Guava: An Effective and Universal Harness for Embodied Manipulation

A new framework called Guava enables compact open-source language models to perform embodied manipulation tasks at levels comparable to frontier proprietary systems, according to research submitted to arXiv in June 2026 [1][2]. The framework, detailed in a paper submitted on 16 June 2026, approaches embodied tool use not as an end-to-end vision-language-action problem but as a harness that combines high-level reasoning with external modules for perception, planning, and control [1][2]. Through a systematic exploration of agent workflows, action spaces, and observation spaces, the researchers identified three design principles: iterative perception-reasoning-action loops, semantic action abstractions, and multimodal observations [2]. To test whether these principles could work with smaller models, the team built a 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 [2]. The resulting system demonstrated performance comparable to frontier proprietary models in both simulated and real-world environments, while generalizing to unseen objects, novel instructions, and long-horizon tasks [2]. The paper appears on arXiv, the open-access repository of electronic preprints that has served the physics, mathematics, and computer science communities since 1991 [6]. As of late 2024, arXiv was receiving roughly 24,000 new articles per month and had surpassed two million total articles by the end of 2021 [6]. Submissions are moderated but not peer-reviewed before posting [6]. The Guava paper is accessible through arXiv's standard abstract page, which includes experimental community features developed under the arXivLabs framework [3][4]. arXivLabs, launched in 2020, provides a formalized structure for third-party collaborators to build tools that enhance the reading and discovery experience while adhering to arXiv's values of openness, community, excellence, and user data privacy [4]. Current Labs integrations include the Bibliographic Explorer for navigating citation trees, the CORE Recommender for surfacing related open-access papers, and Connected Papers for visualizing research landscapes [5]. The Guava results suggest that a well-designed harness can function as a scalable, model-agnostic interface for embodied manipulation, opening the door for compact open-source models to take on physical-world tasks that previously required much larger proprietary systems [2].