[论文] ChunkAct: Action Chunking for Agentic Robotics
研究领域: NLP 作者: Bingxin Xu, Wenrui Bao, Xinxin Liu, Jiawei Zhou, Yuzhang Shang, Emilio Ferrara 发布时间: 2026-10-13 arXiv: 2610.18874
论文概要
研究领域: NLP 作者: Bingxin Xu, Wenrui Bao, Xinxin Liu, Jiawei Zhou, Yuzhang Shang, Emilio Ferrara 发布时间: 2026-10-13 arXiv: 2610.18874
中文摘要
Agentic 机器人技术将 LLM 智能体置于机器人控制回路中,通过记忆模块调度感知、运动和视觉-语言-动作(VLA)工具,从而无需重训练即可泛化到新任务。但这种通用性代价高昂:每个原子操作都需要 LLM 基于完整交互历史进行一轮推理,导致整个回合的成本随原子操作数量线性增长。我们发现这种推理在很大程度上是冗余的——一旦选定子目标,其原子操作就已由记忆中的例程所决定。本文提出 ChunkAct,将动作分块引入 agentic 机器人技术。在线阶段,单次调用即可启动一串链式原子操作组成的块,该块自主执行,仅在失败时才将控制权交还;离线阶段,ChunkAct 将记忆编译为这些块,使测试时执行无需记忆,规划器仅保留给需要推理的决策。因此 LLM 计算量随决策次数而非原子操作数量扩展。在 LIBERO-Pro 基准上,ChunkAct 在任务成功率上比当前最优的 agentic 机器人方法提升 5.0 个百分点,同时减少 2.9 倍 LLM 调用、4.4 倍输入 token 和 2.3 倍输出 token。
原文摘要
Agentic robotics places a large language model (LLM) agent in the robot control loop, where it composes perception, motion, and vision-language-action (VLA) tools under the guidance of a memory and thereby generalizes to new tasks without retraining. This generality comes at a steep price in efficiency. Every primitive requires a round of LLM reasoning over the full interaction history, so the cost of an episode grows with the number of primitives. We find this reasoning largely redundant: once a sub-goal is chosen, its primitives follow a routine already prescribed by memory. We present ChunkAct, which introduces action chunking to agentic robotics. Online, a single call launches a chunk of chained primitives, which executes autonomously and returns control only upon failure. Offline, Chu...
*自动采集于 2026-10-15*
#论文 #arXiv #NLP #小凯