Safe Imitation Learning on Real-Life Highway Data for Human-like Autonomous Driving

Acerbo, Flavia Sofia; Alirczaei, Mohsen; Van Der Auweraer, Herman; Son, Tong Duy · 2021 · Crossref

DOI: 10.1109/itsc48978.2021.9564785

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of developing autonomous driving systems that are both safe and comfortable, aiming to increase user acceptance by mimicking human-like driving behaviors. The authors propose a "safe imitation learning" approach that learns a planning policy from real-life highway driving data. The motivation stems from the limitations of standard behavioral cloning, which suffers from compounding errors and lacks safety guarantees, and end-to-end learning, which is sample-inefficient and prone to catastrophic failures. By incorporating domain-aware safety constraints directly into the training process, the method seeks to ensure risk-free operation while reducing the amount of training data required. The methodology utilizes a neural network policy that outputs coefficients for B-spline trajectories rather than direct control actions. This parametrization leverages the convex hull property of B-splines, allowing the authors to enforce safety constraints via barrier functions in the loss function. These constraints ensure that the trajectory remains within a safe convex hull defined by lane boundaries and collision avoidance limits. The system was trained on data collected from a human-driven Toyota Prius equipped with radar, lidar, and camera sensors on highway roads. The dataset was processed to isolate car-following maneuvers, filtering out lane changes and erratic sensor readings caused by cut-in scenarios. To validate the algorithm without risking real-world deployment, the authors constructed a high-fidelity digital twin using Simcenter Prescan for traffic simulation and Simcenter Amesim for vehicle dynamics modeling. This setup allowed for closed-loop testing that mitigates the domain shift problem between real-world data collection and simulated validation. Experimental results demonstrate that the proposed safe imitation learning significantly outperforms conventional behavioral cloning. In closed-loop simulations, the behavioral cloning agent suffered from compounding errors, causing the vehicle to drift out of its lane and lose track of the leading vehicle within seconds. In contrast, the safe imitation learning agent maintained lane keeping and successfully followed the human driving pattern, albeit with smoothed, averaged dynamics due to the spline parametrization. Safety evaluations across multiple artificial test scenarios revealed that the safe imitation learning agent consistently completed paths while respecting safety constraints, such as maintaining a minimum time-to-collision. Behavioral cloning frequently violated these constraints, resulting in lower path completion rates. The study confirms that integrating barrier functions and B-spline parametrization allows the learning algorithm to remain within safety limits even when encountering states outside the training distribution, without requiring online expert interaction. The significance of this work lies in its demonstration of a robust, sample-efficient method for training autonomous driving policies on real-world data. By formally guaranteeing safety through barrier functions and utilizing a digital twin for validation, the approach addresses critical gaps in current imitation learning techniques. It provides a viable pathway for developing autonomous systems that are not only human-like in behavior but also inherently safe, facilitating easier integration into continuous development loops for Advanced Driver Assistance Systems (ADAS) and autonomous vehicles.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Methodological Resource: tool software, dataset resource
• Theoretical Contribution: computational model