-VLA achieves best performance exactly when the learning difficulty is balanced between the two sub-systems. Combined with the asynchronous design, our approach offers a scalable, data-efficient, and responsive solution for open-world manipulation.