Introduction
Merging onto a highway from an on-ramp is a high-stakes coordination task. In the US alone, nearly 30,000 merging-related collisions occur annually. While humans navigate these interactions by inferring the "internal states" (intents and strategies) of other drivers, autonomous vehicles (AVs) often struggle with this unobservable logic. Current AI approaches like Deep Reinforcement Learning (DRL) are powerful but often act as "black boxes," lacking the interpretability needed for safety-critical systems. This research introduces a probabilistic framework to automatically extract interpretable, mathematically sound internal states from real-world driving data to improve AV decision-making.
The HMM-GMR Framework
The proposed approach, HMM-GMR, treats the merging task as a dynamic, stochastic process by combining two distinct models:
- Hidden Markov Model (HMM): Handles the temporal aspect, assuming the driving process moves through a sequence of "hidden" states where the next state depends only on the current one.
- Gaussian Mixture Regression (GMR): Handles the spatial aspect, mapping observable inputs to a specific output (the ego vehicle's lateral speed) within each state.
By adding the HMM layer, the model ensures that state transitions are smooth and logically consistent with human driving patterns, avoiding the "jittery" or oscillating transitions seen in standard GMR models.
Variable Selection and Data
The study utilized the INTERACTION dataset, focusing on 789 "courteous" merging events from Chinese highway on-ramps. To ensure the model remains efficient and interpretable, the researchers focused on a specific set of critical variables identified in prior research: speed differences with the lead vehicle ($\Delta v_x^{lead}$), distance to the lag vehicle ($\Delta x^{lag}$), and the ego vehicle's longitudinal speed ($v_x^{ego}$).
Key Insights: The Three States of Merging
Using the Bayesian Information Criterion (BIC), the authors discovered that three internal states are optimal for describing the merge process. These states map to clear semantic phases:
- State 1 (Assessment): Initial interaction with high randomness as the driver assesses the gap.
- State 2 (Adjustment): A regular adjustment phase where the ego vehicle moves slower than the lead vehicle to maintain safety.
- State 3 (Execution): The final phase where the lag vehicle actively increases distance to allow the ego vehicle to merge.
Results and Future Work
The HMM-GMR model achieved a Mean Square Error score ($S_{MSE}$) of 0.686, significantly higher than the non-sequential GMM-GMR baseline (0.485). Furthermore, the activation weights for the internal states remained stable over time, leading to more predictable AV behavior.
Future work will focus on expanding the model to include "aggressive" merging behaviors and integrating this framework into real-time online decision-making planners to evaluate performance in live closed-loop simulations.
Summary updated