Collective Efficient Deep Learning and Networked Control for Multiple Collaborative Robot Systems

In recent years, collaborative robot systems have received a great deal of attention in the context of a large number of industrial applications, such as additive manufacturing, automotive manufacturing, material handling, packaging and co-packing, and quality inspection. The collaboration between multiple collaborative robots (Cobots) and human operators is considered to be the most prominent strategy in Industry 5.0, sharing the same space and collaborating on tasks according to their complementary capabilities. DEEPCOBOT project will investigate the design of a new generation of decentralized data-driven Deep Learning-controllers for multiple coexisting Cobots, which interact both between themselves and with human operators in order to collectively learn from each other's experiences and perform cooperatively different complex tasks in large-scale industrial environments. This is motivated by the increasing demand for automation in industry, especially the demand for a safer and more efficient collaboration between multiple Cobots and human operators to integrate the best of human abilities and robotic automation. The vision of this project is that the learning of the optimal local control policies can be substantially accelerated by sharing both information about previous experiences and computation across multiple neighbor Cobots connected through a wireless communication network, providing solutions that satisfy the necessary real-time constraints in the considered robotic applications, as well as providing sufficient robustness and interchangeability to the control solutions. The proposed project targets fundamental research in the fields of deep learning, reinforcement learning, robotics, control theory, embodied Artificial Intelligence, multi-robot cooperation, human-robot collaboration, and cross-layer network design with significant potential in industrial applications.

This project investigates the design of a new generation of decentralized data-driven Deep Learning based controllers for multiple coexisting (possibly mobile) collaborative robots, which interact both between themselves and with human operators in order to collectively learn from each other's experiences and perform cooperatively different complex tasks in large-scale industrial environments. This is motivated by the increasing demand of automation in industry, especially the demand of a safer and more efficient collaboration between multiple cobots and human operators to integrate the best of human abilities (creativity, adaptivity, interaction) and robotic automation (speed, reliability, precision, inexhaustible task execution capability). Our vision is that the learning of the optimal local control policies can be substantially accelerated by sharing both information about previous experiences and computation across multiple neighbor interconnected cobots, providing solutions that satisfy the necessary real-time constraints in robotic applications, as well as sufficient robustness and interchangeability to the control solutions. The main technical novelties are: a) novel local controllers for the collaborative robots based on distributed online deep learning, where each cobot will use a deep reinforcement learning scheme to learn its local policy of shared control, exploiting not only its local experience, but also experience information from other neighbor cobots that will be connected opportunistically through a time-varying inter-cobot wireless network; b) Rich and fluent cobot-human bi-directional interaction and shared control between cobots and human operators, keeping the required safety constraints; c) novel adaptive distributed computation and networking for information aggregation across cobots. The algorithms will be demonstrated at MIL using real-world collaborative robots in cooperation with the world-leading industrial partners ABB and Omron.