Welcome Professor Triebel!
We are pleased to announce the arrival of our new Guest Professor at the AI4EO Future Lab, Prof. Dr. Rudolph Triebel.
Rudolph Triebel’s main research area are machine learning algorithms applied to mobile robotics tasks, in particular robot perception. He investigates probabilistic reasoning methods for object detection and recognition, and he works on efficient learning methods for classification applications which are particularly suited for mobile robots. Recently, his interests are focussed on unsupervised and online learning methods as well as their efficient implementation on real robotic systems.
Rudolph received his PhD in 2007 from the University of Freiburg in Germany, where he worked in the Autonomous Intelligent Systems group of Prof. Wolfram Burgard. The title of his PhD thesis is “Three-dimensional Perception for Mobile Robots”. From 2007 to 2011, he was a postdoctoral researcher in the Autonomous Systems Lab of ETH Zurich with Prof. Roland Siegwart. There, he worked on machine learning algorithms for robot perception within the EU-projects BACS and EUROPA. Then, from 2011 to 2013 he was a member of the Mobile Robotics Group (MRG) under Prof. Paul Newman at the University of Oxford, where he developed unsupervised and online learning techniques for detection and classification applications in mobile robotics. He was also involved in the navigation system of the RobotCar developed at MRG. Since April 2013, Rudolph works as a Senior Researcher at TU Munich in the Computer Vision Group of Prof. Daniel Cremers. In July 2015, he received the venia legendi in Computer Science. Since November 2015, Rudolph is the Head of the Department of Perception and Cognition at the Institute of Robotics and Mechatronics of the DLR and since January 2022 our new Guest Professor at the AI4EO Future Lab.
Here is a small interview about his experience and plans for the AI4EO lab:
Could you please tell us about your current research topics and interests? Can you explain briefly what they are about?
I investigate computer vision and machine perception algorithms, with major focus on applications in robotics. Typical perception tasks that I focus on within my research include object detection and classification, object pose estimation and tracking, 3D reconstruction and grasp analysis, as well as semantic environment mapping.
Can you explain briefly what they are about?
Common to all these tasks is that there is some visual input, e.g. in form of RGB- or depth- or stereo images, and that we seek for an automatic procedure that analyses this data and makes predictions so that a safe and efficient interaction with the environment can be performed. A typical example is the grasping of a robot hand for a given object, e.g. a cup or a tool, which is at reach, but can be known or unknown to the robot. To achieve this, a number of different sub-tasks must be resolved, including an estimation of the exact location of the object in 3D space, a semantic “understanding” of the object and its use, and a good positioning of the hand at the object’s surface to guarantee a stable grasp. All these tasks are comparably easy for humans, but very difficult for robots.
Why have you decided to focus on robotics and machine learning? What is the most interesting aspect about it for you?
For me, the combination of a theoretic analysis and development of data-driven learning methods on one side and challenging robot perception applications on the other side is a perfect interplay between theory and practice, where one influences and motivates the other one. In robotics, many perception problems are so challenging that they demand for advanced and poweful learning mechanisms, and as such they drive the research also in the theoretic domain. On the other hand, novel theoretic developments from machine learning are often very useful for applications in robotics and can therefore be “grounded”, but also refined and adapted for concrete applications. A recent example of the latter are learning methods from the field of natural language processing, which prove also very useful for visual perception tasks such as segmentation of unknown objects. This challenge, namely making scientific contributions in both the theoretic and the applied domain, is what fascinates me most in my research.
Bluntly said you teach roboters how to see. What are the challenges there?
Visual perception is for the most part a cognitive process. We see what we know, and our previous experiences have a significant impact on the way we interpret and reason about our environment. Therefore, it is important to provide this kind of knowledge also to the robot so that it can perform similarly or even better than the human in these perception tasks. In the earlier days of robot and machine perception, a common approach was to “hard code” this knowledge into the software, e.g. by thoroughly designing feature descriptors that can be re-identified and matched across different images. However, in recent years it has become apparent that the generality of the different cases and environments makes most hand-crafted solutions fail. Instead, the idea now is to provide the knowledge in form of large, diverse, and accurate data sets, that can be fed into a learning algorithm so that the system can automatically be adapted to new environments. Still, many classical, model-based approaches show an extraordinary performance in certain cases due to their ability of abstraction, and the challenge is how to find a good combination of the model-based and the data-driven techniques.
Why is your research important for the future?
It is my conviction that the development of novel technologies can help to mitigate some of our most pressing societal challenges, including climate change, health care, nutrition, and societal dis-balance, and robotics will play a major role here. For many of the problems to come we will need intelligent machines that support humans in their labour. Examples of this include harvesting machines, health care and surgical robots, as well as autonomous vehicles on the road and in the air.
What would you wish to be written about your research in textbooks in some 20 years from now?
20 years is a long time, and no one can predict the future for that long. My wish is that with our uncertainty-aware, efficient, and autonomous machine learning algorithms, we can make a significant contribution to the field and also to the society. But I also think that fame should be a measure and not a target, because we all know that mixing those two up most often hinders progress rather than promoting it.
In your opinion what is the biggest challenge of artificial intelligence nowadays
Most currently deployed machine learning approaches still have the typical “black-box” character, i.e. they receive a large training data set and apply some – simple yet often intransparent – magic to come up with a model that makes predictions for future test data samples. This leads to the problem that a detailed interpretation of the predictions is not possible. For example, questions such as “How reliable is the prediction?”, “Was the training data biased?”, or “What are the reasons that the prediction was made the way it was made?” can not be answered by such systems. However, for a safe and fair use of machine learning in real-world applications, these questions must be resolved. Therefore, we need learning algorithms that provide a more detailed insight about their learning process, e.g. by relating learned model parameters with physical entities so that the result is better interpretable.
How would you like to apply your experiences to the AI4EO future lab projects?
To me, the possibility of bringing researchers from different fields together is an exciting endeavour. The areas of robot perception and earth observation are similar in many aspects, but there are also important differences, e.g. regarding the size, quality, and representation of the available data. This fosters fruitful discussions, both from a methodical and an application-oriented aspect with the potential of common scientific contributions. In particular, my group’s expertise in synthetic training data generation and learning algorithms that provide reliable predictive uncertainties offers a promising potential for common achievements.
What are you looking forward to the most about your new job here?
Getting in contact to nice and smart people from different disciplines and having fun in scientific discussions.
We wish Prof. Triebel a fruitful stay at AI4EO!