1 January 2026 in Journal of Robotics
Achieving precise robotic manipulation is difficult when objects are unknown or not pretrained, and the environment is unpredictable. Traditional cameras often struggle with depth perception and lag when a robot gets close to its target. To address these limitations, this study explores the design and implementation of an image‐based visual servoing (IBVS) system for a 5‐degree‐of‐freedom (DOF) robotic arm mounted with an eye‐in‐hand camera and sonar sensor. The proposed system enables real‐time object tracking and autonomous grasp using fusion of visual and depth feedback in the Robot Operating System (ROS) 2 and MoveIt Servo frameworks. A Channel and Spatial Reliability Tracker (CSRT)‐based tracking algorithm is applied for precise object localization and consistent performance after a comparative analysis with some other tracking algorithms for object dynamic movement up to 30 rpm. The depth estimation using the sonar sensor provides an average error of 1.2 cm in the 5–30 cm working range to provide accurate reconstruction of the object’s 3D location. The MoveIt Servo controller generates joint‐space trajectories with smooth velocity changes. Experimental trials with 40 grasping tasks had a total success rate of 80% in manipulation, with consistent performance for objects of different sizes and shapes. Results above confirm the correctness of the presented method in providing adaptive robot control and reducing human operators’ manual operation effort. This hybrid feedback approach provides a reliable foundation for real‐time visual servoing, overcoming the latency and precision constraints of traditional vision‐only methods.
8 December 2025 in Journal of Field Robotics
Safe and reliable mobility over different kinds of ground is important for planetary rovers on space missions. Since terrain changes might affect the mobility of the rover, energy consumption, and safety, detecting the type of ground in real‐time is vital. While past methods utilized cameras, motion sensors, or traditional machine learning, achieving consistent and accurate outputs on mixed terrains has been found to be challenging. Here, a wireless sensor module was developed and placed in the rover wheel, and sampled data from a Force‐Sensitive Resistor and an Inertial Measurement Unit at a 100‐samples/s sampling rate. A suite of supervised machine learning algorithms is used for terrain type labeling: rocky, sandy, hard, and grassy. The system could classify terrains in real‐time with good accuracy. Random Forest performed the best among them with 88% accuracy and an F 1 score of 0.87. As a simple and efficient model, Random Forest is chosen for real‐time classification. This sensor‐based system enables real‐time, lighting‐independent terrain classification, overcoming the visual limitations and latency of deep learning image‐based methods. The solution helps the rover plan its movement based on the terrain sensed, making it energy efficient and responsive. Key contributions of this study include the design of a wireless multisensor module, high‐frequency data acquisition, and the integration of machine learning models for rapid terrain‐adaptive mobility.