面向化学机器人的ACT动作块自适应优化方法

doi:10.16182/j.issn1004731x.joss.26-0061

系统仿真学报 ›› 2026, Vol. 38 ›› Issue (9): 2559-2573.doi: 10.16182/j.issn1004731x.joss.26-0061

面向化学机器人的ACT动作块自适应优化方法

李潭¹^,², 吴聪营¹, 李生华¹, 陈南江¹^,², 宋伟宁³^,⁴, 高方方¹^,⁴, 王琪²^,⁴, 林燕文²^,⁴, 殷润民²^,⁴

^1.南昌大学信息工程学院，江西南昌 330000
^2.青葵智造(北京)科技有限公司，北京 100080
^3.东华理工大学信息工程学院，江西南昌 330000
^4.南昌市具身元宇宙关键技术研发与应用重点实验室，江西南昌 330000

收稿日期:2026-01-20 修回日期:2026-04-27 出版日期:2026-09-30 发布日期:2026-10-09
通讯作者: 陈南江
第一作者简介:李潭(1986-)，男，教授，博士，研究方向为工业互联网、复杂系统仿真、工业元宇宙、具身智能等。

ACT Action Chunk Adaptive Optimization Method for Chemistry-oriented Robots

Li Tan¹^,², Wu Congying¹, Li Shenghua¹, Chen Nanjiang¹^,², Song Weining³^,⁴, Gao Fangfang¹^,⁴, Wang Qi²^,⁴, Lin Yanwen²^,⁴, Yin Runmin²^,⁴

^1.School of Information Engineering, Nanchang University, Nanchang 330000, China
^2.Quiktech (Beijing) Technology Co. , Ltd. , Beijing 100080, China
^3.School of Information Engineering, East China University of Technology, Nanchang 330000, China
^4.Nanchang Key Laboratory of Embodied Metaverse Key Technology R&D and Application, Nanchang 330000, China

Received:2026-01-20 Revised:2026-04-27 Online:2026-09-30 Published:2026-10-09
Contact: Chen Nanjiang

摘要/Abstract

摘要：

化学研发实验室长期面临研发周期长、实验重复性高等问题。随着实验自动化与智能化需求的不断提升，能够学习人类专家操作经验、并在复杂实验环境中执行任务的具身智能机器人，成为提升化学研发效率、降低人为误差的重要发展方向。虽然当前具身智能发展迅猛，但其多聚焦通用操作，针对化学实验室场景的研究匮乏，传统ACT算法因动作块大小固定，难以满足化学操作需求，以优化化学场景模仿学习算法为目标，提出任务感知的ACT动作块自适应优化框架。在硬件平台上，采用宇树G1-29edu机器人(搭载inspire RH56DFXW灵巧手)与UR5的6自由度工业机械臂搭配柔性夹爪，通过遥操作采集化学操作训练数据集；从动作平滑度、操作精度要求、任务时长维度，建立任务特征量化指标体系，支撑动作块大小自适应调整；针对化学操作对抖动的敏感性，在模型输出端增加动作平滑处理模块。实验结果表明，与原始的ACT算法相比，自适应调整动作块大小的策略在任务成功率、动作平滑度以及任务耗时方面都取得了显著的提升，在任务时长方面，大部分任务的耗时均低于原始算法的85%，在部分任务中低于原始算法的66.7%，动作平滑度提升12%~25%，在化学实验场景中的任务成功率提升10%~20%，部分对于精度和平滑度要求较高的任务成功率能提升25%以上。

关键词: 模仿学习, 化学操作, ACT, 自适应, 动作块, 化学实验室机器人

Abstract:

Chemical R&D laboratories face persistent challenges such as long R&D cycles and high experimental repetitiveness. With the growing demand for automation and intelligence in laboratory processes, embodied intelligent robots capable of learning expert operational experience and executing tasks stably in complex environments have become a key direction for improving chemical R&D efficiency and reducing human error. Although the hardware and software for embodied intelligence are advancing rapidly, current research focuses primarily on general-purpose operations, with limited attention to the specific constraints of chemistry laboratory settings. Traditional ACT algorithms, which use fixed action chunk sizes, fail to meet the requirements of chemical operations. Aims to optimize imitation learning algorithms for chemical scenarios by proposing a task-aware adaptive action chunk size optimization framework based on ACT. On the hardware side, the system uses a Unitree G1-29edu robot equipped with an inspire RH56DFXW dexterous hand, along with a UR5 6-degree-of-freedom industrial manipulator fitted with a flexible gripper. A teleoperation system is used to collect a dataset of chemical operation demonstrations. A quantitative indicator system for task features is then established based on three dimensions: motion smoothness, precision requirements, and task duration, enabling adaptive adjustment of action chunk sizes. Additionally, to address the sensitivity of chemical operations to jitter, a motion smoothing module is added at the model output. Experimental results show that compared with the original ACT algorithm, the adaptive action chunk size strategy achieves significant improvements in task success rate, motion smoothness, and task duration. For most tasks, the completion time is reduced to below 85% of that of the original algorithm, and in some tasks, to below 66.7%. Motion smoothness improves by 12% to 25%, and task success rates in chemistry laboratory scenarios increase by 10% to 20%. For tasks requiring high precision and smoothness, the success rate improvement exceeds 25%.

Key words: imitation learning, chemical operations, action chunk with transformer (ACT), adaptivity, action chunks, chemistry laboratory robot

中图分类号:

TP242.6

李潭,吴聪营,李生华等 . 面向化学机器人的ACT动作块自适应优化方法[J]. 系统仿真学报, 2026, 38(9): 2559-2573.

Li Tan,Wu Congying,Li Shenghua,et al . ACT Action Chunk Adaptive Optimization Method for Chemistry-oriented Robots[J]. Journal of System Simulation, 2026, 38(9): 2559-2573.

图/表 14

图1

图2

表1

图3

图4

表2

图5

图6

图7

图8

表3

任务特征量化

任务	精度要求 $P$	平滑度特征 $S$	任务时长 $T$ /s	块大小 $k *$
物品抓取与放置	0.10	0.251 0	55	138
移液	0.67	1.331 5	32	64
打开容量瓶	0.20	0.458 1	25	57
转移试剂瓶	0.08	0.460 1	30	173

表3

表4

图9

图10

参考文献 44

[1]	Yoshikawa Naruki, Skreta Marta, Darvish Kourosh, et al. Large Language Models for Chemistry Robotics[J]. Autonomous Robots, 2023, 47(8): 1057-1086.
[2]	Song Tao, Luo Man, Zhang Xiaolong, et al. A Multiagent-driven Robotic AI Chemist Enabling Autonomous Chemical Research on Demand[J]. Journal of the American Chemical Society, 2025, 147(15): 12534-12545.
[3]	王可, 张辉, 曹意宏, 等. 面向医药生产的智能机器人及其关键技术研究综述[J]. 计算机集成制造系统, 2022, 28(7): 1981-1995.
	Wang Ke, Zhang Hui, Cao Yihong, et al. Overview of Intelligent Robots and Key Technologies for Pharmaceutical Production[J]. Computer Integrated Manufacturing Systems, 2022, 28(7): 1981-1995.
[4]	Zare Maryam, Kebria Parham M, Khosravi Abbas, et al. A Survey of Imitation Learning: Algorithms, Recent Developments, and Challenges[J]. IEEE Transactions on Cybernetics, 2024, 54(12): 7173-7186.
[5]	Zhang Tianhao, Mccarthy Z, Jow O, et al. Deep Imitation Learning for Complex Manipulation Tasks from Virtual Reality Teleoperation[C]//2018 IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2018: 5628-5635.
[6]	Wang Chen, Fan Linxi, Sun Jiankai, et al. MimicPlay: Long-horizon Imitation Learning by Watching Human Play[EB/OL]. (2023-10-13) [2026-01-11]. .
[7]	Fu Zipeng, Zhao Qingqing, Wu Qi, et al. HumanPlus: Humanoid Shadowing and Imitation from Humans[EB/OL]. (2024-06-15) [2026-01-11]. .
[8]	He Tairan, Gao Jiawei, Xiao Wenli, et al. ASAP: Aligning Simulation and Real-world Physics for Learning Agile Humanoid Whole-body Skills[EB/OL]. (2025-04-26) [2026-01-11]. .
[9]	Lin Baihan. Behavioral Cloning and Imitation Learning[M]//Lin Baihan. Reinforcement Learning Methods in Speech and Language Technology. Cham: Springer Nature Switzerland, 2025: 63-67.
[10]	Xie Jun, Wang Zhicheng, Tan Jianwei, et al. Subconscious Robotic Imitation Learning[EB/OL]. (2024-12-29) [2026-01-11]. .
[11]	Li Zezeng, Chapin Alexandre, Xiang Enda, et al. Robotic Manipulation via Imitation Learning: Taxonomy, Evolution, Benchmark, and Challenges[EB/OL]. (2025-09-04) [2026-01-11]. .
[12]	Pomerleau D A. ALVINN: An Autonomous Land Vehicle in a Neural Network[C]//Advances in Neural Information Processing Systems. San Francisco: Morgan Kaufmann Publishers Inc., 1989: 305-313.
[13]	Foster D J, Block A, Misra D. Is Behavior Cloning All You Need? Understanding Horizon in Imitation Learning[C]//38th Conference on Neural Information Processing Systems (NeurIPS 2024). Red Hook: Curran Associates Inc., 2024: 120602-120666.
[14]	Zhao T Z, Kumar V, Levine S, et al. Learning Fine-grained Bimanual Manipulation with Low-Cost Hardware[EB/OL]. (2023-04-23) [2026-01-11]. .
[15]	Kim Heecheol, Ohmura Yoshiyuki, Kuniyoshi Yasuo. Robot Peels Banana with Goal-conditioned Dual-action Deep Imitation Learning[EB/OL]. (2022-03-18) [2026-01-11]. .
[16]	Paradis S, Hwang M, Thananjeyan B, et al. Intermittent Visual Servoing: Efficiently Learning Policies Robust to Instrument Changes for High-precision Surgical Manipulation[C]//2021 IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2021: 7166-7173.
[17]	Zhao Haitao, Chen Wei, Huang Hao, et al. A Robotic Platform for the Synthesis of Colloidal Nanocrystals[J]. Nature Synthesis, 2023, 2(6): 505-514.
[18]	Jiang Ying, Fakhruldeen H, Pizzuto G, et al. Autonomous Biomimetic Solid Dispensing Using a Dual-arm Robotic Manipulator[J]. Digital Discovery, 2023, 2(6): 1733-1744.
[19]	Burger B, Maffettone P M, Gusev V V, et al. A Mobile Robotic Chemist[J]. Nature, 2020, 583(7815): 237-241.
[20]	Zhu Qing, Zhang Fei, Huang Yan, et al. An All-round AI-chemist with a Scientific Mind[J]. National Science Review, 2022, 9(10): nwac190.
[21]	Guo Taicheng, Chen Xiuying, Wang Yaqi, et al. Large Language Model Based Multi-agents: A Survey of Progress and Challenges[EB/OL]. (2024-04-19) [2026-01-11]. .
[22]	Pari J, Shafiullah N M, Arunachalam S P, et al. The Surprising Effectiveness of Representation Learning for Visual Imitation[EB/OL]. (2021-12-06) [2026-01-11]. .
[23]	Rahmatizadeh R, Abolghasemi P, Bölöni Ladislau, et al. Vision-based Multi-task Manipulation for Inexpensive Robots Using End-to-end Learning from Demonstration[C]//2018 IEEE International Conference on Robotics and Automation (ICRA). Piscataway: IEEE, 2018: 3758-3765.
[24]	Liang Wensheng, Xie Jun, Wang Zhicheng, et al. Constrained Behavior Cloning for Robotic Learning[EB/OL]. (2024-08-20) [2026-01-11]. .
[25]	贾政轩, 林廷宇, 肖莹莹, 等. 基于强化学习的最优控制指令模仿生成方法[J]. 系统仿真学报, 2023, 35(11): 2410-2418.
	Jia Zhengxuan, Lin Tingyu, Xiao Yingying, et al. Imitative Generation of Optimal Guidance Law Based on Reinforcement Learning[J]. Journal of System Simulation, 2023, 35(11): 2410-2418.
[26]	顾俊, 程文, 林子昂, 等. 智能驾驶汽车模仿学习技术研究综述[J]. 无人系统技术, 2025, 8(6): 100-113.
	Gu Jun, Cheng Wen, Lin Ziang, et al. A Review of Imitation Learning Technology for Intelligent Driving Cars[J]. Unmanned Systems Technology, 2025, 8(6): 100-113.
[27]	Zhang Tengteng, Mo Hongwei. Reinforcement Learning for Robot Research: A Comprehensive Review and Open Issues[J]. International Journal of Advanced Robotic Systems, 2021, 18(3): 17298814211007305.
[28]	Black K, Brown N, Driess D, et al. π₀ : A Vision-language-action Flow Model for General Robot Control[EB/OL]. (2026-01-08) [2026-01-11]. .
[29]	Chi Cheng, Xu Zhenjia, Feng Siyuan, et al. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion[J]. International Journal of Robotics Research, 2025, 44(10/11): 1684-1704.
[30]	Yanjie Ze, Zhang Gu, Zhang Kangning, et al. 3D Diffusion Policy: Generalizable Visuomotor Policy Learning via Simple 3D Representations[EB/OL]. (2024-09-27) [2026-01-11]. .
[31]	Intelligence P, Black K, Brown N, et al. π₀ 5: A Vision-language-action Model with Open-world Generalization[EB/OL]. (2026-01-08) [2026-01-11]. .
[32]	Kang Bingyi, Ma Xiao, Du Chao, et al. Efficient Diffusion Policies for Offline Reinforcement Learning[C]//Proceedings of the 37th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2023: 67195-67212.
[33]	Janner M, Li Qiyang, Levine S. Offline Reinforcement Learning as One Big Sequence Modeling Problem[C]//Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 1273-1286.
[34]	Chen Lili, Lu K, Rajeswaran A, et al. Decision Transformer: Reinforcement Learning via Sequence Modeling[C]//Proceedings of the 35th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2021: 15084-15097.
[35]	Reed Scott, Zolna Konrad, Parisotto Emilio, et al. A Generalist Agent[EB/OL]. (2022-11-11) [2026-01-11]. .
[36]	Buamanee Thanpimon, Kobayashi Masato, Uranishi Yuki, et al. Bi-ACT: Bilateral Control-based Imitation Learning via Action Chunking with Transformer[C]//2024 IEEE International Conference on Advanced Intelligent Mechatronics (AIM). Piscataway: IEEE, 2024: 410-415.
[37]	Lee A, Chuang I, Chen Lingyuan, et al. InterACT: Inter-dependency Aware Action Chunking with Hierarchical Attention Transformers for Bimanual Manipulation[EB/OL]. (2024-10-16) [2026-01-11]. .
[38]	Liu Yuejiang, Hamid J I, Xie A, et al. Bidirectional Decoding: Improving Action Chunking via Guided Test-time Sampling[EB/OL]. (2025-04-25) [2026-01-11]. .
[39]	Zhang Xinyu, Liu Yuhan, Chang Haonan, et al. Autoregressive Action Sequence Learning for Robotic Manipulation[J]. IEEE Robotics and Automation Letters, 2025, 10(5): 4898-4905.
[40]	Chaudhary S, Ravindran B. Smooth Imitation Learning via Smooth Costs and Smooth Policies[C]//Proceedings of the 5th Joint International Conference on Data Science & Management of Data (9th ACM IKDD CODS and 27th COMAD). New York: ACM, 2022: 63-71.
[41]	Kim Heecheol, Ohmura Yoshiyuki, Kuniyoshi Yasuo. Gaze-based Dual Resolution Deep Imitation Learning for High-precision Dexterous Robot Manipulation[J]. IEEE Robotics and Automation Letters, 2021, 6(2): 1630-1637.
[42]	Oh Hanbit, Matsubara Takamitsu. Leveraging Demonstrator-perceived Precision for Safe Interactive Imitation Learning of Clearance-limited Tasks[J]. IEEE Robotics and Automation Letters, 2024, 9(4): 3387-3394.
[43]	Hu Yingbai, Abu-Dakka Fares J, Chen Fei, et al. Fusion Dynamical Systems with Machine Learning in Imitation Learning: A Comprehensive Overview[J]. Information Fusion, 2024, 108: 102379.
[44]	Qin Yuzhe, Yang Wei, Huang Binghao, et al. AnyTeleop: A General Vision-based Dexterous Robot Arm-hand Teleoperation System[EB/OL]. (2024-05-16) [2026-01-11]. .

时间长度/帧	自相关系数
585	71.51
769	75.67
865	75.92
1 073	80.61
1 275	88.86

面向化学机器人的ACT动作块自适应优化方法

ACT Action Chunk Adaptive Optimization Method for Chemistry-oriented Robots

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 14

参考文献 44

相关文章 15

编辑推荐

Metrics

本文评价

任务	算法模型	总任务数	成功次数	成功率
物品抓取与放置	ACS-ACT	45	42	0.93
	ACT	45	37	0.82
	DP	25	8	0.32
移液	ACS-ACT	50	30	0.60
	ACT	50	17	0.34
	DP	20	3	0.15
打开容量瓶	ACS-ACT	50	44	0.88
打开容量瓶	ACT	50	33	0.66
转移试剂瓶	ACS-ACT	60	59	0.98
转移试剂瓶	ACT	60	52	0.87

[1]	高芳征, 郭旭, 孙权, 黄家才, 张中才. 人工势场引导与ACC自适应调节的RRT*机械臂路径规划[J]. 系统仿真学报, 2026, 38(9): 2755-2772.
[2]	施颖颖, 董骁翀, 傅国斌, 马苗苗, 李延和, 王学斌. 基于数据缺失容忍扩散的风电场景预测方法[J]. 系统仿真学报, 2026, 38(8): 2167-2178.
[3]	王政凯, 于轶蓉, 胡青. 基于二维二阶矢量水听器的线谱增强技术[J]. 系统仿真学报, 2026, 38(8): 2312-2323.
[4]	陈石宇, 陈新民, 胡雄龙, 王亨, 韩业鹏, 陈佳杰. 基于改进超螺旋滑模观测器的PMSM无位置传感器控制[J]. 系统仿真学报, 2026, 38(8): 2324-2337.
[5]	杨永琪, 王聪, 张宏立, 马萍, 孟月, 周天浩. 基于自适应时空图神经网络的居民区净负荷预测[J]. 系统仿真学报, 2026, 38(7): 2007-2019.
[6]	姚磊, 郑子健, 李天皓, 迟玉伦. 基于IGWO-AEKF的永磁同步电机参数辨识[J]. 系统仿真学报, 2026, 38(6): 1613-1627.
[7]	刘满强, 尚自强. 改进多目标差分算法应用于机械臂多目标轨迹规划研究[J]. 系统仿真学报, 2026, 38(6): 1647-1668.
[8]	周子聪, 曾俊杰, 胡越, 朱正秋, 尹全军. 基于次优示例引导的兵棋推演多智能体强化学习方法[J]. 系统仿真学报, 2026, 38(5): 1277-1289.
[9]	刘威, 陈德龙, 刘泽, 王锐, 李凯文, 张涛. 基于自适应混合进化的防空反导火力资源分配优化[J]. 系统仿真学报, 2026, 38(4): 959-973.
[10]	刘沛津, 张闽心, 何林, 孙艺阁, 苏庭琪. 面向城市复杂环境视觉地点识别算法研究[J]. 系统仿真学报, 2026, 38(3): 818-828.
[11]	朱玲, 李靖, 张朝辉. 基于改进REA*算法的机器人自适应路径规划[J]. 系统仿真学报, 2026, 38(2): 332-345.
[12]	魏呈彪, 赵涛岩, 曹江涛, 李平. 深度模糊神经网络的设计和预测[J]. 系统仿真学报, 2025, 37(9): 2200-2210.
[13]	刘子龙, 张磊. 自然环境下改进YOLOv5对小目标苹果的检测[J]. 系统仿真学报, 2025, 37(8): 2124-2138.
[14]	纪志成, 全震, 王艳. 基于混合决策机制的自适应生产调度优化与仿真[J]. 系统仿真学报, 2025, 37(7): 1791-1803.
[15]	李庥甜, 王凌, 陈英武, 邢立宁, 陈盈果. 基于自适应大邻域搜索的多场景多卫星任务规划方法[J]. 系统仿真学报, 2025, 37(7): 1836-1847.