Guava harness unlocks embodied manipulation in small 4B models with just 2K demos

A team of researchers from multiple universities has introduced Guava, a universal harness framework that effectively enables embodied manipulation in small, open-source language models. Through systematic exploration of agent workflows, action spaces, and observation spaces, the authors identified three critical design principles: iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations. Using these principles, they built a training pipeline that distills embodied manipulation capabilities into a 4B parameter model using fewer than 2,000 trajectories collected entirely in simulation.

Evaluation in both simulated and real-world environments shows Guava achieves performance comparable to frontier proprietary models, despite being compact and open-source. The model demonstrates strong generalization to unseen objects, novel instructions, and long-horizon tasks, suggesting that a well-designed harness can serve as a scalable, model-agnostic interface for embodied manipulation. This work provides a practical pathway for bringing advanced robotic control to smaller, more accessible models with minimal training data.