Agent 走向具身世界：从语言智能到机器人「大脑指挥官」

TL;DR - Huawei Noah’s Ark Lab’s RoboHarness uses a coding agent to orchestrate heterogeneous robot policies—VLA, world-action models, reinforcement learning, and task-and-motion planning—for zero-shot, long-horizon tasks. It matters because it improves robotic capability through modular coordination without retraining or modifying the underlying policies.

• RoboHarness decomposes tasks and dynamically routes subtasks using quantitative scene signals, execution history, and online feedback.
• Its Memory Bridge retrieves successful trajectories and plans transitions into states where the next policy can reliably take control.
• Removing Memory Bridge reduced full-task success from 86.0% to 60.4%, highlighting policy handoffs as a key failure point.
• The approach complements embodied foundation-model scaling with an agentic hierarchy, though it remains limited by available policies and historical execution data.