Disengagement Cause-and-Effect Relationships Extraction Using an NLP Pipeline
DOI: 10.48550/arxiv.2111.03511
URL: https://arxiv.org/abs/2111.03511
archive: archived pipeline: cataloged verified
Summary
Constructs an end-to-end NLP pipeline using deep transfer learning to extract cause-and-effect relationships from California DMV Autonomous Vehicle Disengagement (AVD) reports submitted from 2014 to 2020. The authors built a taxonomy with three main cause categories (AV system, human factors, environmental/other) and analyzed 9,511 disengagement events. Chi-square tests confirmed a significant relationship between AVD initiator and cause category (chi^2(2)=571.53, p<0.001) and between initiator and cause subcategory (chi^2(8)=1726.13, p<0.001). Manufacturers tested most intensively in Spring/Winter; test drivers initiated >80% of disengagements; and >75% of disengagements stemmed from AV-system errors in perception, localization & mapping, planning, or control.
Key finding
Most autonomous-vehicle disengagements (>75%) stem from AV-system errors in perception, localization, planning, or control, and test drivers initiate over 80% of takeovers.
Methodology
Exp 1: 10 participants, repeated measures across 6 sessions from 26 total. Exp 2: 20 participants, Old/New sequence comparison. On-road driving paradigm with DRT and NASA-TLX measures.
Sample size: Exp 1: N=10; Exp 2: N=20
Quality score: 5 / 5