Automation Technology
Research
Research
Our goal in field robotics is to increase the autonomy of mobile robots operating in unstructured 3D-environments with as little external infrastructure as possible, e.g. without a given map or a global position given by GPS/GNSS. The SpaceBot Cup organized by the German Aerospace Agency DLR had these constraints and was a perfect test case for our research. The robots had to operate for 60 minutes autonomously on a rugged "planetary" surface, explore and map the environment, find, transport and manipulate two objects, and navigate back to the landing site.
The purpose of the present system is to support people while shopping in the supermarket. For example, the robot provides product information or their position in the market, works as a guide, or as a shopping cart, which follows a customer. However, a major goal of this development is autonomous shopping: with a list of items, the robot collects them efficiently and bring them to the customer or checkout.
Image pixels are the base unit in most image processing tasks. However, they are a consequence of the discrete representation of images and not natural entities. Superpixels are the result of perceptual grouping of pixels, or seen the other way around, the results of an image oversegmentation. Superpixels carry more information than pixels and align better with image edges than rectangular image patches. Superpixel can cause substantial speed-up of subsequent processing since the number of superpixels of an image varies from 25 to 2500, in contrast to hundreds of thousands of pixels. A superpixel segmentation of an input image is illustrated in the left figure. More information on this research subject, Matlab toolboxes for comparing such segmentations and open source superpixel segmentation algorithms can be found here.
Vector Symbolic Architectures (VSA) combine a hypervector space and a set of operations on these vectors. Hypervectors provide powerful and noise-robust representations and VSAs are associated with promising theoretical properties for approaching high-level cognitive tasks. We use them to learn robot behaviours from demonstration.
Circular Convolutional Neural Networks (CCNNs) exploit data with wrap-around structure in 1D (e.g., 360° 2D laserscans), 2D (e.g., panoramic images; projected 360° 3D laserscans), or 3D (e.g., grid cell networks). The circular convolution does not require zero-padding to avoid feature map shrinking due to the kernel size; accordingly, CCNNs prevent that the influence from zeros from padding propagate through the feature maps with an increasing number of layers. We evaluate pros and cons of CCNNs, present how circular convolutional and circular transposed convolutional layers can be implemented, and show how CCNNs can be obtained from available pretrained CNNs without retraining through weight transfer.
The libRSF is an open source C++ library that provides several algorithms that can be applied to sensor fusion problems with non-Gaussian measurement errors. It allows to construct a factor graph with robust error models and solves the corresponding least squares problem by batch or incremental optimization using the Ceres solver.
Our recently published solution to state estimation problems is the adaptive mixture algorithm. We describe the process of robust state estimation as alternating sequence of graph optimization with multimodal error models and Expectation-Maximization (EM). The EM algorithm is applied to adapt the error model to the unknown distribution of the processed measurements.
Non-Gaussian errors are problematic for state of the art sensor fusion algorithms like Kalman filter or factor graphs. We are working on methods to adapt the error distribution inside the fusion algorithm to the properties of the real sensor data. Our recent self-tuning mixtures algorithm was able to improve the estimation results of a real world GNSS application over exiting robust algorithms.
A common challenge for vehicle localization
based on global navigation satellite systems (GNSS) is the
multipath problem when high buildings block the direct line of
sight to one or several satellites. The blocked signals may still
reach the receiver on the ground via one or several reflections on
building structures or the ground. Since the signal path is longer
for the reflected signal, ranging errors occur that can either
prolongate the observed pseudorange or, due to correlation effects,
shorten it. This leads to severely biased position estimates.
We work on methods for mitigating such
effects and apply approaches of robust estimation using graphical
models.
Current state of the art solutions of the SLAM problem are based on
efficient sparse optimization techniques and represent the problem
as probabilistic constraint graphs. For example in pose graphs the
nodes represent poses and the edges between them express spatial
information (e.g. obtained from odometry) and information on loop
closures. The task of constructing the graph is delegated to a
front-end that has access to the available sensor information.
The optimizer, the so called back-end of the system, relies heavily
on the topological correctness of the graph structure and is not
robust against misplaced constraint edges. Especially edges
representing false positive loop closures will lead to the
divergence of current solvers.
We developed a novel
problem formulation that allows the back-end to change parts of
the topological structure of the graph during the optimization
process. The back-end can thereby discard loop closures and
converge towards correct solutions even in the presence of false
positive loop closures. This largely increases the overall
robustness of the SLAM system and closes a gap between the
sensor-driven front-end and the back-end optimizers.
Higher-level navigation algorithms like path planning or mapping are depending on the performance of lower-level algorithms, especially the state estimation. For state estimation usually filter based algorithms like the Extended Kalman Filter (EKF) are used to combine the various sensors in a probabilistic way. However, as this solution has its drawbacks like linearization errors, the handling of delayed measurements or possible inconsistencies, we build our solution upon a factor graph based optimization algorithm.
Recognizing known places in the world is a fundamental capability of autonomous mobile systems. I.e. it is essential for build consistent maps of the environment. We work on different approaches to vision based place recognition. Particular focus is on place recognizing in changing environments.
Changing environments pose a serious problem to current robotic systems aiming at long term operation. While visual navigation systems perform reasonably well in static or low-dynamic environments, severe appearance changes that occur between day and night, between different seasons or different local weather conditions remain a challenge. We aim at adapting the idea of landmark-based navigation for changing environments.
Changing environments pose a serious problem to current robotic systems aiming at long term operation. While place recognition systems perform reasonably well in static or low-dynamic environments, severe appearance changes that occur between day and night, between different seasons or different local weather conditions remain a challenge. We propose to learn to predict the changes in an environment. Our key insight is that the occurring scene changes are in part systematic, repeatable and therefore predictable.
Navigation within a 3D map is a capability which comes along with multiple advantages. First, it provides a global localization which, e.g., is suited for robot swarm tasks as relative poses are immediately known, too. Secondly, since our approach is camera based, simpler and cheaper robots can perform a global localization solely equipped with a camera; again this could be advantageous for swarm robotics applications. Also heterogenous robotic teams can benefit from a camera-based navigation in a 3D map through one robot equipped with a 3D laserscanner which maps the environment and concurrently defines a global coordinate system; so, more robots equipped with a camera can perform their navigation within this global frame.
We present algorithms for a camera-based navigation within a 3D map either merely building upon depth information given a point cloud as map, or building upon additional semantic information given an extruded semantic floor plan as 3D map.
We describe the concept of a mobile robotics course for undergraduate students from an educational point of view in terms of learning goals, experiences, and hardware design. The course as well as the hardware was continuously improved over more than a decade. Hence, we like to describe our motivation and the current structure of the course in order to share our experiences as an inspiration for similar courses.
At the chair of process automation, we have been working on autonomous systems for more than 15 years (UGVs, airship/blimp, automated driving). In 2007 we started working with quadrotors by using 3 AscTec Hummingbirds shortly after they became available. From the beginning, our research has focused on fully autonomous systems in indoor/GPS-denied environments without external sensors or computation. The results of several projects using these MAVs with additional sensors and modifications have been published since 2009, e.g. an autonomous landing procedure by using a camera and a self-made optical flow sensor or in 2010, one of the first autonomous indoor flights using a Kinect on the Pelican.
3D laserscanners are well suited sensors for different perception tasks like navigation and object recognition. However, ready-to-use 3D laserscanners are expensive and offer a low resolution as well as a small field of view. Therefore, many groups design their own 3D laserscanner by rotating a 2D laserscanner. Since this whole process is done frequently, this paper aims at fostering other groups’ future research by offering a list of necessary hardware including an online-accessible mechanical drawing, and available software. As it is possible to align the rotation axis and the 2D laserscanner in many different ways, we present an approach to optimize these orientations. A corresponding Matlab toolbox can be found at our website. The performance of the 3D laserscanner is shown by multiple matched point clouds acquired in outdoor environments.
Autonomous navigation of our rough-terrain rovers
implies the need of a good representation of their near surrounding.
In order to archive this we fuse several of their sensors into one
representation called OctoMap. But moving obstacles can produce
artefacts, leading to untraversable regions. Furthermore, the map
itself is increasing in size while discovering new places. Even though
we are only interested in the near surrounding of the rovers.
Our approach to these problems is the usage of timestamps within the
map. If a certain region was not updated within a given interval, it
will be set to free space or deleted from the map. This first option
is an existing solution and the second option reflects our new
alternative. The proposed approach is provided as open source.
Many of the impressive capabilities of the human visual system surpass all existing technological solutions. However, some of the mechanisms of the biological system are well investigated and robotics may benefit from them. A prominent example is human bottom up visual attention. We focus on implementing the principle mechanisms into technical solutions rather than mimicing the exact biological systems.
The same principles applied for environmental perception in mobile robots, can be used for industrial applications. Automated processing and evaluation of imagery is state of the art in many areas of process automation and quality assurance. Nevertheless, there are many areas for which no solutions exist, particluarly applications with challenging lighting conditions (e.g. endoscope imagery) or rapidly changing requirements (e.g. production of small quantities).