More and more industrial businesses are adopting trained autonomous robots with the goal to improve worker safety and reduce operational costs. These inspection robots need a high degree of autonomy, which requires operational decision-making. How robots learn to make these assessments in real-time is a complex process based on computer vision. ANYbotics, with the assistance of Digica, focusses on refining this process to optimize the value that autonomous robots bring to our end-user customers.
Computer Vision Enables Safe and Automated Industrial Inspection
Globally, a growing number of industrial site managers are adopting autonomous robots to perform routine inspection missions. The goals are to improve worker safety, automate repetitive tasks, and reduce operational costs. Using artificial intelligence (AI) to learn, ANYmal, ANYbotics’ legged robot, autonomously performs inspection tasks. The main areas where AI plays a crucial role are locomotion, navigation, and inspection intelligence. The focus of this article is on the latter.
ANYmal’s inspection capabilities include thermography, gas detection, visual and acoustic measurement. ANYbotics developed an algorithm, i.e. a tool, that accurately captures digital information from industrial objects like manometers, square and linear gauges (e.g., thermometers), digital and analog counters, LCD displays, and valves. Using cutting-edge computer vision, ANYmal learns how to self-determine and self-regulate data acquisition quality, make calculated adjustments to its own actions, and repetitively, and autonomously, extract information from industrial objects. The result is consistent data at the highest level of accuracy, retrieved without operator intervention.
ANYmal capturing and verifying digital information from industrial gauges.
Synthetic Data Enables Efficient Learning
Together with Digica, a leading provider of AI-powered software solutions, ANYbotics developed its intelligent visual inspection framework. The framework was refined using synthetic data like computer-generated images based on real images. Like this, ANYmal learns how to interact with inspection objects from different angles and lighting conditions with less real data by generating thousands of synthetic images in real-time.
With Digica’s insight and expertise, we are able to improve our object recognition intelligence. High-quality data sets always help, but Digica helped us learn and apply how to be less dependent on real image data for training object detection. Dr. Robert Mackenzie, CTO Solutions, ANYbotics.
Programming New Inspection Tasks Made Simple
ANYmal’s inspection intelligence extends far beyond how objects are deciphered and includes how new inspection tasks can be programmed easily after an initial mission is set up.
ANYbotics uses a practical balance between automated and ‘one-time human input’ approaches to teaching ANYmal exactly what and how to inspect. Operators simply guide ANYmal and create a 3D digital twin route model including inspection points, terrain plotting, task sequences, and path planning. With a simple gesture, ANYmal can automatically capture an object’s data and convey the information to an operator. Operators can adjust or confirm this information via a smart device when finalizing the inspection mission set-up. After this initial set up, ANYmal will carry out the execution of its missions autonomously.
ANYmal using a detailed digital twin model to simulate, plan, and commission autonomous robotic inspection missions.
A New Era of Robotic Intelligence for Smarter, Faster Industrial Inspection
ANYmal robots have superhuman abilities: they can see and record temperature, have perfect memory and can store vast amounts of data, they can precisely and reliably perform an exact set of actions repeatedly, and accurately measure their environment via dozens of sensors and share that sensor input with human coworkers (i.e. multi-modal sensing).
These capabilities are shaping a new era in industrial inspection. An era where intelligent, reliable, mobile and autonomous, legged robots, such as ANYmal, only need to be taught what and how to inspect – once!