[论文] PGX: Physically Grounded Execution for Long-Horizon Humanoid Loco-Mani...
论文概要
研究领域: ML
作者: Kai Lu, Chenyang Ma, Ioannis Havoutis
发布时间: 2026-11-19
arXiv: 2611.26596
中文摘要
机器人视觉-语言-动作(VLA)模型能执行多样的单个操作技能,但在长时程移动操作任务上仍显吃力——时间结构、技能和精度在单一模型中紧密耦合。我们提出物理落地执行(PGX)框架,将时间和精度控制卸载给两个经典免训练组件,使VLA专注于阶段性技能动态。PGX将长时程移动操作分解为三层:(i)基于谓词的状态机管理任务推进,用物理谓词门控阶段转换、触发恢复并协调导航与等待,共享谓词语义连接离线示范分割、在线执行和阶段评估;(ii)VLA技能在活跃操作阶段内运行,扁平或分阶段训练的模型专注于任务相关动作;(iii)世界坐标锁定(WCL)通过基座位置保持、物体相对定位和世界锚定接触目标维持空间一致性。我们在三个模拟任务和一个真实世界任务上通过大量实验证明PGX的有效性,使用Unitree G1人形机器人执行长时程移动操作,在阶段推进和任务成功率上均有提升。项目页面:https://kl-research.github.io/pgx
原文摘要
Robotic vision-language-action (VLA) models can execute diverse individual manipulation skills, yet they still struggle with long-horizon loco-manipulation tasks. We propose Physically Grounded Execution (PGX), a framework that offloads temporal and precision control to two classical training-free components. PGX decomposes long-horizon loco-manipulation into three layers: predicate-based state machine, VLA skills, and world-coordinate locking. Experiments on three simulated tasks and one real-world task using Unitree G1 show improvements in stage progression and task success rates.
自动采集于 2026-11-21
#论文 #arXiv #ML #小凯