For a specified operational design domain (ODD) automated driving systems (ADS) will be capable to take over the complete driving task. The vast number of environmental conditions that ADS needs to handle must be systematically integrated into the development process. We present in this paper our approach to solve this integration using the ODD as single point of knowledge. The paper covers: 1) Requirement analysis of the need for ODD data in subsequent development processes. 2) Description of the modelling approach for an ODD implementation. 3) Providing an ODD toolchain covering the functional needs of multiple organization units. 4) Present a method for validating defined ODD against measurement data. We tested our concept by using the toolchain on sample data to create and compare ODD. Our results suggest using this framework supports a more adequate requirement definition and scenario generation.

We are concerned with analysis of Automated Driving Systems (ADS) behavior against formal specifications of desired guardrails. We present methods for determining whether observed behaviors are consistent with formal requirements. This paper is intended to be used as a component within, or otherwise inform, the development of a BigData system continuously monitoring a large fleet of diverse vehicles. Our results include: 1) Definition of behaviors and their relationships to scenarios. 2) Definition of guardrails to be validated and their relationship to ambiguity. 3) Reduce guardrail validation to simple fuzzy constraint satisfaction. 4) Determine degree of behavior compliance with guardrails. 5) Enable validation against partially specified guardrails and situations describing a small fraction of the world. 6) Enable support for uncertainty specifications. 7) Support decomposition into modules aligned with organizational structures. 8) Provide a low cost scalable solution for assessment of compliance, coverage and acceptable risk. The complexity of validation is linear in the number of requirements and situations extracted from logs.

The development of Automated Driving Systems (ADSs) is driven by the many benefits they offer. However, the complexities associated with ADSs and their interactions with the environment pose challenges for their safety assurance. A key aspect during its development process is knowing the capabilities, limitations, and being able to convey them in a clear manner for various types of stakeholders. The Operational Design Domain (ODD) concept was introduced to define the operating boundaries where a system can operate safely. It is therefore a key element for the safety assurance of ADSs. Efforts have been made to define the scope and the content an ODD for ADSs should cover, however there remains the need for a common, exchangeable, executable, and human-readable format for the description. This paper presents a language for the description of the ODD of ADSs, in a textual format that leans on natural language influence. Such format is intended to be both human and machine-readable and would be relevant to end users such as regulators and systems designers. The two-level abstraction approach – a structured natural language representation and a formal representation (covered across two papers) -- has been developed to have the ability to describe complex ODD conditionalities and utilize a well-defined domain ontology to achieve rich semantics. It is aimed to support ODD related activities throughout the development cycle of ADSs (specification as well as verification and validation), while covering a diverse range of stakeholders.

A formal representation for the Operational Design Domain (ODD) of Automated Driving Systems (ADSs) is presented in this paper. An ODD specification determines for every situation whether it is included or excluded from the ODD. We focus on methods to provide unambiguous specification in a programmatic format which are simultaneously machine and human readable. We present a logical framework with intuitive and well-defined semantics which directly supports safety engineering process through specification stage to defining uncertainty and acceptable risk. Its rich and diverse feature set include 1) parsimonious permissive and restrictive constraints, 2) the ability to import OWL ontologies, 3) ability to bind to non-uniform complex variable length structures within situation data, and 4) ability for components to control the scope of usage by integrators. The presentation of syntax and formal semantics is illustrated with example demonstrating key concepts and language capabilities. \n

We consider workloads reducible to membership checks against predicates over continuous variables, for which a scan is required. We explore trading-off storage of re-usable components to avoid repeated computation for each entry of a full scan. Our method renders effective the storage of reusable results in smaller faster memory. Upon receipt of data, a build step produces a compressed representation. Subsequently, upon receipt of a query, a compilation step constructs lookup tables which are used to determine membership, with "one-sided" error guarantees; when misses occur, the full predicate evaluation is performed. We mitigate the exponential complexity by providing a recursive decomposition generating compressed representation reusable across numerous queries, useful e.g. for quantile membership checks. We experiment with a number of knobs, distributions and selectivity rates. We develop an analytic performance model that relies on metrics measurable on a workload sample, specify analytically the conditions for which speedups are achievable, and demonstrate concordance with the empirical evaluation. The key advantages of the proposed methods are: (1) Unlocking the value of GPU massive parallelism; (2) Lossy compression which enables loading into memory a much larger number of entries as compared to using the raw data; (3) The size and speed are superior to Bloom Filters with similar miss-rates; and (4) Predictable query latency regardless of the width of the raw data or predicate computation cost or workload distribution or selectivity.

Whereas autonomous driver engineering has limited control on outcomes for individual scenarios, engineering processes must exert control over the performance statistics. A key challenge impeding statistical confidence is the need to test for a long list of infrequent hazardous events. Using "smart miles" promises to reduce the evaluation cost by increasing the frequency of those infrequent hazardous events. We propose a simulation based Bayesian approach which increases the frequency of those infrequent events by many orders of magnitudes as compared to naturalistic miles. We represent the Operational Design Domain (ODD) using a population of scenarios, and provide methods for sampling the ODD. We propose a quantile function based sampling approach which is able to generate a single sample with thousands of instances using a single forward pass of a deep neural network (DNN). We develop a practical sampler training method for an "inverted DNN" architecture. The resulting sampler is capable of generating a skewed distribution comprising of "smart miles" in which hazardous events of interest occur at a frequency >95%. To gauge the quality of the generated sample we propose the quality metrics of histogram diversity, histogram homoscedasticity and average sample distance.

We are concerned with the task of learning the model of a system given some abstract understanding of its behavior encoded in the form of a family functions. The approaches of both system identification and supervised learning are concerned with such tasks. We propose a unified perspective leading to the Particle Swarm Regression (PSR) algorithm. This algorithm boasts, among other features, interpretable results that identifies a symbolic relationship between independent and dependent variables. In particular, the PSR produces a model comprising of a primary predictor supported by numerous correctors. PSR is capable of fitting families of functions to multi-dimensional data. It supports polynomials, transcendental functions, logarithmic functions, as well as combinations thereof. It can approximate non-convex data using families of convex functions, including posynomials. The effectiveness of PSR is demonstrated using a damped oscillator and benchmark F16 vibration data.