Research map / Multi-agent systems and games / Human-AI and goal-conditioned
The Virtual Driving Instructor: Multi-Agent System Collaborating via Knowledge Graph for Scalable Driver Education
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Describes a deployed Virtual Driving Instructor, a hybrid multi-agent, knowledge-graph and Bayesian-network system that gives real-time feedback in driving simulators at a Norwegian driving school, evaluated through student surveys and interviews.
Problem
Traditional driver education requires a human instructor's full attention per student, limiting scalability, standardization, and cost-efficiency.
Approach
A hybrid AI system where over 70 subsumption-layered rule-based agents (plus a CNN gaze-estimation agent) write alerts into a property knowledge graph, which feeds a tutoring agent for real-time feedback, a decision-support interface for 1-to-N teacher monitoring, and a Bayesian-network student model updated after each session.
Key results
- 74% of simulator sessions were rated good or very good in a survey of 316 responses from 87 students.
- 52.2% of students found VDI feedback understandable, and 39.6% partially; under 5% said no.
- Deployed in 5 simulators at 3 locations of Way AS since June 2021; 623 students trained in 2022; 1:3 teacher-student ratio found feasible.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide AI-based instruction and real-time feedback through the VDI | Driver education in simulator · Driving instruction trials | students monitored per instructor: 1:3 | Helped |
| increase the number of students managed by one instructor to 1:5 | VDI teacher-student instruction · Driving instruction trials | instructor manageability: 1:5 | Hurt |
| provide simulator-based VDI instruction | Students using the simulator · Simulator sessions, February–April 2023 | good or very good session rating: 74% | Helped |
| deliver feedback through the VDI | Students using the simulator · VDI feedback survey responses | agreement with feedback: 55% partially agreed; 32% fully agreed; 10% disagreed | Mixed |
| use the eye-tracking system to assess mirror-checking | VDI feedback system · Driving simulator sessions | reliability of mirror-check detection | Hurt |
| give audio feedback rather than relying on dashboard text | VDI tutoring agent · Driving simulator sessions | student feedback preference | Helped |
| display skill progression in the mobile application | VDI student model · Post-session student skill progression | perceived accuracy of displayed data | Mixed |
Surprising findings
- Students sometimes expected negative feedback for mistakes and were surprised when none was given, since the system avoids continuous criticism while still logging errors.
- Students preferred audio feedback over dashboard text, and some never noticed the dashboard feedback.
Limitations
- Evaluation relies on student self-reported surveys and six interviews rather than measured learning outcomes.
- Eye-tracking inconsistency limits reliability of mirror-check detection.
- No evaluation of effect on on-road driving performance yet.
Experimental setup
Datasets: student survey responses (316 responses from 87 students) student interviews (6 interviewees) conditional probability distributions from datasets generated by professional driving instructors
Metrics: student-reported simulator session rating feedback comprehensibility feedback agreement number of students trained
Builds on: subsumption architecture property graph model Bayesian networks
Closest papers by meaning
- A Virtual Driving Instructor That Generates Personalized Driving Lessons Based on Student Skill Level (AAAI 2024 · similarity 0.88)
- DriveAgent-R1: Advancing VLM-based Autonomous Driving with Active Perception and Hybrid Thinking (ICLR 2026 · similarity 0.85)
- MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving (NeurIPS 2026 · similarity 0.84)
- Knowledge Distillation from Single-Task Teachers to Multi-Task Student for End-to-End Autonomous Driving (AAAI 2024 · similarity 0.83)
- Hybrid-Driving: An Autonomous Driving Decision Framework Integrating Large Language Models, Knowledge Graphs and Driving Rules (AAAI 2025 · similarity 0.83)
- World Knowledge-Enhanced Reasoning Using Instruction-Guided Interactor in Autonomous Driving (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.