Skip to main content Skip to main navigation

Robots learn with qubits: Bremen scientists develop innovative approach to quantum-based space exploration

| Press release | Autonomous Systems | Machine Learning & Deep Learning | Robotics | Robotics Innovation Center | Bremen

Quantum computers hold unimagined potential for numerous fields of application – including robotics. But research in this area is still in its infancy. In the recently completed QINROS project, researchers from the German Research Center for Artificial Intelligence (DFKI) and the University of Bremen have succeeded for the first time in using reinforcement learning methods with quantum algorithms for robot navigation in the context of space exploration. Their work lays important foundations for research into this future-oriented technology for robotic applications. The project was funded by the German Federal Ministry of Economics and Climate Protection (BMWK).

© DFKI, Florian Cordes
What is the potential of quantum machine learning for robotic navigation in space? In the picture: DFKI robot SherpaTT on an exploration mission during field tests in Morocco in 2018.

The biggest challenge in the field of autonomous robotics is the huge amount of data that must be processed in a very short time so that robots can act autonomously and react quickly to unforeseen situations. Quantum computers are capable of computing a large number of solution paths in parallel, which is why they process information much faster, and could handle much more complex tasks than classical digital computers. However, research into quantum-based computational methods in robotics is still in its very early stages. To advance this, the DFKI Robotics Innovation Center and the robotics group at the University of Bremen have defined a research agenda to develop quantum-based concepts and solutions for application fields in artificial intelligence (AI) and robotics.

Prof. Dr. Dr. h.c.. Frank Kirchner, Director of the DFKI Robotics Innovation Center: "Quantum technology, especially quantum machine learning, has the potential to enable significant developments in the field of efficient computation of highly complex processes. In robotics, we are always working at the limit of computers – the more computing power we have at our disposal, the better. However, there is still a lot of basic research to be done here, as well as corresponding educational work. We would like to contribute to this with current and future projects."

The QINROS project (Quantum Computing and Quantum Machine Learning for Intelligent and Robotic Systems), launched in September 2020, focused on the question of whether and to what extent computational processes for robotic navigation tasks can already be outsourced to quantum computers today. Since the computational capacity of current quantum computers is not yet sufficient for fully quantum-based processing, the distribution of computational processes is of particular relevance. In addition, the Bremen researchers investigated the advantages of quantum-based execution of machine learning and optimization methods over classical methods. They used the example of a mobile Turtlebot system, whose task is to independently explore an unknown environment with the help of reinforcement learning. In the process, the robot collects sensor data from the environment as well as information on its internal state, which serves as the basis for reinforcement learning rewarding desirable robot behavior such as successfully avoiding an obstacle. In this way, the system learns to gradually find its way around the unfamiliar environment.

Based on this scenario, it was first evaluated theoretically which parts of the reinforcement learning can be calculated with the help of quantum-supported methods and which parts should, but also must be preprocessed with classical methods. The implementation was carried out in the simulated environment with increasing complexity. For this purpose, the researchers developed algorithms for parameterizable quantum circuits, which, among other things, enable the calculation of new trajectory targets with qubits. It turned out that with the help of quantum circuits equivalent results as with classical neural networks can be achieved. Moreover, initial results indicate that problems can be represented in a much more compact way: For example, instead of 2000 parameters in the neural network, the scientists needed only 200 parameters in the quantum circuit to solve one and the same problem. The researchers published their results in the paper "Quantum Deep Reinforcement Learning for Robot Navigation Tasks" (arXiv:2202.12180).

Another key component of QINROS was a two-day virtual workshop on February 17 and 18, 2022, which aimed to provide interested individuals from (AI) software development, executives from industry and research, as well as students and doctoral candidates with an initial insight into the complex topic. With a total of 150 participants, consistently positive feedback, and the desire for further training opportunities of this kind, the event was a complete success.

Frank Kirchner: "With QINROS, we have succeeded for the first time in demonstrating quantum machine learning methods for robot behavior in a simulated environment, thus showing the performance of parameterizable quantum circuits in a space robotics use case. We are very pleased with the promising results, which we now plan to pursue. The active participation in our workshop has once again shown us the need and the enormous interest in the topic."

The QINROS project was funded by the German Federal Ministry of Economics and Climate Protection (BMWK) through the German Aerospace Center (DLR) from September 1, 2020, to February 28, 2022.

Further Information:
Paper: Quantum Deep Reinforcement Learning for Robot Navigation Tasks Dirk Heimann, Hans Hohenfeld, Felix Wiebe, Frank Kirchner           
Project page: