Design Space of Behaviour Planning for Autonomous Driving

Ilievski, Marko; Sedwards, Sean; Gaurav, Ashish; Balakrishnan, Aravind; Sarkar, Atrisha; Lee, Jaeyoung; Bouchard, F.; De Iaco, Ryan; Czarnecki, Krzysztof · 2019 · OpenAlex-citations

DOI: 10.48550/arxiv.1908.07931

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 complexity of behavior planning (BP) in autonomous driving, specifically focusing on the trade-offs inherent in the design space. The authors argue that design choices optimizing one aspect of BP often critically constrain others, complicating the development process. To clarify these relationships, the study decomposes the design space into three principal axes: environment representation, system architecture, and decision logic representation. This decomposition aims to aid the design process by mapping out qualitative trade-offs and identifying directions for future research. The methodology involves a structured review of current state-of-the-art approaches, organized around the three identified axes. The first axis concerns how the driving environment is represented, ranging from raw sensor data (e.g., lidar point clouds) to feature-based models (e.g., road maps), grid-based occupancy maps, and latent representations (e.g., compressed vectors via autoencoders). The second axis examines the architecture of the motion planner, specifically the level of integration between the behavior planner and the local planner (separated vs. integrated) and the incorporation of prediction modules (explicitly defined external/internal models vs. implicitly defined models). The third axis analyzes the representation of decision logic, categorized into programmed logic (imperative systems like state machines, and declarative systems like expert systems or optimization) and learned logic (learning from examples or interaction via neural networks or reinforcement learning). The findings highlight significant trade-offs across these dimensions. Regarding environment representation, there is a tension between fidelity and computational burden; raw data offers maximum information but requires opaque, end-to-end learning, while abstracted features improve modularity but risk losing critical details. In terms of architecture, separated planners offer simplicity but may suffer from computational redundancy and conflicting solutions, whereas integrated, learned systems improve efficiency but rely heavily on large datasets and lack interpretability. For decision logic, programmed systems provide safety guarantees and interpretability but struggle to scale to complex, noisy real-world scenarios, while learned systems generalize better but are difficult to verify and validate. The paper notes that implicit prediction models, often used in reinforcement learning, eliminate the need for separate prediction modules but sacrifice human interpretability. The significance of this work lies in its systematic mapping of the behavior planning design space, providing a framework for researchers and engineers to navigate the inevitable compromises in autonomous vehicle development. By explicitly linking design choices to their consequences, the paper identifies gaps in current approaches, such as the need for scalable, interpretable learned systems and robust methods for handling noisy, partially observable environments. It serves as a guide for selecting appropriate architectural and logical paradigms based on specific operational requirements, balancing safety, computational efficiency, and adaptability.

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.

• automation surprise