Brief

Researchers have developed Guava, a framework designed to enhance embodied manipulation capabilities in AI agents by integrating high-level reasoning with external modules for perception, planning, and control. This harness identifies iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations as key components for effective embodied agents. Guava has demonstrated its ability to distill complex manipulation skills into a compact 4B open-source model with minimal training data, achieving performance comparable to frontier proprietary models in both simulated and real-world environments. Separately, the PhysVLA framework offers a plug-and-play solution that wraps existing Vision-Language-Action models to enforce physical principles like rigid-body dynamics and contact constraints without retraining, significantly improving robotic manipulation success rates and stability. AI