SINADRA: Situation-Aware Dynamic Risk Assessment for Safe and Permissive Autonomous Systems

  • Runtime safety monitors are a key safety measure for autonomous systems, but they are often designed for worst-case behavior of other agents and the environment. This keeps safety arguments simple, yet it can make systems overly conservative, with unnecessary stops, large distances, and reduced throughput. The core problem addressed in this dissertation is how to increase monitor permissiveness without weakening safety justification: when a monitor relaxes an environment assumption, engineers must be able to argue why this relaxation is safe in the relevant hazardous situations and within the stated operational design domain. This dissertation proposes Situation-Aware Dynamic Risk Assessment (SINADRA), an engineering method that extends hazard analysis and risk assessment (HARA) work products so that they support monitorable low-risk situation hypotheses for adaptive constraint relaxation. SINADRA targets risk variability that stems from external agent behavior like human workers in warehouses or human drivers in traffic. The method has three main elements. First, Structured Environment Models (SEM) scope each hazardous situation into an explicit, risk-driven environment view, including initiator scenarios and environment factors that matter for risk. Second, Critical Behavior Cause Trees (CBCT) refine hazardous situations into behavior risk scenarios and relate these scenarios to risk implausibility hypotheses – monitorable situation conditions indicating low risk – and other safety measures from the safety concept like system behavior restrictions. CBCTs form a Boolean model that supports minimal cut set analysis as reviewable verification evidence that, within the modeled scope, risk scenarios are sufficiently blocked by hypotheses and safety measures. Third, SINADRA provides a runtime model transformation that turns argued hypotheses into executable inference components, using either Boolean logic or Bayesian networks (BNs). SINADRA is evaluated in two case studies. A warehouse robotics study executes the full design-time chain, integrates Boolean runtime models into a ROS-based safety monitoring system, and demonstrates hypothesis-driven mode switching with worst-case as the default. An analytic comparison reports estimated corridor travel-time gains of about 2% for passing and 21% for overtaking under stated assumptions. An urban automated-driving study demonstrates component-level runtime feasibility in CARLA by inferring behavior intent likelihoods with BNs and computing an intent-weighted collision-risk signal that differs from a worst-case baseline in uncritical situations. Overall, SINADRA provides a traceable path from HARA artifacts to runtime adaptation logic for safer, situation-dependent permissiveness.

Download full text files

Export metadata

Metadaten
Author:Jan ReichORCiD
URN:urn:nbn:de:hbz:386-kluedo-137215
DOI:https://doi.org/10.26204/KLUEDO/13721
Advisor:Karsten BernsORCiD, Mario TrappORCiD
Document Type:Doctoral Thesis
Cumulative document:No
Language of publication:English
Date of Publication (online):2026/10/14
Year of first Publication:2026
Publishing Institution:Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau
Granting Institution:Rheinland-Pfälzische Technische Universität Kaiserslautern-Landau
Acceptance Date of the Thesis:2026/03/27
Date of the Publication (Server):2026/10/15
GND Keyword:Autonomes System; Risikoanalyse; Funktionssicherheit; Autonomes Fahrzeug; Fahrerassistenzsystem; Künstliche Intelligenz; Mobiler Roboter; Autonomer Roboter
Page Number:XV, 173
Faculties / Organisational entities:Kaiserslautern - Fachbereich Informatik
CCS-Classification (computer science):C. Computer Systems Organization / C.3 SPECIAL-PURPOSE AND APPLICATION-BASED SYSTEMS (J.7)
D. Software / D.2 SOFTWARE ENGINEERING (K.6.3) / D.2.10 Design** (D.2.2) / Methodologies**
I. Computing Methodologies / I.2 ARTIFICIAL INTELLIGENCE / I.2.9 Robotics / Autonomous vehicles (NEW)
I. Computing Methodologies / I.6 SIMULATION AND MODELING (G.3) / I.6.5 Model Development / Modeling methodologies
DDC-Cassification:0 Allgemeines, Informatik, Informationswissenschaft / 004 Informatik
Licence (German):Creative Commons 4.0 - Namensnennung, nicht kommerziell, keine Bearbeitung (CC BY-NC-ND 4.0)