← 首页|学术|Instruct-to-Act: Decoupling Planning and Control
cs.AI, cs.CL, cs.MA, cs.RO · 2608.26788 · 2026-08-27

Decoupling Planning and Control for Instructable Agents

Zineng Tang, Kelsey R. Allen, Sjoerd van Steenkiste, Ishita Dasgupta, Alane Suhr
规划-控制解耦VLA具身AI架构
💬 把「想清楚该干嘛」和「快速执行动作」拆成两个系统:VLM 做低频高层规划,world-model 控制器做高频低延迟执行。

🎯 背景

预训练、指令微调的视觉-语言模型(VLM)擅长把指令和观测映射成高层计划,但很难在陌生环境里把计划落实成可靠的低延迟动作序列;世界模型控制器则擅长快速的观测到动作控制,但缺乏开放式任务引导能力。

🔬 方法

Instruct-to-Act 把两者的优势合并进一个系统:训练一个世界模型控制器,在以稀疏、高延迟、高层文本指令(由 VLM 规划器生成)为条件时能高频自主行动。训练时用合成指令重新标注控制器策略 rollout 片段,把行为克隆目标和现有的奖励最大化、世界建模目标联合优化。

📊 结果与意义

在七个具身环境(含三个多智能体环境,VLM 规划器通过语言协调、训练好的控制器充当执行器)上评测,解耦方案在匹配观测/动作空间下持续优于纯控制器和直接 VLM 动作生成变体,保持了快速控制,还能不经微调直接替换不同的预训练 VLM 规划器——这种「高层慢规划 + 低层快执行」的双层解耦,与 duplex agent 里 Thinking Layer / Interaction Layer 的异步分工思路高度呼应。
▶ 原文摘要 Abstract
Recent work shows that pre-trained, instruction-tuned vision-language models (VLMs) perform well at mapping from instructions and observations to high-level plans, but struggle to realize such plans as reliable low-latency action sequences in unfamiliar environments. At the same time, world-model controllers excel at fast observation-to-action control, but lack open-ended task guidance. In this work, we combine these strengths into a single system, Instruct-to-Act, where we train a world-model controller to act autonomously at high frequency when conditioned on sparse, higher-latency, and high-level text instructions generated by a VLM planner. To train controllers to be language-instructable, we relabel segments of controller policy rollouts with synthetic instructions and jointly optimize a behavior-cloning objective along with existing reward-maximizing and world-modeling objectives. We evaluate our proposed approach across seven embodied environments, including three multi-agent environments where VLM planners coordinate through language while trained controllers serve as their actuators. Under matched observation and action spaces, our decoupled approach consistently outperforms controller-only and direct VLM action-generation variants, preserves fast control, and lets us swap in different pretrained VLM planners without fine-tuning, while remaining competitive with strong vision-language-action and multi-agent RL baselines on six of seven tasks.
来源:arXiv:2608.26788 · 精读基于摘要与 arXiv HTML/abs 页信息生成,未解析 PDF 全文