Willi Richert

Increasing the Autonomy of Mobile Robots by On-line Learning Simultaneously at Different Levels of Abstraction

We present a framework that is able to handle system and environmental changes by learning autonomously at different levels of abstraction. It is able to do so in continuous and noisy environments by 1) an active strategy learning module that uses reinforcement learning and 2) a dynamically adapting skill module that proactively explores the robots own action capabilities and thereby providing actions to the strategy module. We present results that show the feasibility of simultaneously learning low-level skills and high-level strategies in order to reach a goal while reacting to disturbances like hardware damages. Thereby, the robot drastically increases its overall autonomy.

Combining Autonomous Exploration, Goal-Oriented Coordination and Task Allocation in Multi-UAV Scenarios

Successful rescue operations after big accidents or natural disasters require a fast and efficient overview of the overall situation. With recent advances, unmanned aerial vehicles (UAVs) are more and more a viable choice under such circumstances.With the number of employed UAVs, the problem of coordination arises as well as proper task allocation among possibly heterogeneous UAVs. This paper presents a hybrid approach for UAV coordination and covers the exploration of unknown terrains as well as goal-oriented coordination and simultaneous task allocation. The approach combines the simplicity of the gradient method with informed A* search and supports prioritized task assignment. The system is suited for highly dynamic environments requiring frequent path recalculations.

The Paderkicker Team: Autonomy in Realtime Environments

The Paderkickers are a robot soccer team that makes heavy use of automotive technology like C167 micro-controllers or communication over CAN bus. All sensor data is processed on these decentralized embedded nodes to yield a high degree of reliability and hardware layer abstraction. In this paper, we describe how the complex system copes with perception and action in real-time and integrates it in the higher strategy layer to achieve autonomous behavior.

Learning action sequences through imitation in behavior based architectures

In this paper a new architecture for learning action sequences through imitation is proposed. Imitation occurs by means of observing and applying sequences of basic behaviors. When an agent has observed another agent and applied the observed action sequence later on, this imitated action sequence can be seen as a meme. Agents that behave similarly can therefore be grouped by their typical behavioral patterns. This paper thus explores imitation from the view of memetic proliferation. Combining imitation learning with meme theory we show by simulating agent societies that with imitation significant performance improvements can be achieved. The performance is quantified by using an entropy measure to qualitatively evaluating the emerging clusters. Our approach is demonstrated by the example of a society of emotion driven agents that imitate each other to reach pleasant emotional state.

Adaptivity at every layer: a modular approach for evolving societies of learning autonomous systems

We describe a developmental architecture that enables individual robots to fulfill tasks assigned to the robot society in a robust, decentralized manner. The architecture is meant to show emergent properties according to Organic Computing principles that are positive for the societys robustness and performance. This requires the architecture to feature those adaptation and learning processes that are not only selfishly useful for the individual robot, but also incorporate the robot societys actual needs at all layers.

Layered understanding for sporadic imitation in a multi-robot scenario

With imitation robots have a powerful means to drastically cut down the exploration space. However, as existing imitation approaches usually require repetitive demonstrations of the skill to learn in order to be useful, those are typically not applicable in groups of robots. In these scenarios usually each robot has its own task to accomplish and should not be disturbed by teaching others. Therefor, most of the time an imitating robot has only one observed performance of the behavior from which it can learn. Utilisation of these sparse observation data has largely been ignored. We present an approach that allows an individually learning robot to make use of such cases of sporadic imitation which is often the only possibility to learn from other robots in a group. The power of the algorithm comes from the fact that it uses the robots already known skills and strategies to understand the observed behavior. Thereby, a robot can use imitation in order to guide its exploration efforts towards more rewarding areas in the exploration space. This is inspired by imitation often found in nature where animals or humans try to map observations into their own capability space. We show the feasibility by realistic simulation of Pioneer robots.

Towards Robust Layered Learning

In his landmark work introducing layered learning Stone presented a new way of handling complex application domains suitable especially for mobile robots. We extend his framework by introducing robust layered learning- a framework that is able to handle system and environmental changes at every layer. We present first results of a lower level implementation of such a framework for mobile robots and discuss how all available sources of information regarding unforeseen changes can be integrated in such a framework in order to reach maximal robustness.

Increasing Learning Speed by Imitation in Multi-robot Societies

The paradigm of imitation provides a powerful means for increasing the overall learning speed in a group of robots. While separately exploring the environment in order to learn how to behave with respect to a pre-defined goal, a robot gathers experience based on its own actions and interactions with the surroundings, respectively. By accumulating additional experience via observing the behaviour of other robots, the learning process can be significantly improved in terms of speed and quality. Within this article we present an approach, that enables robots in a multi-robot society to imitate any other available robot without imposing unnecessary restrictions regarding the robots’ design. Therefore, it benefits not only from its own actions, but also from actions that an observed robot performs. In order to realise the imitation paradigm, we solve three main challenges, namely enabling a robot to decide whom and when to imitate, to interpret and thereby understand the behaviour of an observed robot, and to integrate the experience gathered by observation into its individual learning process.

Optic-tactile robotics and medical applications

This paper gives an overview of new developed hardware and software framework for fiber-optical constructional sensory applied to robotics and medical applications. Different sensor types are shown and used together on an robotic arm and are also applied to human spine movement detection. The sensory bases on fiber-optical ribbons. After presenting the different sensors and both application areas the overall system is shown beginning with real-time measurement, processing and calibration, visualization and analysis of the movements. The control loop is closed to the actuators of the robot application. The correlation between different sensor segments will be mapped onto a Petri-net topology to handle interactions. The software suite which puts all parts together will be outlined afterwards.

Measurement of Robot Similarity to Determine the Best Demonstrator for Imitation in a Group of Heterogeneous Robots

Imitation is not only a powerful means to drastically downsize the exploration space when learning behavior. It also helps to align the learning efforts of a robot group towards a common goal. However, one prerequisite in imitation, the decision of which robot to imitate, is often factored out in current research.