Learning to Navigate Intersections with Unsupervised Driver Trait Inference

Liu, Shuijing; Chang, Peixin; Chen, Haonan; Chakraborty, Neeloy; Driggs-Campbell, Katherine · 2021 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2109.06783

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 autonomous vehicle navigation through uncontrolled intersections, where vehicles must implicitly negotiate right-of-way without traffic signals. The authors argue that successfully navigating these environments requires inferring the hidden traits of other drivers, such as aggressiveness or cooperativeness, to predict their future behaviors. Existing methods often rely on supervised learning with expensive, noisy trait labels or fail to capture long-term persistent traits, leading to performance degradation when inference errors cascade into the navigation policy. To solve this, the authors propose an unsupervised pipeline that learns a latent representation of driver traits from observed trajectories and uses this representation to train a reinforcement learning (RL) navigation policy. The methodology consists of two stages. First, the authors employ a Variational Autoencoder (VAE) with Gated Recurrent Units (GRUs) to encode vehicle trajectories into a two-dimensional latent space representing driving styles. This network is trained unsupervised using reconstruction loss and KL divergence, requiring no ground truth trait labels. The input features are longitudinal offsets relative to preceding vehicles, derived from simulations using the Intelligent Driver Model (IDM) with conservative and aggressive parameters. Second, the inferred latent traits are fed into a navigation policy network, which uses an attention mechanism to weigh the influence of surrounding vehicles. This policy is trained using Proximal Policy Optimization (PPO) in a T-intersection simulation where the ego vehicle must merge into traffic. The policy is trained directly on the inferred traits rather than ground truth labels, reducing sensitivity to inference errors. Experimental results demonstrate that the proposed unsupervised method effectively distinguishes between conservative and aggressive driving styles. A supervised classifier trained on the learned latent representations achieved 98.08% accuracy, significantly outperforming a baseline method by Morton et al., which achieved only 60.22% accuracy and suffered from model collapse. In navigation tasks, the proposed method achieved success rates within 2% of an oracle policy that had access to true trait labels. It outperformed a supervised baseline by Ma et al., which suffered from severe performance drops due to cascading errors between the trait classifier and the RL policy. Additionally, an ablation study confirmed the importance of the attention mechanism, showing that removing it reduced success rates by 3% to 14%. The significance of this work lies in demonstrating that unsupervised trait inference can provide robust, actionable information for autonomous navigation without requiring labeled data. By training the navigation policy directly on inferred representations, the method mitigates the distribution shift and error accumulation common in modular pipelines. This approach offers a scalable solution for handling complex, interactive driving scenarios where understanding the persistent behavioral traits of other agents is critical for safety and efficiency.

Key finding

An unsupervised variational autoencoder-based method for inferring driver traits from trajectories enables an autonomous vehicle to navigate uncontrolled T-intersections more safely and efficiently than supervised baselines.

Methodology

simulation_modeling

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. Discovered via author_sweep_intake on 2026-05-28.

StageOutcomeToolModelPromptAttemptsCompleted
discover success author_sweep 2 2026-05-28
archive success canonical_url 1 2026-06-04
extract success cached 3 2026-06-10
clean success clean 1 2026-06-04
chunk success chunk 1 2026-06-04
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-04
enrich success 1 2026-05-28
promote success 1 2026-06-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 2 2026-06-10
tag success vector_similarity 15 2026-06-11
verify success 2 2026-06-10

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; 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).