Contributions published at Artificial Intelligence Lab

Rolling shutter cameras are present in virtually every mobile device nowadays and even on high-end cameras. Their line-by-line readout of the image sensors, greatly weakens the assumption of instantaneous exposure made by most computer vision algorithms until recent years. Even short exposure rolling shutter images and videos can exhibit considerable visual distortion when acquired in presence of motion, either from the camera itself or from the scene objects. Traditional structure from motion pipelines have been proven to fail under these conditions, notably in its typical refining step, the bundle adjustment. Even after over 50 years of research, bundle adjustment is still the state of the art technique for simultaneous optimal estimation of camera poses and scene 3D structure. It has demonstrated its flexibility and robustness over many different kinds of visual models. Nevertheless, handling the extra freedom of the rolling shutter imagery in presence of diverse levels of noise and outliers can be extremely challenging. In this work, we present a very simple and general linear camera model that allows rolling shutters but, at the same time, constrains it to an usable parametrization. We further investigate the amount of information necessary for each level of noise and propose a weak motion prior to additionally constraint the reconstruction.

Soft materials are not only highly deformable, but they also possess rich and diverse body dynamics. Soft body dynamics exhibit a variety of properties, including nonlinearity, elasticity and potentially infinitely many degrees of freedom. Here, we demonstrate that such soft body dynamics can be employed to conduct certain types of computation. Using body dynamics generated from a soft silicone arm, we show that they can be exploited to emulate functions that require memory and to embed robust closed-loop control into the arm. Our results suggest that soft body dynamics have a short-term memory and can serve as a computational resource. This finding paves the way towards exploiting passive body dynamics for control of a large class of underactuated systems.

This bachelor-thesis discusses the optimization of a kinematic model of the human thumb. Myoelectric hand prostheses have high rejection rates, which increase with the time a prosthesis is worn and can even reach 50% after five years. Often stated reasons for rejection are the prosthesis’ weight, the actuation speed, how durable the device is and how anthropomorphicit looks. The optimization of thumb models with the help of genetic algorithms is presented. It is hypothesized that these optimized models will allow to reduce rejection rate of myoelectric prostheses by providing upper limb amputees with lighter and more functional devices. Based on the results of the here presented research, it can be concluded that genetic algorithms are an effective way to generate thumb models which can perform activities of daily living very well, while at the same time reduce the weight and the material usage of these thumb models.

Today manufacturers lack the possibility to directly interact with their customers. Due to the inability to have direct access to loyalty information from retailers or point-of-sales, manufacturers define own loyalty programs with the main purpose to get a better and more direct path to understand the buying behaviors of their customers. Existing marketing activities are very inefficient, expensive and still do not bring the desired effect of closer and more in-depth knowledge of consumers. Current research and applications have shown that there exists a gap of a more advanced loyalty design. Recent trends demonstrate that including gamification principles lead to improved long-term user retention rates. We have identified this gap and developed a mobile loyalty application that is product-focused, customer-centric and considers gamification principles. The focus of this thesis was to inquire on the impact of a product-based mobile loyalty application. To achieve this goal, we identified three key focal points, concerning the manufacturers, the consumers and the system-design. 1. Manufacturers: First we conducted qualitative surveys amongst important Swiss companies to understand the basic marketing problems and the impact of our approach. 2. System design: Secondly we followed a Design Science Research (DSR) approach, where we designed a product-centric loyalty system including gamification principles. 3. Consumers: The qualitative surveys at the end of the field study led to the understanding of usability and acceptance of the mobile application. Our findings show that a customer-centric, product-focused and gamified mobile loyalty application can lead to increased sales, bring a better understanding of buying behavior of consumers and enhance the customer-manufacturer-relationship. With this research we were able to show how future loyalty systems for manufacturers could be designed. By applying our findings, the proposed solution delivers benefits for manufacturers, retailers and consumers.

This thesis documents the design of an experimental platform based on a quadrotor micro-helicopter. A 2~GHz quad-core CPU board is fitted. The design and assembly of an intermediate printed circuit board which acts as a power supply, a Hi Speed USB hub and a serial interface level shifter is discussed. Furthermore, a state of the art visual reconstruction algorithm is extended to take into account image gradients from captured images. Experiments show improvements in precision and completeness of the reconstructed scene when image gradients are considered. Lastly, experiments are conducted to address aspects of using the method described in this work for next best view planning.

Traditionally, in cognitive science the emphasis is on studying cognition from a computational point of view. Studies in biologically inspired robotics and embodied intelligence, however, provide strong evidence that cognition cannot be analyzed and understood by looking at computational processes alone, but that physical system–environment interaction needs to be taken into account. In this opinion article, we review recent progress in cognitive developmental science and robotics, and expand the notion of embodiment to include soft materials and body morphology in the big picture. We argue that we need to build our understanding of cognition from the bottom up; that is, all the way from how our body is physically constructed.

Interfacing robots with real biological systems is a potential approach to realizing truly adaptive machines, which is a long-standing engineering challenge. Although plants are widely spread and versatile, little attention has been given to creating cybernetic systems incorporating plants. Producing such systems requires two main steps: the acquisition and interpretation of biological signals, and issuing the appropriate stimulation signals for controlling the physiological response of the biological part. We investigate an automated physiological recovery of young avocado plants by realizing a closed interaction loop between the avocado plant and a water-control device. The study considers the two aforementioned steps by reading out postural cues (leaf inclination) and electrophysiological (biopotential) signals from the plant, and controlling the water resource adaptive to the drought condition of an avocado plant. Analysis of the two signals reveals time-frequency patterns of increased power and global synchronization in the narrow bands when water is available, and local synchronization in the broad bands for water shortage. The results indicate the feasibility of interface technologies between plants and machines, and provide preliminary support for achieving adaptive plant-based ‘machines’ based on plants’ large and robust physiological spectrum and machines’ control scheme diversity. We further discuss fundamental impediments hindering the use of living organisms like plants for artificial systems.

null

Human infants show remarkably complex motor behaviors already a few weeks after birth. However, it is still unclear what the principles underlying the development of these behaviors are. One principle which might be involved in the development of some behaviors is the maximization of sensory stimulation in a non-uniform sensor morphology. In this thesis, we test this hypothesis by studying the development of the visual fixation behavior and of the hand-mouth touching behavior. We apply the principle in a simple robotic setup, using retina-like vision and a non-uniform arrangement of tactile sensors as the sensor morphologies. The total sensory activation is used as the reward signal in a reinforcement learning framework based on artificial neural networks. Our experimental results show that both behaviors can be developed from the maximization of sensory stimulation in a non-uniform sensor morphology and that the principle can be applied independently of the sensory modality.

A new class of sensors, inspired by biology, promises significant savings in terms of power consumption and computation. But for these savings to really be meaningful, the systems to which such sensors are connected for data processing, must exhibit low-power consumption too. A new software architecture has to be created to both run efficiently on such systems and take advantage of the event-based nature of these sensors' output. This thesis discusses the details of such an architecture and presents a working implementation, showing that the goals of portability to a wide range of systems and efficient resource usage have been successfully met.

Biological research has concluded that the actuation of the spine contributes significantly to the performance of quadrupeds in terms of controlling body posture, and integrating limbs and trunk actions. Inspired by this biological findings, we develop a pneumatically-driven quadruped robot called Renny with configurable spine morphology to study how the spine contributes to cheetah-like running. Three spine morphologies: rigid spine, passive spine, and actuated spine, are introduced and tested in Renny robot. In addition, we investigate the effect of the stiffness distribution of the spine muscles in the passive case. The experimental results show that the passive one where the dorsal stiffness is higher than the ventral stiffness can run faster, even faster than the rigid case. Moreover, the coordination between the legs and the actuated spine is studied in actuated spine morphology. We found that when the spinal movements are synchronized with the legs movements, the speed is much faster. In the actuated case, both flexion and extension benefit the increase of the speed by advancing limbs rapidly and increasing the limb swing.

null

null

In human babies there is currently a rather clear picture of what behaviours appear at what stages of development (eg. reaching, walking). However, information about how these behaviours are brought to life is much more scarse. Currently my supervisor is working on a hypothesis where these behaviors are the result of a single rule: maximization of receptor activity. In his hypothesis he expects to produce behaviours entailed by different sensor modalities: vision, audition, and tactile.