Article
Chemistry, Multidisciplinary
Liang Yang, Guanyu Lai, Yong Chen, Zhihui Guo
Summary: This paper proposes a new online walking controller for biped robots, which integrates a neural-network estimator and an incremental learning mechanism to improve control performance in dynamic environment. An interval type-2 fuzzy weight identifier is developed to address the imbalanced distribution problem of training data. The effectiveness of the control scheme is verified through full-dynamics simulation and practical robot experiment.
APPLIED SCIENCES-BASEL
(2021)
Article
Automation & Control Systems
Qiang Huang, Chencheng Dong, Zhangguo Yu, Xuechao Chen, Qingqing Li, Huanzhong Chen, Huaxin Liu
Summary: Compliance control is crucial for disturbance absorption in biped robots, but it can cause balance deterioration due to the robot's floating base nature. To address this issue, we propose a strategy called resistant compliance, inspired by how humans resist external disturbance by reconciling their posture and pushing back. This strategy allows the robot to initially comply with the disturbance and then repel it to reduce imbalance caused by motion adjustments, resulting in improved stability and human-like reactions in locomotion and interactions.
IEEE-ASME TRANSACTIONS ON MECHATRONICS
(2022)
Article
Automation & Control Systems
Qiang Huang, Chencheng Dong, Zhangguo Yu, Xuechao Chen, Qingqing Li, Huanzhong Chen, Huaxin Liu
Summary: Compliance control is crucial for biped robots to maintain balance under disturbance, and the proposed resistant compliance strategy allows robots to react more like typical humans, improving environmental stability and safety.
IEEE-ASME TRANSACTIONS ON MECHATRONICS
(2022)
Article
Engineering, Mechanical
Zebang Pan, Shan Yin, Guilin Wen, Zhao Tan
Summary: Designing a high-performance controller for biped robots' walking gaits is a research area that is still open due to their nonlinearity and non-smooth responses. To overcome these challenges, a humanoid robot with a torso is developed first, followed by the adoption of the twin delayed deep deterministic policy gradient algorithm to design the reinforcement learning controller. A reward function utilizing the Poincare map and the power function is constructed for the specified control targets, providing guidelines for the controller. The proposed controller can adaptively output accurate cosine torques and achieve the goal without relying on pre-designed reference trajectories or unstable periodic gaits.
ACTA MECHANICA SINICA
(2023)
Article
Engineering, Multidisciplinary
Koray K. Safak, Turgut Batuhan Baturalp, Selim Bozkurt
Summary: This study proposes a design approach and develops a low-power planar biped robot named YU-Bibot. The robot's kinematic structure consists of six independently driven axes with a low power-to-weight ratio of 30 W/kg. It mimics the natural human walking gait and features spring-supported knee and ankle joints.
Article
Computer Science, Artificial Intelligence
Weiyi Zhang, Yancao Jiang, Fasih Ud Din Farrukh, Chun Zhang, Debing Zhang, Guangqi Wang
Summary: A gait controlling framework based on reinforcement learning is proposed, leveraging the prior knowledge of reference motion to train the robot. The trained agent outperforms traditional methods and achieves improved performance through optimization in a finely crafted RL environment. The proposed method is validated in different tasks and environments, demonstrating its performance and robustness.
PEERJ COMPUTER SCIENCE
(2022)
Article
Computer Science, Information Systems
Jakub Bernat, Pawel Czopek, Szymon Bartosik
Summary: This study presents a Deep Reinforcement Learning algorithm for controlling a differentially driven mobile robot, and investigates the effectiveness of different environment definitions on the learning process.
Article
Computer Science, Software Engineering
Hwangpil Park, Ri Yu, Yoonsang Lee, Kyungho Lee, Jehee Lee
Summary: The primary goal of this study is to address questions regarding the robustness of deep control policies compared with human walking, and to evaluate the effectiveness of different variants of DRL algorithms.
Article
Chemistry, Analytical
Zemin Cui, Yaxian Xin, Shuyun Liu, Xuewen Rong, Yibin Li
Summary: This article proposes a decoupled control framework to improve the balance and dynamic locomotion capabilities of a wheeled biped robot (WBR). By decoupling the WBR into a wheeled inverted pendulum and a multi-rigid body system, time-varying linear quadratic regulators and a model predictive controller are designed. The Kalman filter is also used to optimize the estimation of system state, enabling the WBR to achieve various functions.
Article
Chemistry, Multidisciplinary
Jaeuk Cho, Jong Hyeon Park
Summary: This paper proposes a method for online motion control of a running biped robot on an uneven terrain based on a dual linear inverted pendulum model (D-LIPM) and hierarchical control. The method generates the trajectory of the center of mass (COM) using linear model predictive control (MPC) and generates the angular motions of the robot using quadratic-problem (QP) based momentum control, ensuring stable bipedal running on uneven terrains.
APPLIED SCIENCES-BASEL
(2022)
Article
Robotics
Julian Ibarz, Jie Tan, Chelsea Finn, Mrinal Kalakrishnan, Peter Pastor, Sergey Levine
Summary: Deep reinforcement learning has shown promise in enabling physical robots to learn complex skills in the real world, which presents numerous challenges in perception and movement. Real-world robotics provides a unique domain for evaluating deep RL algorithms, addressing challenges that are often overlooked in mainstream RL research.
INTERNATIONAL JOURNAL OF ROBOTICS RESEARCH
(2021)
Article
Computer Science, Artificial Intelligence
Saeed Saeedvand, Hanjaya Mandala, Jacky Baltes
Summary: The research introduces a novel hierarchical deep learning algorithm that learns how to drag heavy objects with an adult-sized humanoid robot for the first time. The algorithm utilizes a Three-layered Convolution Volumetric Network for 3D object classification, a lightweight real-time instance segmentation method for floor surface detection and classification, and a deep Q-learning algorithm for policy control of the robot's Center of Mass.
APPLIED SOFT COMPUTING
(2021)
Article
Computer Science, Information Systems
Tengteng Zhang, Hongwei Mo
Summary: This paper presents a robotic grasp architecture based on attention-based deep reinforcement learning. Prominent characteristics of input images are automatically extracted using a full convolutional network to prevent the loss of local information. Unlike previous model-based and data-driven methods, the reward is remodeled to address sparse rewards. Experimental results show that our method can double the learning speed in grasping randomly placed objects. In real-world experiments, the grasping success rate of the robot platform reaches 90.4%, outperforming several baselines.
Article
Engineering, Electrical & Electronic
Landong Hou, Bin Li, Weilong Liu, Yiming Xu, Shuhui Yang, Xuewen Rong
Summary: This paper modifies the SRB model and proposes a DRL-based MPC method to resist disturbances caused by swinging legs, improving the robustness of the model.
Article
Robotics
Julian Whitman, Matthew Travers, Howie Choset
Summary: Modular robots can be rearranged into new designs for different tasks, but each design requires its own control policy. To address this, we propose a modular policy framework that creates a policy conditioned on the hardware arrangement, allowing one training process to control various designs.
IEEE TRANSACTIONS ON ROBOTICS
(2023)
Article
Chemistry, Multidisciplinary
Xinyang Jiang, Xuechao Chen, Zhangguo Yu, Weimin Zhang, Libo Meng, Qiang Huang
APPLIED SCIENCES-BASEL
(2018)
Article
Chemistry, Multidisciplinary
Tianqi Yang, Weimin Zhang, Xuechao Chen, Zhangguo Yu, Libo Meng, Qiang Huang
APPLIED SCIENCES-BASEL
(2018)
Article
Automation & Control Systems
Runming Zhang, Libo Meng, Zhangguo Yu, Xuechao Chen, Huaxin Liu, Qiang Huang
Summary: This article proposes a method of integrating center of mass stabilization control into the optimization of stepping variables to achieve robust walking under perturbations. The method involves offline and online procedures, with benchmark coefficients and transition matrices derived in the offline procedure. In the online procedure, real-time prediction of subsequent step states and optimization of stepping variables are achieved using model predictive control. The combination of stepping adjustment and CoM stabilization control is demonstrated to have advantages.
IEEE-ASME TRANSACTIONS ON MECHATRONICS
(2022)
Article
Engineering, Biomedical
Lei Wang, Libo Meng, Ru Kang, Botao Liu, Sai Gu, Zhihao Zhang, Fei Meng, Aiguo Ming
Summary: This paper presents the design of a parallel quadrupedal robot capable of versatile dynamic locomotion and perception-less terrain adaptation. The robot is implemented with symmetric legs and a powerful actuator for highly dynamic movement. A fast and reliable method based on generalized least square is proposed to estimate the terrain parameters. The optimal foot force for terrain adaptation is achieved using virtual model control with the quadratic program method. Simulation and experiments demonstrate the robot's robust and versatile dynamic locomotion on uneven terrain, proving the effectiveness and robustness of the proposed method.
CYBORG AND BIONIC SYSTEMS
(2022)