Libra-VLA: Achieving Learning Equilibrium via Asynchronous Coarse-to-Fine Dual-System
A dual-system AI balances coarse and fine actions, boosting robot performance by 40%.
Libra-VLA, accepted at the ACL 2026 main conference, tackles a fundamental flaw in Vision-Language-Action (VLA) models: the flat, monolithic generation of motor commands. Traditional VLAs directly map visual-linguistic features to high-frequency actions, ignoring the natural hierarchy of manipulation. This creates a semantic-actuation gap, forcing models to simultaneously handle high-level intent and low-level precision, leading to inefficiency and instability.
Libra-VLA introduces a Coarse-to-Fine Dual-System architecture. The Semantic Planner predicts discrete tokens for macro-directional intent (e.g., moving an arm toward a cup), while the Action Refiner generates continuous, high-frequency actions for precise alignment (e.g., gripping the handle). Crucially, the researchers found that performance follows an inverted-U curve relative to action decomposition granularity, peaking when learning difficulty is balanced between the two sub-systems. This asynchronous execution strategy offers scalable, robust, and responsive open-world manipulation, potentially reducing task completion time by 2x and improving success rates by 40% over monolithic baselines.
- Libra-VLA splits VLA models into two sub-systems: a Semantic Planner (discrete macro-actions) and an Action Refiner (continuous micro-alignment).
- Performance peaks at an optimal decomposition granularity, following an inverted-U curve, achieving a learning equilibrium.
- Accepted at ACL 2026, the architecture enables asynchronous execution for more robust and responsive robotic manipulation.
Why It Matters
This dual-system approach could make robots faster and more reliable for real-world tasks like assembly or surgery.