Testing, Validation, and Verification of Robotic and Autonomous Systems: A Systematic Review

Abstract

We perform a systematic literature review on testing, validation, and verification of robotic and autonomous systems (RAS). The scope of this review covers peer-reviewed research papers proposing, improving or evaluating testing techniques, process, or tools that address the system-level qualities of RAS. Our survey is performed based on a rigorous methodology structured in three phases. First, we made use of a set of 26 seed papers (selected by domain experts) and the SERP-TEST taxonomy to design our search query and (domain-specific) taxonomy. Second, we conducted a search in three academic search engines and applied our inclusion and exclusion criteria to the results. Respectively, we made use of related work and domain specialists (50 academics and 15 industry experts) to validate and refine the search query. As a result, we encountered 10,735 studies, out of which, 195 were included, reviewed and coded. Our objective is to answer four research questions, pertaining to (1) the type of models, (2) measures for system performance and testing adequacy, (3) tools and their availability, and (4) evidence of applicability, particularly in industrial contexts. We analyse the results of our coding to identify strengths and gaps in the domain and present recommendations to researchers and practitioners. Our findings show that variants of temporal logics are most widely used for modelling requirements and properties, while variants of state-machines and transition systems are used widely for modelling system behaviour. Other common models concern epistemic logics for specifying requirements and belief-desire-intention models for specifying system behaviour. Apart from time and epistemics, other aspects captured in models concern probabilities (e.g., for modelling uncertainty) and continuous trajectories (e.g., for modelling vehicle dynamics and kinematics). Many papers lack any rigorous measure of efficiency, effectiveness, or adequacy for their proposed techniques, processes, or tools. Among those that provide a measure of efficiency, effectiveness, or adequacy the majority use domain-agnostic generic measures such as number of failures, size of state-space or verification time were most used. There is a trend in addressing the research gap in this respect by developing domain-specific notions of performance and adequacy. Defining widely-accepted rigorous measures of performance and adequacy for each domain is an identified research gap. In terms of tools, the most widely used tools are well-established model-checkers such as Prism and Uppaal, as well as simulation tools such as Gazebo; Matlab/Simulink is another widely used toolset in this domain. Overall there is very limited evidence of industrial applicability in the papers published in this domain. There is even a gap considering consolidated benchmarks for various types of autonomous systems.

Causal Models to Support Scenario-Based Testing of ADAS

“…As a consequence, the use of Machine Learning (ML)-based subsystems has been increasing [19], creating new challenges for an appropriate safety assessment [20]; in large parts, this is due to their interpretation as black-box elements. A key aspect of any testing endeavor is the identification of a suitable set of test cases [21], [22], [23]. Regardless of the capabilities of the SUT, this requires knowledge about its targeted ODD.…”

Probabilistic Risk Assessment of an Obstacle Detection System for GoA 4 Freight Trains

“…Given an image 𝑝 ∈ 𝑀 0 that is classified by the CNN as "obstacle", so that Λ(𝑝) = 1, all images 𝑝 ′ that can be reached from 𝑝 on a null curve, that is, a piecewise smooth curve of length null in the degenerate metric 𝑔 0 , are also classified by the CNN as obstacles. 4 The obstacle image space O = Λ −1 ({1}) ⊆ 𝑀 0 of all images classified by the CNN as obstacles, however, is not null-connected: for some images 𝑝, 𝑝 ′′ (points) that are both classified as obstacles, every piecewise smooth curve connecting 𝑝 and 𝑝 ′′ traverses one or more regions of points mapped to "no obstacle". Each submanifold of O consisting of pairwise null-connectible points represents an equivalence class, say 𝑐, of the CNN.…”

“…For the application of the CCP in the context of this article, we assume the availability of a large database 𝐷 of 'obstacleon-track' sample images representing the urn in the CCP. We assume that there exists a random selection mechanism 4 The length of differentiable curve 𝛾 in 𝑀 0 is obtained by integrating over the length of its tangent vectors is some curve parametrisation, say, 𝛾 (𝑡 ), 𝑡 ∈ [0, 1] [9]. The length of a tangent vector 𝑣 = • 𝛾 (𝑡 ) is obtained by calculating √︁ 𝑔 0 (𝑣, 𝑣): the metric 𝑔 0 on 𝑀 0 induces a bilinear form (also denoted by 𝑔 0 on the tangent space at 𝛾 (𝑡 ).…”

“…A major obstacle preventing the immediate deployment of autonomous transportation systems in their designated operational environments is their safety assessment. The latter poses several technical challenges [4,20,21], in particular, the trustworthiness of AI-based methods involving ML. As pointed out in ISO 21448 [19], the safety of the intended functionality (SOTIF) is not necessarily ensured by the correctness of the design and its implementation in hardware and software (HW, SW) alone, since the implemented functionality also depends on the training strategy and the training data applied.…”

Trustworthy Autonomous Systems (TAS): The Verifiability Approach

“…Establishing the right level of trust involves gathering and communicating sufficient evidence for the system's safety and usefulness. A holistic validation and verification process is an essential ingredient for providing such evidence (Mousavi et al, 2022;Araujo et al, 2022).…”