Researchers develop novel dataset to boost robot grasping accuracy

Researchers form the College of Engineering, Shibaura Institute of Technology, Japan and Hanoi University, Vietnam has developed designed dataset aimed at enhancing the performance of 6D pose estimation algorithms. Accurate object pose estimation refers to the ability of a robot to determine both the position and orientation of an object.

Increasingly complex operations require a robot to precisely determine the six degrees of freedom (6D pose) of an objects, position, and orientation, becomes critical. This ability ensures that robots can interact with objects in a reliable and safe manner. However, despite advancements in deep learning, the performance of 6D pose estimation algorithms largely depends on the quality of the data they are trained on.

The dataset addresses a major gap in robotic grasping and automation research by providing a comprehensive resource that allows robots to perform tasks with higher precision and adaptability in real-world environments.

The study was led by Associate Professor Phan Xuan Tan, College of Engineering, Shibaura Institute of Technology, Japan, along with his team of researchers, Dr. Van-Truong Nguyen, Mr. Cong-Duy Do, and Dr. Thanh-Lam Bui from the Hanoi University of Industry, Vietnam, Associate Professor Thai-Viet Dang from Hanoi University of Science and Technology, Vietnam

Associate Professor Tan comments, “Our goal was to create a dataset that not only advances research but also addresses practical challenges in industrial robotic automation. We hope it serves as a valuable resource for researchers and engineers alike.”

The research team created a dataset that not only met the demands of the research community but is also applicable in practical industrial settings. Using the Intel RealSenseTM depth D435 camera, they captured high-quality RGB and depth images, annotating each with 6D pose data rotation and translation of the objects. The dataset features a variety of shapes and sizes, with data augmentation techniques added to ensure its versatility across diverse environmental conditions. This approach makes the dataset highly applicable to a wide range of robotic applications.

“Our dataset was carefully designed to be practical for industries. By including objects with varying shapes and environmental variables, it provides a valuable resource not only for researchers but also for engineers working in fields where robots operate in dynamic and complex conditions,” adds Associate Professor Tan.

The dataset was evaluated using state-of-the-art deep learning models, EfficientPose and FFB6D, achieving accuracy rates of 97.05% and 98.09%, respectively. Thee result proves that the dataset provides reliable and precise pose information, which is crucial for applications such as robotic manipulation, quality control in manufacturing, and autonomous vehicles. The strong performance of these algorithms on the dataset underscores the potential for improving robotic systems that require precision.

Associate Professor Tan states, “While our dataset includes a range of basic shapes like rectangular prisms, trapezoids, and cylinders, expanding it to include more complex and irregular objects would make it more applicable for real-world scenarios”.

“While the Intel RealSense™ Depth D435 camera offers excellent depth and RGB data, the reliance of the dataset on it may limit its accessibility for researchers who do not have access to the same equipment.”

The results clearly demonstrate that a well-designed dataset significantly improves the performance of 6D pose estimation algorithms, allowing robots to perform more complex tasks with higher precision and efficiency.

Looking ahead, the team plans to expand the robot dataset by incorporating a broader variety of objects and automating parts of the data collection process to make it more efficient and accessible. These efforts aim to further enhance the applicability and utility of the dataset, benefiting both researchers and industries that rely on robotic automation.

Image: Researchers from SIT, Japan, developed a novel dataset to enhance robot precision in 6D pose estimation, improving pick-and-place tasks in industrial settings. Credit Phan Xuan Tan from SIT, Japan.

 Paper: https://doi.org/10.1016/j.rineng.2024.103459