Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Risk-Aware Human-in-the-Loop Framework with Adaptive Intrusion Response for Autonomous Vehicles

Dawood Wasif, Terrence J. Moore, Seunghyun Yoon, Hyuk Lim, Dan Dongseong Kim, Frederica F. Nelson, Jin-Hee Cho

TL;DR

RAIL addresses safety in autonomous driving under rare and adversarial conditions by computing an Intrusion Risk Score from three cues and leveraging shielded action blending to enforce graded control. It couples a contextual bandit for shield selection with SAC trained via risk-prioritized replay, treating takeovers and near misses as informative supervision signals. Empirical results in MetaDrive and CARLA show improved safety, reduced disengagement, and robustness to CAN and LiDAR intrusions compared with RL, safe RL, offline RL, IL, and HITL baselines, while preserving task performance. The framework provides interpretable, online adaptations that can scale to additional cues and attack scenarios, reducing operator burden and enhancing real-world practicality.

Abstract

Autonomous vehicles must remain safe and effective when encountering rare long-tailed scenarios or cyber-physical intrusions during driving. We present RAIL, a risk-aware human-in-the-loop framework that turns heterogeneous runtime signals into calibrated control adaptations and focused learning. RAIL fuses three cues (curvature actuation integrity, time-to-collision proximity, and observation-shift consistency) into an Intrusion Risk Score (IRS) via a weighted Noisy-OR. When IRS exceeds a threshold, actions are blended with a cue-specific shield using a learned authority, while human override remains available; when risk is low, the nominal policy executes. A contextual bandit arbitrates among shields based on the cue vector, improving mitigation choices online. RAIL couples Soft Actor-Critic (SAC) with risk-prioritized replay and dual rewards so that takeovers and near misses steer learning while nominal behavior remains covered. On MetaDrive, RAIL achieves a Test Return (TR) of 360.65, a Test Success Rate (TSR) of 0.85, a Test Safety Violation (TSV) of 0.75, and a Disturbance Rate (DR) of 0.0027, while logging only 29.07 training safety violations, outperforming RL, safe RL, offline/imitation learning, and prior HITL baselines. Under Controller Area Network (CAN) injection and LiDAR spoofing attacks, it improves Success Rate (SR) to 0.68 and 0.80, lowers the Disengagement Rate under Attack (DRA) to 0.37 and 0.03, and reduces the Attack Success Rate (ASR) to 0.34 and 0.11. In CARLA, RAIL attains a TR of 1609.70 and TSR of 0.41 with only 8000 steps.

Risk-Aware Human-in-the-Loop Framework with Adaptive Intrusion Response for Autonomous Vehicles

TL;DR

RAIL addresses safety in autonomous driving under rare and adversarial conditions by computing an Intrusion Risk Score from three cues and leveraging shielded action blending to enforce graded control. It couples a contextual bandit for shield selection with SAC trained via risk-prioritized replay, treating takeovers and near misses as informative supervision signals. Empirical results in MetaDrive and CARLA show improved safety, reduced disengagement, and robustness to CAN and LiDAR intrusions compared with RL, safe RL, offline RL, IL, and HITL baselines, while preserving task performance. The framework provides interpretable, online adaptations that can scale to additional cues and attack scenarios, reducing operator burden and enhancing real-world practicality.

Abstract

Autonomous vehicles must remain safe and effective when encountering rare long-tailed scenarios or cyber-physical intrusions during driving. We present RAIL, a risk-aware human-in-the-loop framework that turns heterogeneous runtime signals into calibrated control adaptations and focused learning. RAIL fuses three cues (curvature actuation integrity, time-to-collision proximity, and observation-shift consistency) into an Intrusion Risk Score (IRS) via a weighted Noisy-OR. When IRS exceeds a threshold, actions are blended with a cue-specific shield using a learned authority, while human override remains available; when risk is low, the nominal policy executes. A contextual bandit arbitrates among shields based on the cue vector, improving mitigation choices online. RAIL couples Soft Actor-Critic (SAC) with risk-prioritized replay and dual rewards so that takeovers and near misses steer learning while nominal behavior remains covered. On MetaDrive, RAIL achieves a Test Return (TR) of 360.65, a Test Success Rate (TSR) of 0.85, a Test Safety Violation (TSV) of 0.75, and a Disturbance Rate (DR) of 0.0027, while logging only 29.07 training safety violations, outperforming RL, safe RL, offline/imitation learning, and prior HITL baselines. Under Controller Area Network (CAN) injection and LiDAR spoofing attacks, it improves Success Rate (SR) to 0.68 and 0.80, lowers the Disengagement Rate under Attack (DRA) to 0.37 and 0.03, and reduces the Attack Success Rate (ASR) to 0.34 and 0.11. In CARLA, RAIL attains a TR of 1609.70 and TSR of 0.41 with only 8000 steps.
Paper Structure (24 sections, 6 equations, 4 figures, 4 tables)

This paper contains 24 sections, 6 equations, 4 figures, 4 tables.

Table of Contents

  1. INTRODUCTION
  2. Related Work
  3. Human-in-the-Loop (HITL) Learning Methods
  4. Risk-Aware Decision Making
  5. Intrusion Detection and Response for AVs
  6. Proposed Approach: RAIL
  7. System Overview
  8. Intrusion Risk Score (IRS)
  9. Curvature uncertainty cue
  10. Time-to-collision cue
  11. Observation shift cue
  12. Risk-Aware Action Blending
  13. Risk-Prioritized Replay and Training Procedure
  14. Experiments and Results
  15. Experimental Setup
  16. ...and 9 more sections

Figures (4)

  • Figure 1: Paradigm of Risk-Aware HITL Learning
  • Figure 2: RAIL System Overview
  • Figure 3: Training Performance Comparison of RAIL, HAIM huang2024human, and HACO li2022efficient on MetaDrive
  • Figure 4: Qualitative Cue-Driven IRS Responses