Improving embodied agents usually focuses on training the underlying model, while the harness that handles planning, context, and tool use is largely hand‑engineered. We ask whether this harness can self‑improve directly from sparse environmental feedback in the form of experience traces. To answer this, we introduce Embodied Agent Harness Optimization (EMHO). EMHO keeps the embodied model frozen and repeatedly analyzes execution trajectories together with past harness configurations to iteratively refine the harness logic. The optimization goes beyond adding skills or recovery prompts; it modifies progress monitoring, vision‑tool invocation, observation grounding, and failure response.
To address the conflict of supporting multiple subtasks with a single harness, EMHO‑Merge employs an evidence‑driven mechanism that uses episode‑level gains and losses to decide when and how to apply revised behaviors. We evaluate EMHO on EmbodiedBench across navigation and manipulation tasks using Qwen‑9B and Qwen‑27B models, and EMHO consistently raises task success rates. Qualitative analysis shows that EMHO not only recovers from failures but fundamentally reshapes how the embodied agent perceives and interacts with its environment.
Review