Personal details
| Title | [Sweden, with Volvo] Intelligent diagnosis of failure root causes for autonomous mining systems |
| Description | Together with Volvo Autonomous Solutions in Sweden (Gothenburg), we plan to send someone (1 person) from UOL to Sweden to do a "research internship"/"master thesis" on the topic of “Intelligent diagnosis of failure root causes in autonomous mining systems”. The candidate must be an EU citizen, as this work will be conducted in Sweden and there will be visa issues. Volvo is currently planning next year's budget and will work to secure this funding. To be on the safe side, Volvo colleagues and I agreed that the UOL candidate should also apply for the Erasmus+ internship. The candidate may need to assist me with administrative and regulatory matters. The result will always lead to a research paper submission (the result in the 2025 thesis is now under review), and optionally a patent. If you are interested, please send a short CV and your transcript to chih-hong.cheng@uni-oldenburg.de. For selected candidates, Volvo colleagues & I will have a chat with you. Many thanks, and have a great study. |
| Home institution | Department of Computing Science |
| Associated institutions |
|
| Type of work | practical / application-focused |
| Type of thesis | Bachelor's or Master's degree |
| Author | Prof. Dr. Chih-Hong Cheng |
| Status | available |
| Problem statement | In autonomous mining scenario-based testing, engineers must localize the root cause of failed test cases across perception, prediction, planning, and control—quickly and with evidence. The challenge is that failures are buried in vast, multimodal artifacts (sensor/video logs, system traces, KPIs, scenario metadata) and may only manifest under specific environment or fleet states. We seek a method to efficiently isolate the bug by orchestrating: (1) a ReAct-style agent to search and chain reasoning over distributed logs/alerts, trigger targeted queries, and call tools for scenario replay with altered seeds and toggled components; and (2) a VLM to “watch” simulation/video outputs to detect and explain visual anomalies (missed detections, occlusions, lateral drift) and align them with trace events. The system should output a ranked set of root-cause hypotheses with aligned evidence (log snippets, replay diffs, VLM observations) and repro steps. Success is measured by time-to-isolation, diagnosis precision/recall, engineer effort reduction, and ability to generalize across scenarios (dust, night, tight corridors, mixed traffic). Constraints include real-time triage for nightly regressions, uncertain/missing data, and safety-critical explainability. |
| Requirement |
|
| Created | 27/10/25 |
