Robotic end-to-end dynamic grasping method based on curriculum reinforcement learning

doi:10.11772/j.issn.1001-9081.2025060749

Journal of Computer Applications ›› 2026, Vol. 46 ›› Issue (7): 2307-2317.DOI: 10.11772/j.issn.1001-9081.2025060749

• Multimedia computing and computer simulation • Previous Articles

Robotic end-to-end dynamic grasping method based on curriculum reinforcement learning

Yanyang LIANG¹, Wenxuan XIE²(), Wei CUI², Hongfei LYU², Da LI², Dongzhou ZHONG¹

^1.School of Electronics and Information Engineering，Wuyi University，Jiangmen Guangdong 529020，China
^2.School of Mechanical and Automation Engineering，Wuyi University，Jiangmen Guangdong 529020，China

Received:2025-07-09 Revised:2025-09-12 Accepted:2025-09-24 Online:2025-10-27 Published:2026-07-10
Contact: Wenxuan XIE
About author:LIANG Yanyang， born in 1980， Ph. D.， associate professor. His research interests include machine vision， robotic motion control.
CUI Wei， born in 2000， M. S. candidate. His research interests include machine vision， image processing.
LYU Hongfei， born in 1998， M. S. candidate. Her research interests include robotic motion control， machine learning.
LI Da， born in 2002， M. S. candidate. His research interests include embodied control.
ZHONG Dongzhou， born in 1977， Ph. D.， professor. His research interests include laser and optoelectronics， photonic neural networks.
Supported by:
National Natural Science Foundation of China(62075168)

基于课程强化学习的机器人端到端动态抓取方法

梁艳阳¹, 谢文轩²(), 崔伟², 吕洪妃², 李达², 钟东洲¹

^1.五邑大学电子与信息工程学院，广东江门 529020
^2.五邑大学机械与自动化工程学院，广东江门 529020

通讯作者: 谢文轩
作者简介:梁艳阳（1980—），男，广东阳江人，副教授，博士，主要研究方向：机器视觉、机器人运动控制
崔伟（2000—），男，河南周口人，硕士研究生，主要研究方向：机器视觉、图像处理
吕洪妃（1998—），女，河南新乡人，硕士研究生，主要研究方向：机器人运动控制、机器学习
李达（2002—），男，湖南岳阳人，硕士研究生，主要研究方向：具身控制
钟东洲（1977—），男，江西瑞金人，教授，博士，主要研究方向：激光与光电子、光子神经网络。
基金资助:
国家自然科学基金资助项目(62075168)

Abstract

Abstract:

To address the low efficiency and the difficulty of balancing high success rate with motion smoothness in robotic end-to-end dynamic grasping tasks， a robotic end-to-end dynamic grasping method based on Curriculum Reinforcement Learning （CRL） was proposed. First， a multi-modal input network that fuses color images， depth maps， and robot proprioceptive states was constructed to map raw sensory data directly to continuous action commands for the end-effector. Second， a curriculum mechanism with synchronously increasing difficulty and smoothness constraints was designed， and combined with a staged reward function， the agent was guided to master grasping from static to dynamic ones progressively. Finally， Domain Randomization （DR） was employed to enhance the policy's transfer capability of Simulation-to-Reality （Sim-to-Real）. Simulation results show that the proposed method achieves a grasping success rate of nearly 100% at target speeds ranging from 0.15 to 0.40 m/s， elevating the upper speed limit for stable grasping from 0.25 m/s of the object detection-based baseline method to 0.40 m/s. Compared to Simple Curriculum Learning （SimpleCL） with only increasing difficulty， the proposed method increases the success rate by 3.6 percentage points and reduces the average joint acceleration and jerk norm by 58.34% and 69.25%， respectively， in the most difficult test. In physical experiments， the grasping success rates of the proposed method for static scene and two dynamic scenes are 95.0%， 90.0%， and 70.0%， respectively. It can be seen that this method effectively coordinates success rate and smoothness in robotic dynamic grasping tasks by jointly optimizing task difficulty and behavioral constraints.

Key words: dynamic grasping, reinforcement learning, curriculum learning, end-to-end learning, robotic manipulation

摘要：

针对机器人学习端到端动态抓取任务时效率低且难以兼顾高成功率与运动平滑性的问题，提出一种基于课程强化学习（CRL）的机器人端到端动态抓取方法。首先，构建融合彩色图、深度图与机器人本体状态的多模态输入网络，将原始感知直接映射为末端执行器的连续动作指令；其次，设计难度与平滑性约束同步递增的课程机制，并结合阶段化的奖励函数引导智能体从静态抓取逐步掌握至动态抓取；最后，采用域随机化（DR）增强策略从仿真到现实（Sim-to-Real）的迁移能力。仿真实验结果表明，在0.15~0.40 m/s的目标速度下，本文方法的抓取成功率接近100%，将稳定抓取的速度上限从基于目标检测的基线方法的0.25 m/s提升至0.40 m/s；相较于仅有难度递增的简单课程学习（SimpleCL），本文方法在最高难度测试中的抓取成功率提升了3.6个百分点，且关节加速度与加加速度的平均范数分别降低了58.34%和69.25%；在物理实验中，本文方法对静态场景及2种动态场景的抓取成功率分别为95.0%、90.0%和70.0%。本文方法通过协同优化任务难度与行为约束，在机器人动态抓取任务中实现了成功率与平滑性的有效协调。

关键词: 动态抓取, 强化学习, 课程学习, 端到端学习, 机器人操作

CLC Number:

TP242.6

Yanyang LIANG, Wenxuan XIE, Wei CUI, Hongfei LYU, Da LI, Dongzhou ZHONG. Robotic end-to-end dynamic grasping method based on curriculum reinforcement learning[J]. Journal of Computer Applications, 2026, 46(7): 2307-2317.

梁艳阳, 谢文轩, 崔伟, 吕洪妃, 李达, 钟东洲. 基于课程强化学习的机器人端到端动态抓取方法[J]. 《计算机应用》唯一官方网站, 2026, 46(7): 2307-2317.

Figures/Tables 21

References 33

[1]	Keshvarparast A， Battini D， Battaia O， et al. Collaborative robots in manufacturing and assembly systems： literature review and future research agenda ［J］. Journal of Intelligent Manufacturing， 2024， 35（5）： 2065-2118.
[2]	Duan J， Yu S， Tan H L， et al. A survey of embodied AI： from simulators to research tasks ［J］. IEEE Transactions on Emerging Topics in Computational Intelligence， 2022， 6（2）： 230-244.
[3]	Elguea-Aguinaco Í， Serrano-Muñoz A， Chrysostomou D， et al. A review on reinforcement learning for contact-rich robotic manipulation tasks ［J］. Robotics and Computer-Integrated Manufacturing， 2023， 81： No.102517.
[4]	Xie P， Chen S， Chen Q， et al. GAP-RL： grasps as points for RL towards dynamic object grasping ［J］. IEEE Robotics and Automation Letters， 2025， 10（1）： 40-47.
[5]	Bohg J， Morales A， Asfour T， et al. Data-driven grasp synthesis — a survey ［J］. IEEE Transactions on Robotics， 2014， 30（2）： 289-309.
[6]	Asif U， Tang J， Harrer S. GraspNet： an efficient convolutional neural network for real-time grasp detection for low-powered devices ［C］// IJCAI 2018. California： ijcai.org， 2018： 4875-4882.
[7]	李淦，牛洺第，陈路，等.融合视觉特征增强机制的机器人弱光环境抓取检测［J］.计算机应用， 2023， 43（8）： 2564-2571.
	Li Gan， Niu Mindi， Chen Lu， et al. Robotic grasp detection in low-light environment by incorporating visual feature enhancement mechanism ［J］. Journal of Computer Applications， 2023， 43（8）： 2564-2571.
[8]	陈路，王怀瑶，刘京阳，等.融合空间-傅里叶域信息的机器人低光环境抓取检测［J］.计算机应用， 2025， 45（5）： 1686-1693.
	Chen Lu， Wang Huaiyao， Liu Jingyang， et al. Robotic grasp detection with feature fusion of spatial-Fourier domain information under low-light environments ［J］. Journal of Computer Applications， 2025， 45（5）： 1686-1693.
[9]	Kalashnikov D， Irpan A， Pastor P， et al. Scalable deep reinforcement learning for vision-based robotic manipulation ［C］// CoRL 2018. New York： JMLR.org， 2018： 651-673.
[10]	余家宸，杨晔.基于裁剪近端策略优化算法的软机械臂不规则物体抓取［J］.计算机应用， 2024， 44（11）： 3629-3638.
	Yu Jiachen， Yang Ye. Irregular object grasping by soft robotic arm based on clipped proximal policy optimization algorithm ［J］. Journal of Computer Applications， 2024， 44（11）： 3629-3638.
[11]	Huang B， Yu J， Jain S. EARL： eye-on-hand reinforcement learner for dynamic grasping with active pose estimation ［C］// IROS 2023. Piscataway： IEEE， 2023： 2963-2970.
[12]	Huang Y， Liu D， Liu Z， et al. A novel robotic grasping method for moving objects based on multi-agent deep reinforcement learning ［J］. Robotics and Computer-Integrated Manufacturing， 2024， 86： No.102644.
[13]	Chen P， Lu W. Deep reinforcement learning based moving object grasping ［J］. Information Sciences， 2021， 565： 62-76.
[14]	Morrison D， Corke P， Leitner J. Learning robust， real-time， reactive robotic grasping ［J］. The International Journal of Robotics Research， 2020， 39（2/3）： 183-201.
[15]	Ibarz J， Tan J， Finn C， et al. How to train your robot with deep reinforcement learning： lessons we have learned ［J］. The International Journal of Robotics Research， 2021， 40（4/5）： 698-721.
[16]	Du G， Wang K， Lian S， et al. Vision-based robotic grasping from object localization， object pose estimation to grasp estimation for parallel grippers： a review ［J］. Artificial Intelligence Review， 2021， 54（3）： 1677-1734.
[17]	Liu J， Zhang R， Fang H S， et al. Target-referenced reactive grasping for dynamic objects ［C］// CVPR 2023. Piscataway： IEEE， 2023： 8824-8833.
[18]	Marturi N， Kopicki M， Rastegarpanah A， et al. Dynamic grasp and trajectory planning for moving objects ［J］. Autonomous Robots， 2019， 43（5）： 1241-1256.
[19]	陈佳盼，郑敏华.基于深度强化学习的机器人操作行为研究综述［J］.机器人， 2022， 44（2）： 236-256.
	Chen Jiapan， Zheng Minhua. A survey of robot manipulation behavior research based on deep reinforcement learning ［J］. Robot， 2022， 44（2）： 236-256.
[20]	Wu T， Zhong F， Geng Y， et al. GraspARL： dynamic grasping via adversarial reinforcement learning ［PP/OL］. V2. arXiv （2022-03-14）［2025-01-23］. .
[21]	Christen S， Kocabas M， Aksan E， et al. D-Grasp： physically plausible dynamic grasp synthesis for hand-object interactions ［C］// CVPR 2022. Piscataway： IEEE， 2022： 20545-20554.
[22]	Soviany P， Ionescu R T， Rota P， et al. Curriculum learning： a survey ［J］. International Journal of Computer Vision， 2022， 130（6）： 1526-1565.
[23]	Höfer S， Bekris K， Ankur H， et al. Sim2Real in robotics and automation： applications and challenges ［J］. IEEE Transactions on Automation Science and Engineering， 2021， 18（2）： 398-400.
[24]	Song S， Zeng A， Lee J， et al. Grasping in the wild： learning 6DoF closed-loop grasping from low-cost demonstrations ［J］. IEEE Robotics and Automation Letters， 2020， 5（3）： 4978-4985.
[25]	Zimmermann S， Poranne R， Coros S. Go fetch！ — dynamic grasps using boston dynamics spot with external robotic arm ［C］// ICRA 2021. Piscataway： IEEE， 2021： 4488-4494.
[26]	Tobin J， Fong R， Ray A， et al. Domain randomization for transferring deep neural networks from simulation to the real world ［C］// IROS 2017. Piscataway： IEEE， 2017： 23-30.
[27]	Tobin J， Biewald L， Duan R， et al. Domain randomization and generative models for robotic grasping ［C］// IROS 2018. Piscataway： IEEE， 2018： 3482-3489.
[28]	Muratore F， Ramos F， Turk G， et al. Robot learning from randomized simulations： a review ［J］. Frontiers in Robotics and AI， 2022， 9： No.799893.
[29]	Schulman J， Wolski F， Dhariwal P， et al. Proximal policy optimization algorithms ［PP/OL］. V2. arXiv （2017-08-28）［2025-06-06］. .
[30]	Shakya A K， Pillai G， Chakrabarty S. Reinforcement learning algorithms： a brief survey ［J］. Expert Systems with Applications， 2023， 231： No.120495.
[31]	Schulman J， Moritz P， Levine S， et al. High-dimensional continuous control using generalized advantage estimation ［PP/OL］. V6. arXiv （2018-10-20）［2025-06-17］. .
[32]	Mittal M， Yu C， Yu Q， et al. Orbit： a unified simulation framework for interactive robot learning environments ［J］. IEEE Robotics and Automation Letters， 2023， 8（6）： 3740-3747.
[33]	Serrano-Muñoz A， Chrysostomou D， Bøgh S， et al. skrl： Modular and flexible library for reinforcement learning ［J］. Journal of Machine Learning Research， 2023， 24： 1-9.

层类型	输入通道	输出通道	卷积核	步长	激活函数
卷积层	3或1	16	8×8	4	ReLU
卷积层	16	16	4×4	2	ReLU
卷积层	16	32	3×3	1	ReLU

层类型	输入通道	输出通道	卷积核	步长	激活函数
卷积层	3或1	16	8×8	4	ReLU
卷积层	16	16	4×4	2	ReLU
卷积层	16	32	3×3	1	ReLU

阶段	物体生成位置	物体运动模式	直线运动概率	水平速度范围/（m·s^-1）	加速度惩罚	动作变化惩罚
静态阶段	视野内	静止	0.0	0.00	-0.01	-0.001
低速阶段	视野外	直线	1.0	0.15~0.20	-0.03	-0.003
中速阶段	视野外	直线、正弦	0.8	0.15~0.30	-0.10	-0.010
高速随机	视野外	直线、正弦	0.5	0.15~0.40	-0.30	-0.030

阶段	物体生成位置	物体运动模式	直线运动概率	水平速度范围/（m·s^-1）	加速度惩罚	动作变化惩罚
静态阶段	视野内	静止	0.0	0.00	-0.01	-0.001
低速阶段	视野外	直线	1.0	0.15~0.20	-0.03	-0.003
中速阶段	视野外	直线、正弦	0.8	0.15~0.30	-0.10	-0.010
高速随机	视野外	直线、正弦	0.5	0.15~0.40	-0.30	-0.030

参数名称	随机化方法与范围	对物理实验性能的影响
初始关节位置	在默认姿态上叠加标准差为0.005的零均值高斯噪声	模拟物理机器人任务重置时存在的微小姿态偏差，避免策略依赖固定的初始位置
夹爪初始张合程度	在整个行程范围内均匀采样	模拟物理夹爪在连续操作中可能存在的非完全复位状态，增强策略对执行器状态不确定性的适应能力
夹爪张合速度	在0.1~0.5 rad/s范围内均匀采样	模拟物理夹爪电机性能的波动，使策略不依赖于某个特定的执行器速度
光照强度	在5 000~15 000 lm范围内均匀采样	模拟从阴影到明亮的真实光照条件，使策略学习对光照不敏感的几何与形状特征
桌面纹理	在包含7种不同木质纹理的预设库中随机选择1种	避免策略过拟合仿真环境中单一的桌面外观，提升策略在真实桌面上的泛化能力
物体颜色	在指定RGB空间中随机采样（R：0.7~1.0， G：0.7~1.0， B：0.0~0.2）	针对物理实验中的网球进行颜色随机化，确保策略学习到的是物体的形状和动态特征，而非某一特定的颜色

Robotic end-to-end dynamic grasping method based on curriculum reinforcement learning

基于课程强化学习的机器人端到端动态抓取方法

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 21

References 33

Related Articles 15

Recommended Articles

Metrics

超参数	值	超参数	值
Rollouts	16	Ratio clip	0.2
Learning epochs	6	Value clip	0.2
Mini batches	4	Entropy loss scale	0.01
Discount factor	0.95	Value loss scale	2.0
Lambda	0.95	Kl threshold	0.0
Learning rate	0.000 3	Timesteps	120 000

方法	本文方法	YOLO-RL	GAP-RL
技术路线	端到端	模块化	模块化
核心思想	从原始像素直接学习	目标检测+强化学习	抓取位姿检测+强化学习
信息流	RGB图像和深度图像→动作	RGB图像→目标框→动作	点云→抓取位姿→动作
优点	响应快；无误差累积；能学习隐式特征	概念简单；可利用成熟的目标检测器	泛化性好；可利用成熟的抓取检测器
局限	训练初期探索难度大（本文通过课程学习缓解）；模型可解释性较差	性能受限于检测器；信息延迟与误差累积；目标框信息不足以指导复杂抓取	性能受限于检测器；信息延迟与误差累积；系统复杂度高

物体运动模式	物体速度/（m·s^-1）	尝试次数	成功次数	成功率/%	平均耗时/s
总体	—	40	35	87.5	15.57
静止	0.00	20	19	95.0	12.45
右→左	约0.05	10	9	90.0	17.91
左→右	约0.05	10	7	70.0	20.98

[1]	Xiayu WU, Hong ZHANG. Review of evolution and changes in crowd evacuation calculation methods [J]. Journal of Computer Applications, 2026, 46(4): 1309-1322.
[2]	Tianyu XUE, Aiping LI, Liguo DUAN. Vehicular edge computing scheme with task offloading and resource optimization [J]. Journal of Computer Applications, 2025, 45(6): 1766-1775.
[3]	Pengcheng XU, Lei HE, Chuan LI, Weiqi QIAN, Tun ZHAO. Deep symbolic regression method based on Transformer [J]. Journal of Computer Applications, 2025, 45(5): 1455-1463.
[4]	Huahua WANG, Liang HUANG, Jiajie CHEN, Jiening FANG. Dynamic allocation algorithm for multi-beam subcarriers of low orbit satellites based on deep reinforcement learning [J]. Journal of Computer Applications, 2025, 45(2): 571-577.
[5]	Jing WANG, Xuming FANG. Intelligent joint power and channel allocation algorithm for Wi-Fi7 multi-link integrated communication and sensing [J]. Journal of Computer Applications, 2025, 45(2): 563-570.
[6]	Xiang KUANG, Zhen MA, Wanchun ZHU, Zhi ZHANG, Yunfei CUI. Secure and reliable service function chain deployment based on encoder-decoder structured reinforcement learning [J]. Journal of Computer Applications, 2025, 45(12): 3947-3956.
[7]	Chengyi WANG, Lei XU, Jinyin CHEN, Hongjun QIU. Cyber anti-mapping method based on adaptive perturbation [J]. Journal of Computer Applications, 2025, 45(12): 3896-3908.
[8]	Xiaojuan CHEN, Wei ZHANG. Task allocation of unmanned aerial vehicle for rural last-mile delivery based on reinforcement learning [J]. Journal of Computer Applications, 2025, 45(12): 4055-4063.
[9]	Jun ZENG, Yinghua TONG, Defang WANG. Anomaly detection method based on cumulative probability fluctuation and automated clustering [J]. Journal of Computer Applications, 2025, 45(12): 3864-3871.
[10]	Haoxiang XU, Dunhui YU, Yichen DENG, Kui XIAO. Knowledge graph constrained question answering model based on hierarchical reinforcement learning [J]. Journal of Computer Applications, 2025, 45(12): 3764-3770.
[11]	Lin WEI, Shihao ZHANG, Mengyang HE. Workflow task optimization and energy-efficient offloading method for computing power network [J]. Journal of Computer Applications, 2025, 45(12): 3916-3924.
[12]	Lin WEI, Jinyang LI, Yajie WANG, Mengyang HE. Highly reliable matching method based on multi-dimensional resource measurement and rescheduling in computing power network [J]. Journal of Computer Applications, 2025, 45(11): 3632-3641.
[13]	Shuai ZHOU, Hao FU, Wei LIU. Spatial-temporal Transformer-based hybrid return implicit Q-learning for crowd navigation [J]. Journal of Computer Applications, 2025, 45(11): 3666-3673.
[14]	Yanpeng ZHANG, Yuqian ZHAO, Fan ZHANG, Tenghai QIU, Gui GUI, Lingli YU. Capacitated vehicle routing problem solving method based on improved MAML and GVAE [J]. Journal of Computer Applications, 2025, 45(11): 3642-3648.
[15]	Jinghua ZHAO, Zhu ZHANG, Xiting LYU, Huidan LIN. Multiscale information diffusion prediction model based on hypergraph neural network [J]. Journal of Computer Applications, 2025, 45(11): 3529-3539.