Status Quo: Automated systems will replace the human operator at different tasks in everyday life. From today’s perspective, these new technologies offer predicted but also unknown benefits. However, as every other new technology, also automated systems will have drawbacks for some stakeholders in our society. As long as new technologies are within readiness levels of research, their impact is mostly negligible. The technology readiness level of automated driving in road traffic is pushed forward strongly by many researchers and developers all over the world. Consequently, the demand for safety assurance gets urgent. From today’s perspective, a concept that evaluates the safety of automated driving in an affordable and meaningful way is missing. However, this concept is necessary to enable the introduction of automated driving to public road traffic. Objectives: The objective of this thesis is to improve the understanding of the challenge for safety assurance on automated vehicles. Therefore a concept is aimed for, that estimates the safety impact for the stakeholders of automated driving. Estimations are always based on assumptions and suffer from uncertainty. For that reason the concept needs to consider and express the underlying assumptions and uncertainties. Methodology: The methodology for reaching the objectives is formed around the core assumption of the concept: The safety of an Object under Test can be described by the parameter of a probability distribution. This parameter connects the number of events that result from driving a distance with the safety performance of the OuT. Based on this core assumption a model for safety evaluation is developed iteratively. First of all the relevant stakeholders that are influenced by the technology are identified and analyzed. The second step identifies measurable requirements for the safety of automated vehicles from the stakeholder’s perspectives. Based on this preliminary work on the one hand a usage strategy is defined that controls the introduction of automated vehicles. On the other hand an examination strategy is developed to evaluate whether this strategy enables the automation to meet the requirements. In step four the usage strategy is examined for the Autobahn automation being one representative use case. The results, meaning testing effort and introduction possibilities, are compared and discussed. A refinement of stakeholders as well as requirements is performed. Such a refinement is necessary as only a more precise and subtle analysis will lead to a share between efforts and benefits of the introduction of automated vehicles that forms a basis for the discussion on the safety assurance challenge. Results: The results of the thesis can be grouped into four mayor insights. Firstly, the number of rare events like accidents can be handled as being a product of a random experiment that depends on a safety performance of a traffic participant and the number of driven kilometers. From today’s perspective a falsification of this approach was not found and thus builds a simple first approach. Secondly, the statistical proof of safety based on real-world driving is not economically feasible before mass application of the automated vehicle. Thirdly, refinement of the requirements is necessary and justifiable to reduce the safety requirements. Splitting up the requirements of society and vehicle users leads to reduced testing efforts and an uncertainty-based usage strategy. This uncertainty most likely will reduce during usage, thus also enabling a statistical statement on safety at one point in future. Lastly, a method consisting of evaluation criteria as well as an introduction simulation is developed to examine proposed usage strategies. Thereby the possible safety impacts of the usage are studied. Conclusion: As the safety of automated driving cannot be proven statistically before introduction, the introduction needs to be performed despite and under consideration of an estimated uncertainty. This does not mean that the introduced vehicles are less safe compared to their benchmark; however during introduction it will be uncertain. As long as the uncertainty stays above a threshold a usage strategy that is included into the safety assurance concept is necessary. Such a usage strategy would be cautious and based on regular observation of the events encountered by introduced vehicles. Several challenges have been identified for the developed introduction concept of automated vehicles. Based on these challenges further work should mainly address two topics: 1. The identification and collection of data that is necessary for concept application. 2. The answer of an unavoidable question: How much harm, caused by a human built machine, is acceptable for the exposed humans?
Robotics has the potential to be one of the most revolutionary technologies in human history. The impact of cheap and potentially limitless manpower could have a profound influence on our everyday life and overall onto our society. As envisioned by Iain M. Banks, Asimov and many other science fictions writers, the effects of robotics on our society might lead to the disappearance of physical labor and a generalized increase of the quality of life. However, the large-scale deployment of robots in our society is still far from reality, except perhaps in a few niche markets such as manufacturing. One reason for this limited deployment of robots is that, despite the tremendous advances in the capabilities of the robotic hardware, a similar advance on the control software is still lacking. The use of robots in our everyday life is still hindered by the necessary complexity to manually design and tune the controllers used to execute tasks. As a result, the deployment of robots often requires lengthy and extensive validations based on human expert knowledge, which limit their adaptation capabilities and their widespread diffusion. In the future, in order to truly achieve an ubiquitous robotization of our society, it is necessary to reduce the complexity of deploying new robots in new environments and tasks. The goal of this dissertation is to provide automatic tools based on Machine Learning techniques to simplify and streamline the design of controllers for new tasks. In particular, we here argue that Bayesian modeling is an important tool for automatically learning models from raw data and properly capture the uncertainty of the such models. Automatically learning models however requires the definition of appropriate features used as input for the model. Hence, we present an approach that extend traditional Gaussian process models by jointly learning an appropriate feature representation and the subsequent model. By doing so, we can strongly guide the features representation to be useful for the subsequent prediction task. A first robotics application where the use of Bayesian modeling is beneficial is the accurate learning of complex dynamics models. For highly non-linear robotic systems, such as in presence of contacts, the use of analytical system identification techniques can be challenging and time-consuming, or even intractable. We introduce a new approach for learning inverse dynamics models exploiting artificial tactile sensors. This approach allows to recognize and compensate for the presence of unknown contacts, without requiring a spatial calibration of the tactile sensors. We demonstrate on the humanoid robot iCub that our approach outperforms state-of-the-art analytical models, and when employed in control tasks significantly improves the tracking accuracy. A second robotics application of Bayesian modeling is automatic black-box optimization of the parameters of a controller. When the dynamics of a system cannot be modeled (either out of complexity or due to the lack of a full state representation), it is still possible to solve a task by adapting an existing controller. The approach used in this thesis is Bayesian optimization, which allows to automatically optimize the parameters of the controller for a specific task. We evaluate and compare the performance of Bayesian optimization on a gait optimization task on the dynamic bipedal walker Fox. Our experiments highlight the benefit of this approach by reducing the parameters tuning time from weeks to a single day. In many robotic application, it is however not possible to always define a single straightforward desired objective. More often, multiple conflicting objectives are desirable at the same time, and thus the designer needs to take a decision about the desired trade-off between such objectives (e.g., velocity vs. energy consumption). One framework that is useful to assist in this decision making is the multi-objective optimization framework, and in particular the definition of Pareto optimality. We propose a novel framework that leverages the use of Bayesian modeling to improve the quality of traditional multi-objective optimization approaches, even in low-data regimes. By removing the misleading effects of stochastic noise, the designer is presented with an accurate and continuous Pareto front from which to choose the desired trade-off. Additionally, our framework allows the seamless introduction of multiple robustness metrics which can be considered during the design phase. These contributions allow an unprecedented support to the design process of complex robotic systems in presence of multiple objective, and in particular with regards to robustness. The overall work in this thesis successfully demonstrates on real robots that the complexity of deploying robots to solve new tasks can be greatly reduced trough automatic learning techniques. We believe this is a first step towards a future where robots can be used outside of closely supervised environments, and where a newly deployed robot could quickly and automatically adapt to accomplish the desired tasks.