EMS framework lets robots react at 93.4 Hz while keeping task accuracy

Embodied AI systems have long struggled to balance long-horizon reasoning with real-time closed-loop control. Recent Vision-Language-Action (VLA) architectures adopt a dual-system approach—a slow, deliberative model for global planning and a fast, reactive model for immediate actions. But existing designs tightly couple these modules, forcing end-to-end joint training and limiting flexibility. A new paper introduces Environment-aware Model Selection (EMS), which completely decouples the two systems and uses a reinforcement-learning-based switching policy to decide which model to invoke based on real-time environmental feedback. This sparse activation of the slow system preserves pretrained knowledge while keeping runtime efficiency high.

Evaluated on the LIBERO benchmark, EMS achieves success rates comparable to large-scale baselines while raising effective action frequency to 93.4 Hz—far beyond what typical deliberative systems allow. The framework also proves extensible in real-world dual-arm manipulation tasks, where it accelerates task completion without sacrificing robust performance. Because the two systems are fully decoupled, users can swap in different models without retraining the whole stack, making EMS a modular, adaptable foundation for next-generation embodied agents that need to both reason carefully and react instantly.