Intelligent Cloud Robotics for Real-Time Manipulation at Scale

Machine learning based robotic manipulation approaches attempt to enable robotic systems to acquire manipulation skills using methods such as deep reinforcement learning. Data-driven approaches have emerged as a predominant paradigm in particular since noise, uncertainty and incomplete sensor data prohibit analytic modelling or perfect simulation of the robot-environment interaction in many scenarios. Unfortunately, however, no general purpose robot exists today that can reliably manipulate the wide variety of objects that humans interact with on a daily basis and conducting data-driven robotic manipulation research is costly and time-consuming since running real-world robotic experiments is error-prone and costly in terms of time and hardware. Research today is typically done on a single robot, resulting in current experimental setups having to rely on only a limited set of real-world training data which is still orders of magnitudes smaller than the amount of sensory data a human child gathers in the first few months after birth. Progress in fields such as computer vision and speech recognition however provides evidence that breakthroughs in machine learning-based methods require scale. In particular machine learning at scale e.g. via deep learning, training data at scale e.g. ImageNet, and infrastructure and parallelization at scale e.g. via cloud-computing.

We argue that robotic manipulation research is still starved for scale, which makes it difficult to obtain statistically reliable and repeatable results and has fundamentally hindered progress in this area. In our view, a concentrated cross-disciplinary Cloud
Robotics effort towards robotic manipulation research is required to overcome these challenges and to unlock the potential of robotic automation beyond today's limited robotic manipulation applications.

To endow robotic systems with capabilities to reliably and safely grasp and manipulate the vast variety of objects encountered in both industrial and everyday home settings, significant interdisciplinary breakthroughs are required before such systems can be deployed. While machine learning relying on large scale image, text and audio data is progressing at rapid pace, robotic manipulation research has in comparison been starved for data since data acquisition on a single robotic system is expensive and slow. This project tackles challenges in this domain by focusing in particular on the emerging paradigm of Cloud Robotics, where robots can communicate over a network to share observations and jointly learn from past observations.

Autonomous robotic systems that can reliably grasp and manipulate objects in their environment in order to perform complex tasks "in the wild" have been the goal of robotic manipulation research for decades, promising breakthrough productivity and safety improvements to industrialized and aging societies with potential applications ranging from industrial automation to robotic health- and elderly care and personal robotic assistants. While approximately 2.7 million industrial robotic systems are estimated to be in operation worldwide as of 2020, these systems currently do not yet benefit from a network effect at scale as they are mostly manually programmed and do not share information or learn collaboratively in a large scale data-driven manner. As a result, the predominant industrial robotic manipulation applications are still very limited to specific tasks in controlled environments such as industrial welding, pick-and place and palletizing tasks.

Recent progress in machine learning applications, particularly in fields such as speech recognition and computer vision, have been powered by a confluence of factors chiefly related to developing systems at scale. In particular, performance
breakthroughs have been enabled by scalable machine learning algorithms, large scale training data collection and processing and cloud computation that enables dynamic resource allocation depending on the machine learning task requirements.

In this interdisciplinary project spanning machine learning, robotics, cloud-computing and real-time control, we focus on achieving a breakthrough in the foundational algorithms, machine learning and system design requirements for scaling networked, distributed, robotic manipulation systems to a large network of cloud-connected robots. We envision this network to continuously collect very large scale manipulation training data and to dynamically learn from past experience using federated machine learning. Our approach will allow robots to balance between centralized machine learning in the cloud and local processing of information using the computational resources of each individual robot while incorporating real-time control and network bandwidth constraints.

Our consortium proposes to pioneer this emerging paradigm of cloud robotics using a robotic demonstrator system featuring a large number of low cost miniature robotic arm systems connected to remote cloud computing resources. The final scale of this system is envisioned to reach approximately 100 low-cost robotic arms that will operate in parallel as well as an initial demonstration of our approach on a smaller set of available higher precision ABB Yumi robots. We will utilize our demonstrator system to study fundamental research challenges, including questions about the foundational data requirements of machine learning methodologies
in the robotic manipulation context, the development of federated machine learning methodologies in a cloud-edge network of robots, resource management, and the development of rigorous large scale experiment design, testing and real-time control strategies for data-driven robotic manipulation at previously unseen levels of network-orchestrated parallelism.