Guava Framework Enables Compact Language Models to Perform Complex Embodied Manipulation Tasks

Researchers have introduced Guava, a harness framework that enables language models to perform embodied manipulation tasks through a systematic approach to agent design. The framework represents an alternative to end-to-end vision-language-action systems by combining high-level reasoning capabilities with external modules for perception, planning, and control.

The study, submitted to arXiv on June 16, 2026, identifies three critical design principles for effective embodied agents: iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations. Through systematic exploration of the design space encompassing agent workflows, action spaces, and observation spaces, the researchers developed a comprehensive framework applicable across different language model architectures.

A key innovation of the research is an end-to-end training pipeline that distills embodied manipulation capabilities into a 4-billion-parameter open-source model using fewer than 2,000 trajectories collected entirely in simulation. This approach demonstrates that effective embodied capabilities can be achieved with minimal training data and computational resources. The researchers tested the framework in both simulated and real-world environments, showing performance comparable to proprietary frontier models.

Experimental results indicate strong generalization across unseen objects, novel instructions, and long-horizon tasks. The framework's model-agnostic design enables it to serve as a scalable interface for embodied manipulation, allowing compact open-source models to achieve emergent embodied capabilities without requiring extensive proprietary training infrastructure. These findings suggest that well-designed harnesses can democratize access to embodied AI systems by reducing the gap between small and large language models in manipulation tasks.