Cognitive Robotics: Intelligent Software Solutions for Adaptive Robotic Systems

Fraunhofer IFF's Department of Robotic Systems develops innovative AI solutions to automate complex processes, such as industrial assembly and disassembly or object manipulation in the health care sector. Our technologies enable robots to respond reliably to changes and adapt to new tasks. In contrast to conventional robotics that are limited to specific, narrowly defined tasks, we are enabling new fields of use that were previously not possible. We make this possible through adaptive, AI-based generation of robot movements during execution. This involves recording the current status of the process and the environment at all times and deriving the optimal action from this information. This opens wholly new fields of use for robotic applications that are now unfeasible with common programming methods. Robots acquire the ability to understand their environment, make decisions, solve problems independently and learn from experience. To do this, we combine findings and methods from different disciplines, such as artificial intelligence, machine learning, control systems engineering and computer vision.

A key priority when developing AI-based robot motions is the generation of requisite training data. Advanced AI systems need a large number of training examples to develop robust and generalizable skills. The volume of data would not only make direct training on physical robots time-consuming and expensive but would also pose the risk of damage caused by malfunction. This is why we use advanced simulation environments in which the robotic system, its environment and the process are simulated. This makes it possible to train AI models efficiently and safely with a virtually unlimited number of scenarios and variants. There is always a certain difference between simulation and reality, the so-called sim2real gap. Our research is consequently concentrated on methods to close this gap. By systematically using simulated variability, we generate robust AI models that can handle these differences. We were able to demonstrate the successful sim2real transfer in numerous practical applications.

In a current research project, we are demonstrating the performance of our AI-based motion generation for robots, taking automated disassembly of end-of-life electronics as an example. This task places particular requirements on the robot controller’s flexibility and adaptability since neither the product being disassembled nor its condition is known in detail in advance. Since this variety makes standard, rules-based programming virtually impossible, we are opting for reinforcement and imitation learning methods that make it possible to learn from experience. The resultant possibilities are demonstrated in the video below, taking the automated removal of a motherboard from a computer’s housing as an example.

Our services:

• Development of AI-based robot motions for processes that cannot be implemented with common programming methods
• Efficient generation of training data in simulations and transfer of trained models to real robotic systems
• Reduction of the work required to integrate robotic systems in highly variable environments or processes
• Increased resilience of robotic applications
• Development of new capabilities to automate hitherto unautomatable applications