Safe and Computational Efficient Imitation Learning for Autonomous Vehicle Driving
DOI: 10.23919/acc45564.2020.9147256
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 safe, computationally efficient, and human-like driving policies for autonomous vehicles. Traditional hierarchical approaches, which separate planning and control, often struggle with real-time constraints and simplified vehicle models, while end-to-end deep learning methods suffer from sample inefficiency, lack of interpretability, and insufficient safety guarantees. To bridge these gaps, the authors propose a "mid-to-mid" imitation learning framework that combines machine learning with model-based control. The system uses a neural network to predict reference trajectories, which are then tracked by a model-based feedback controller to ensure safety and stability. The methodology employs a constrained neural network trained using the Dataset Aggregation (DAgger) algorithm. The expert policy is generated by a safety-critical nonlinear Model Predictive Control (MPC) system, which provides optimal trajectories based on a high-fidelity 15-degree-of-freedom vehicle simulation. A key innovation is the use of B-spline parametrization for the network’s output. Instead of predicting discrete trajectory points, the network outputs B-spline coefficients. This leverages the convex hull property of B-splines, allowing safety constraints (such as lane boundaries) to be enforced directly on the coefficients, ensuring the entire trajectory remains within safe limits. Additionally, barrier functions are incorporated into the training loss function to explicitly penalize violations of safety constraints, rather than relying solely on mean squared error. The study evaluates the approach in a lane-keeping scenario using simulated data. Results demonstrate that B-spline parametrization produces smoother, more comfortable trajectories compared to point-based prediction, which often yields jerky motions. The unconstrained policy network (UPN) initially performs well but degrades during DAgger iterations, eventually causing crashes. In contrast, the constrained policy network (CPN) maintains safety from the first rollout. However, the CPN’s performance worsens after the second iteration due to the rigid nature of the barrier functions. To address this, the authors introduce an Adaptive Constrained Policy Network (ACPN), which decreases the slope of the barrier functions over iterations. The ACPN achieves superior convergence and performance, maintaining a low error margin (on the order of $10^{-3}$) compared to the expert across ten rollouts. The significance of this work lies in its ability to integrate safety guarantees directly into the learning process, improving both the convergence speed and the reliability of imitation learning for autonomous driving. By using B-splines and augmented loss functions, the approach reduces computational complexity and ensures that the learned policy respects hard constraints. The findings suggest that combining model-based safety mechanisms with machine learning can overcome the limitations of pure end-to-end systems, offering a viable path toward safer and more interpretable autonomous driving solutions. Future work includes applying this framework to human driving datasets and more complex scenarios.
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.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | Crossref | — | — | 1 | 2026-06-25 |
| archive | success | unpaywall | — | — | 2 | 2026-06-26 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-26 |
| chunk | success | chunk | — | — | 1 | 2026-06-26 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-26 |
| enrich | success | openalex | — | — | 1 | 2026-06-26 |
| promote | success | — | — | — | 1 | 2026-06-25 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-26 |
| verify | success | — | — | — | 1 | 2026-06-26 |
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.
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).
- Theoretical Contribution: computational model