Image: AdobeStock, Gorodenkoff Productions OU
Research project We are combining the latest findings in machine learning, cloud technology and robotics to create new, advanced solutions. It represents a major paradigm shift, with the aim of using cloud-based resources to make robots smarter, more flexible and capable of working in real time. This solution is scalable, which means it can manage and coordinate many robots simultaneously.
Autonomous robotic systems that can grasp and manipulate objects in their environment to perform complex tasks have been the target of research in this area for decades. Such developments would bring new productivity and safety improvements to an industrialised and rapidly ageing IT society. Our focus is on achieving a breakthrough to update the basic algorithms but also to develop new solutions in machine learning and system design requirements for the scaling of networked, distributed robotic manipulation systems to a large network of cloud-connected robots.
It is estimated that more than 3 million industrial robot systems are in operation worldwide. However, these systems do not currently benefit from the network effect on a large scale as they are mostly manually programmed, not sharing information, or learning from each other in a large-scale data-driven manner. As a result, the dominant applications of industrial robot manipulation are still very much limited to specific tasks in controlled environments, such as industrial welding, pick-and-place and palletising tasks.
The rapid progress in machine learning applications, particularly in areas such as speech recognition and computer vision, has been driven by a variety of factors mainly related to developing systems at scale.
In particular, performance breakthroughs have been enabled by scalable machine learning algorithms, large-scale collection and processing of training data, as well as cloud computing. This makes it possible to dynamically allocate resources locally, depending on the requirements of the machine learning task.
Intelligent Cloud Robotics for Real-Time Manipulation at Scale is a collaboration with , Associate Professor, Division of Robotics, Perception and Learning, KTH, (main PI) and , Professor at Department of Atuomatic Control, Lund University.
In this interdisciplinary project, funded with 20 MSEK by , the Wallenberg AI, Autonomous Distributed Systems and Software Programme, we develop new opportunities with a combination of machine learning, robotics, cloud computing and real-time control. Our focus is on achieving a breakthrough by updating the basic algorithms and developing new solutions in machine learning and system design requirements for scaling network-based, distributed robot manipulation systems to a large community of cloud-connected robots.
We envision this network to be able to continuously collect very large-scale training data for manipulation and to dynamically learn from experience using so-called federated machine learning. This approach allows robots to balance between centralised machine learning in the cloud and local processing of information using each robot's computational resources. At the same time, it incorporates real-time control and network bandwidth constraints.
Our consortium intends to pioneer this new paradigm of cloud-based robotics using a robotic system where a large number of cost-effective miniature robotic arm systems are connected to remote cloud-based computing resources. The envisaged scope is about 100 cost-effective robotic arms that will be able to work in parallel. A first demonstration of our approach on a smaller set of available higher-precision ABB Yumi robots will be presented in addition.
We will then use our demonstration system to study fundamental research challenges, including questions on: The fundamental data requirements for machine learning methods in the context of robotic manipulation; the development of federated machine learning methods in a cloud-based network of robots; resource management as well as; the development of rigorous large-scale experiment design and testing and real-time control strategies for data-driven robotic manipulation of network orchestrated parallelism for the very first time.
