A shorter version of this article was published in the monthly magazine Acteurs du franco-allemand, as part of an editorial partnership.
Robots must learn to communicate better if they want to earn their spot in the factory of the future. This will be a necessary step in ensuring the autonomy and flexibility of production systems. This issue is the focus of the German-French Academy for the Industry of the Future’s SCHEIF project. More specifically, researchers must choose appropriate forms of communication technology and determine how to best organize the transmission of information in a complex environment.
“The industry system is monolithic for robots. They are static, and specialized for a single task, but it is impossible for us to change their specialization based on the environment.” This observation was the starting point for the SCHEIF project. SCHEIF, conducted in the framework of the German-French Academy for the Industry of the Future, seeks to allow robots to adapt more easily to function changes. To achieve this, the project brings together researchers from EURECOM, the Technical University of Munich (TUM) and IMT Atlantique. The researchers’ goal is “to create a ‘plug and play’ robot that can be deployed anywhere, easily understand its environment, and quickly interact with humans and other robots,” explains Jérôme Härri, a communications researcher with EURECOM participating in this project.
The robots’ communication capacities are particularly critical in achieving this goal. In order to adapt, they must be able to effectively obtain information. The machines must also be able to communicate their actions to other agents—both humans and robots—in order to integrate into their environment without disruption. Without these aspects, there can be no coordination and therefore no flexibility.
This is precisely one of the major challenges of the SCHEIF project, since the industrial environment imposes numerous constraints on machine communications. They must be fast in the event of an emergency, and flexible enough to prioritize information based on its importance for production chain safety and effectiveness. They must also be reliable, given the sensitivity of the information transmitted. The machines must also be able to communicate over the distances of large factories, not just a few meters. They must combine speed, transmission range, adaptability and security.
Solving the technology puzzle
“The solution cannot be found in a single technology,” Jérôme Härri emphasizes. Sensor technology, for example, like Sigfox and LoRa, which are dedicated to connected objects, have high reliability and a long range, but cannot directly communicate with each other. “There must be a supervisor in charge of the interface, but if it breaks down, it becomes problematic, and this affects the robustness criterion for the communications,” the researcher adds. “Furthermore, this data generally returns to the operator of the network base stations, and the industrialist must subscribe to a service in order to obtain it.”
On the other hand, 4G provides the reliability and range, but not necessarily the speed and adaptability needed for the industry of the future. As for 5G, it provides the required speed and offers the possibility of proprietary systems. This would free industrialists from the need to go through an operator. However, its reliability in an industrial context is still under specification.
Faced with this puzzle, two main approaches emerge. The first is based on increasing the interoperability and speed of sensor technology. The second is based on expanding 5G to meet industrial needs, particularly by providing it with features similar to those of sensor technologies. The researchers chose this second option. “We are improving 5G protocols by examining how to allocate the network’s resources in order to increase reliability and flexibility,” says Jérôme Härri.
To achieve this, the teams of French and German researchers can draw on extensive experience in vehicular communication, which uses 4G and 5G networks to solve transport and mobility issues. The cellular technology used for vehicles has the advantage of featuring a cooperative scheduling specification. This information system feature decides who should communicate a message and at what time. A cooperative scheduler is essential for fleets of vehicles on a highway, just like fleets of robots used in a factory. It ensures that all robots follow the same rules of priority. For example, thanks to the scheduler, information that is urgent for one robot is also urgent for the others, and all the machines can react to free the network from traffic and prioritize this information. “One of our current tasks is to develop a cooperative scheduler for 5G adapted to robots in an industrial context,” explains Jérôme Härri.
Deep learning for added flexibility
Although the machines can rely on a scheduler to know when to communicate, they still must know which rules to follow. The goal of the scheduler is to bring order to the network, to prevent network saturation, for example, and collisions between data packets. However, it cannot determine whether or not to authorize a communication solely by taking communication channel load into account. This approach would mean blindly communicating information: a message would be sent when space is available, without any knowledge of what the other robots will do. Yet in critical networks, the goal is to plan for the medium term in order to guarantee reliability and reaction times. When robots move, the environment changes. It must therefore be possible to predict whether all the robots will start suddenly communicating in a few seconds, or if there will be very few messages.
Deep learning is the tool of choice for teaching networks and machines how to anticipate these types of circumstances. “We let them learn how several moving objects communicate by using mobility datasets. They will then be able to recognize similar situations in their actual use and will know the consequences that can arise in terms of channel quality, or number of messages sent,” the researcher explains. “It is sometimes difficult to ensure learning datasets will match the actual situations the network will face in the future. We must therefore add additional learning on the fly during use. Each decision taken is analyzed. System decisions therefore improve over time.”
The initial results on this use of deep learning to optimize the network have been published by the teams from EURECOM and Technical University of Munich. The researchers have succeeded in organizing communication between autonomous mobile agents in order to prevent the collision of the transmitted data packets. “More importantly, we were able to accomplish this without each robot being notified of whether the others would communicate,” Jérôme Härri adds. “We succeeded in allowing one agent to anticipate when the others will communicate based solely on behavior that preceded communication in the past.”
The researchers intend to pursue their efforts by increasing the complexity of their experiments to make them more like actual situations that occur in industrial contexts. The more agents, the more the behavior becomes erratic and difficult to predict. The challenge is therefore to enable cooperative learning. This would be a further step towards fully autonomous industrial environments.
 SCHEIF is an acronym for Smart Cyber-physical Environments for Industry of the Future.