Advances in legged robots control, perception and learning

Abstract

This is the Institution of Engineering and Technology (IET) Cyber-systems and Robotics Special Issue of Advances in Legged Robots: Control, Perception and Learning.

Legged mammals are found everywhere in nature. These legged animals can reach anywhere on Earth, adapt to uneven, discontinuous and obstructed environment. Their locomotion patterns are flexible and diverse to better adapt to the living environment. Imitating these real animals, legged robots have potential advantages over wheeled or tracked vehicles in regard to the traversal of rough and unstructured terrain. However, there are still many challenges for legged systems in solving the technical problems of the real world.

Control, perception and learning are the key technologies in the field of legged robots. Control is the basis of the stable and flexible locomotion of the legged robot. The combination of control and mechatronics machines will show excellent passing ability in continuous stairs, discrete terrain and vertical obstacle environments. The control methods of legged robots mainly include model-based control and learning-based control. After decades of development, research results have made robots more flexible and stable. The latter is a new method combining artificial intelligence, exploring how robots can acquire new motor skills in the process of interacting with the environment and achieve expected motor abilities. Perception allows the robot to understand the world. Autonomous navigation, behavioural decision-making and task operation, all require environmental awareness and understanding. This ability is an unattainable component of the legged robot. For example, different road surfaces require different gait modes, which is the most direct perceptual demand for legged robots.

Paper 1 by Haochen Xu, paper 2 by Qiuyue Luo and paper 3 by Wenhan Cai investigated the control problems of biped robots, paper 4 by Linqi Ye studied the leg–arm and wheel reconfiguration design and control strategy problems and paper 5 by Jinmian Hou extended the multi-leg hexapod robot problems. The design, control and strategy of the legged robot are discussed.

In paper 1, Haochen Xu et al. studied the disturbance rejection for biped robots during walking and running using CMG. They used the CMG as an auxiliary stabilisation device for fully actuated biped robots and integrated the CMG into the balance control framework. This method can effectively help the robot resist disturbance and remain stable over time.

In paper 2, Qiuyue Luo et al. exploited a self-stabilised walking gait for humanoid robots based on the essential model with internal states. They extended an essential dynamic model to the full dynamic model of humanoid robots based on the zero dynamics concept. By adjusting the step timing and landing position of the swing foot automatically and following the intrinsic dynamic characteristics, the humanoid robot can achieve robust walking.

In paper 3, Wenhan Cai et al. proposed the squat motion of a bipedal robot using RKP and whole-body control. The RKP method considers upcoming reference motion trajectories and combines it with quadratic programming (QP)-based whole body control (WBC). It greatly reduces the computational cost compared to model predictive control with WBC and exhibits high adaptability to rough planning with much less computation time.

In paper 4, Linqi Ye et al. designed a robotic system with legs, wheels and a reconfigurable arm, which takes advantage of the wheel and legs. In this robotic system, a leg–arm reconfiguration design allows the robot to walk, reducing the total weight of the robot, and the multi-task control strategy was described based on variable configuration to complicate environments.

In paper 5, Jinmian Hou et al. introduced a novel heuristic whole-body motion control framework for the heavy-duty hexapod robot to traverse complex terrain. They designed a whole-body motion planning and whole-body torque controller, and virtual model control was used to optimise the ground reaction forces for tracking the pre-planned motion based on single rigid-body dynamics.

Paper 6 by Zhicheng Wang et al. investigated the deep reinforcement learning (DRL) problem, while paper 7 by Chaoyue Xu et al. studied the neural network control problem. Both are essential aspects of learning to achieve control. The DRL approach can help learn robust and variety gaits more efficiently and better terrain adaptation. The neural network can help improve the dynamic performance of non-linear actuators.

In paper 6, Zhicheng Wang et al. proposed the efficient learning of robust quadruped bounding using pretrained neural networks. In their method, they designed a reward function to enforce the gait symmetry and periodicity to improve the bounding performance and learnt feedback controller by simulation where they can build variety of environments for simulation learning. This method has been deployed on the real quadruped robot.

In paper 7, Chaoyue Xu et al. described a new control for a PM bionic legged robot based on the neural network. In their method, a double closed-loop control strategy of the PM bionic leg was designed. Based on the three-element model, a feedforward neuron proportion-integral-derivative controller is designed as the inner control loop, and a sliding mode robust controller with local model approximation is designed using the radial basis function neural network as the outer control loop.

Paper 8 by Guangyu Fan et al. investigated the SLAM problem in dynamic environments, while paper 9 by Jiamin Guo et al. studied autonomous recognition and navigation problems. Both are essential aspects of perception. The SLAM method can help improve the robustness of localization, enabling navigation tasks in dynamic environments. Additionally, the autonomous recognition and navigation method can help understand the environment, make decisions and move to destinations.

In paper 8, Guangyu Fan et al. proposed sampling visual SLAM with a wide-angle camera for legged robots. Their method sampled image blocks with clear texture and enhanced the image details to extract the feature points, and the matched feature points were extracted from the images. These points were selected as the template points, the relationship between the template points and the images was established through the wide-angle camera model, and the pixel coordinates of the template in the images and the descriptors were calculated.

In paper 9, Jiamin Guo et al. researched a quadruped robot with a manipulator to realise leader-following, object recognition, navigation and operation. In this paper, the authors developed a systematic solution including the hardware and software system and system architecture. This can achieve recognition, autonomous navigation and operation on a quadruped robot with a manipulator, and it can be extended to other forms of mobile robots.

All of the papers selected for this special issue indicate trends in the emerging field of control, perception and learning in legged robots. We hope that this special issue will benefit researchers around the world by exchanging and sharing the latest results.