Hierarchical Learned Risk-Aware Planning Framework for Human Driving Modeling

TL;DR

This work addresses the challenge of realistically modeling human driving for safe, scalable autonomous-vehicle (AV) simulation by introducing a hierarchical learned risk-aware planning framework. It combines LSTM-based trajectory prediction with Convolutional Social Pooling to generate multimodal forecasts of surrounding vehicles, and a risk cost that aggregates over predicted futures weighted by probabilities $P_{ik}$ and a planning threshold $H_P$. A planning module uses a Dijkstra-based planner on a risk-aware grid, supplemented by data-driven lane- and speed-related costs learned from NGSIM data to capture diverse driving styles and long-horizon decisions. Validation on real-world data and simulation cases demonstrates the model's ability to reproduce human-like behaviors across highway scenarios and adapt via adjustable risk parameters. The proposed framework thus enables more realistic, diverse, and safety-conscious AV testing by accounting for multimodal future interactions and driver variability.

Abstract

This paper presents a novel approach to modeling human driving behavior, designed for use in evaluating autonomous vehicle control systems in a simulation environments. Our methodology leverages a hierarchical forward-looking, risk-aware estimation framework with learned parameters to generate human-like driving trajectories, accommodating multiple driver levels determined by model parameters. This approach is grounded in multimodal trajectory prediction, using a deep neural network with LSTM-based social pooling to predict the trajectories of surrounding vehicles. These trajectories are used to compute forward-looking risk assessments along the ego vehicle's path, guiding its navigation. Our method aims to replicate human driving behaviors by learning parameters that emulate human decision-making during driving. We ensure that our model exhibits robust generalization capabilities by conducting simulations, employing real-world driving data to validate the accuracy of our approach in modeling human behavior. The results reveal that our model effectively captures human behavior, showcasing its versatility in modeling human drivers in diverse highway scenarios.

Figures (5)

• Figure 1: A visualization of the normalized risk computed using the higher-order rectangular Gaussian function for a single vehicle at one time step. Lower risk areas are visualized in blue, and higher risk areas in yellow.
• Figure 2: The planning grid with nodes and edges for the ego vehicle (pictured in blue) that is used to find a safe path through the surrounding agents.
• Figure 3: A visualization of the risk in each lane relative to the velocity of the vehicle. This is a visualization of data seen in \ref{['tab_learned_parameters']}.
• Figure 4: A demonstration of the model's ability to weave through traffic. The ego vehicle starts behind all vehicles on the road, but with a higher velocity and risk threshold $\mathcal{H}_{P}$ than all surrounding vehicles. The risk-aware planner navigates a safe human-like route to weave through the traffic. https://youtu.be/El7k68taNEY
• Figure 5: A demonstration of the model's ability to move out of the way of an approaching high-speed vehicle to maintain safety. The ego vehicle can be seen starting in front of a high-speed vehicle and traveling at a slower velocity and risk threshold $\mathcal{H}_{P}$ than the approaching vehicle. The planner navigates a safe human-like route to move out of the way. https://youtu.be/SjX5KsTUTF8