Guava: An Effective and Universal Harness for Embodied Manipulation
This work provides a scalable, model-agnostic interface for embodied manipulation, enabling compact open-source models to achieve strong embodied capabilities with minimal data.
Guava introduces a harness framework for embodied tool use that identifies three key design principles—iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations—and demonstrates that distilling these into a 4B open-source model with fewer than 2K simulation trajectories achieves performance comparable to frontier proprietary models in both simulation and real-world tasks, with strong generalization.
Language models trained on large-scale vision-language data have demonstrated strong potential for embodied agents. Harnessing models through embodied tools use offers a promising alternative to end-to-end vision-language-action systems by combining high-level reasoning with external modules for perception, planning, and control. However, it remains unclear what makes an effective harness for embodied manipulation, and to what extent such a harness can unlock embodied capabilities in a wide range of reasoning models. In this work, we present Guava, a harness framework for embodied tool use developed through systematic exploration of the design space of agent workflows, action spaces, and observation spaces. Our study identifies three key ingredients for effective embodied agents: iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations. To understand whether these design principles are universal even to small models, we develop an end-to-end training pipeline that distills embodied manipulation capabilities into a 4B open-source model using fewer than 2K trajectories collected entirely in simulation. Experimental results in both simulation and real-world environments show performance comparable to frontier proprietary models while exhibiting strong generalization to unseen objects, novel instructions, and long-horizon tasks. Results suggest that a well-designed harness can serve as a scalable, model-agnostic interface for embodied manipulation, enabling strong emergent embodied capabilities in compact open-source models with minimal training data.