外观
论文arXiv 2602.142551 个月前
面向工业机器人视觉运动策略学习的延迟感知框架
原文链接：A latency-aware framework for visuomotor policy learning on industrial robots（arXiv 2602.14255 · 发布日期：2026-09-21）
A latency-aware framework for visuomotor policy learning on industrial robots
针对工业机械臂观测—执行延迟导致的部署问题，提出延迟感知系统框架，并在接触丰富装配任务上验证。
核心要点
- 工业机械臂的高层控制接口和较慢闭环动态会放大观测、推理、执行延迟影响。
- 框架集成延迟校准多模态感知、数据同步、统一通信流水线和遥操作示范采集接口。
- 为策略预测动作序列打时间戳，仅调度时间可行的动作，无需修改策略架构或训练流程。
- 在100–500 ms推理延迟下，相比阻塞和朴素异步基线保持运动、接触和任务推进表现。
验证范围与局限：验证范围为工业机械臂上的接触丰富装配任务，并系统改变推理延迟。
英文摘要 (Abstract)
Industrial robots are increasingly deployed in construction and manufacturing tasks, where the deployment of end-to-end visuomotor policies is challenged by the observation–execution gap induced by observation, inference, and execution latencies. This gap is often significant on industrial robotic arms due to high-level control interfaces and slower closed-loop dynamics, making execution timing a dominant system-level concern. This paper presents a system-level, latency-aware framework for deploying and evaluating visuomotor policies on industrial robotic arms. The framework integrates latency-calibrated multimodal sensing, data synchronization, a unified communication pipeline, and a teleoperation interface for collecting expert demonstrations. Within this framework, we formalize a latency-aware execution strategy that assigns timestamps to policy-predicted action sequences and schedules only temporally feasible actions according to their intended execution time, enabling asynchronous inference and execution without modifying policy architectures or training procedures. We evaluate the framework on a contact-rich assembly task while systematically varying inference latency and compare latency-aware execution against blocking and naive asynchronous baselines using identical policies and sensing modalities. Results show that latency-aware execution preserves smooth motion, compliant contact behavior, and task progression consistent with demonstrations across inference latencies of 100–500 ms. Latency-aware execution maintained task duration and motion smoothness within 13% and 9% of the demonstration reference, respectively, while avoiding the latency-dependent slowdown observed under blocking execution and the large contact-force overshoots produced by naive asynchronous execution. These findings demonstrate that explicit handling of the observation–execution gap is essential for reproducible, closed-loop deployment of visuomotor policies on industrial robotic platforms.
作者：Daniel Ruan, Salma Mozaffari, Sigrid Adriaenssens, Arash Adel
来源
本页事实对应的公开页面，以原文为准。