TL;DR
ImIn-GAIL is a novel framework that learns realistic driving behaviors from sparse, real-world trajectory data by integrating trajectory interpolation directly into the imitation learning process. By treating interpolation and policy learning as a joint task rather than separate steps, the model significantly outperforms traditional calibration and standard imitation learning methods in simulating realistic traffic.
What Problem Does This Paper Address?
To build reliable traffic simulators for testing autonomous vehicles or traffic signal policies, we need "driver models" that behave like real humans. Traditionally, researchers used rule-based car-following models, but these are often too simplistic. Imitation Learning (IL)—specifically Generative Adversarial Imitation Learning (GAIL)—has emerged as a powerful alternative that learns behavior directly from "expert" (real-world) trajectories.
However, there is a major catch: standard IL requires dense data (high-frequency GPS pings). In the real world, trajectory data is often sparse. For example, a car might only be "observed" when it passes through intersections equipped with cameras, leaving long unobserved gaps in between. Directly applying IL to sparse data causes the model to fail at understanding what happens in those gaps. While one could interpolate the data first and then run IL, this "two-step" approach often introduces errors because the interpolation doesn't account for the underlying driving logic.
Methods
The authors propose ImIn-GAIL (Imitation-Interpolation GAIL), which merges interpolation and imitation into a single end-to-end pipeline. The architecture consists of three main parts:
- Generator (The Policy): A neural network ($\pi_\theta$) that acts as the "driver" in a simulator (CityFlow). It observes the state (speed, distance to lead car) and decides on an action (acceleration).
- Downsampler: This component takes the dense trajectories produced by the generator and purposefully makes them sparse. This creates a "synthetic sparse" dataset where the ground truth (the original dense path) is known.
- Interpolator-Discriminator: This is the core innovation.
- The Interpolator learns to fill in the gaps of sparse trajectories. It uses the data from the Downsampler to practice, effectively learning "how to guess the path" based on the generator's current behavior.
- The Discriminator then takes these newly "densified" expert trajectories and compares them against the generator's trajectories.
By training these together, the interpolator becomes better at reconstructing expert paths, which in turn provides a much cleaner and more accurate "reward signal" for the generator to learn from.
Datasets
The researchers evaluated their framework on four distinct environments:
- Synthetic - Ring: A simple circular road to test basic car-following logic.
- Synthetic - Intersection: A 4-way intersection with bi-directional traffic.
- Real-world - $LA_{1 \times 4}$: Data from Lankershim Boulevard in Los Angeles, featuring four successive intersections.
- Real-world - $HZ_{4 \times 4}$: A larger $4 \times 4$ grid from the Gudang area in Hangzhou, China, captured via surveillance cameras.
In all cases, the authors took dense ground-truth data and sampled it at low frequencies to simulate the "sparsity" challenge found in real-world infrastructure.
Key Insights and Ideas
- The Synergy of Joint Learning: The "aha moment" is that interpolation and imitation are two sides of the same coin. A model that knows how a human drives (the policy) is naturally better at guessing where a human was during an unobserved gap (the interpolation).
- Avoiding the "Two-Step" Trap: Traditional two-step methods (interpolate then learn) fail because the initial interpolation is "blind" to the simulator's physics. ImIn-GAIL ensures the interpolated expert paths are physically plausible within the simulation environment.
- Self-Supervision via the Generator: The generator serves as a source of "labeled data" for the interpolator. Because the generator's full path is known, the system can train the interpolator to reconstruct sparse versions of those paths without needing extra real-world labels.
Key Results
ImIn-GAIL consistently outperformed all baselines, including Behavioral Cloning (BC), standard GAIL, and calibrated rule-based models (CFM).
- Accuracy: In the Hangzhou ($HZ_{4 \times 4}$) dataset, ImIn-GAIL reduced the Root Mean Squared Error (RMSE) of vehicle positions to 0.130 km, compared to 0.196 km for standard GAIL—a nearly 34% improvement.
- Travel Time: The error in predicted travel time was also significantly lower. In the Ring scenario, ImIn-GAIL achieved an RMSE of 16.97s, while standard GAIL lagged at 42.06s.
- Robustness to Sparsity: The authors tested various sampling rates (from 2% to 100%). While performance naturally improves with more data, ImIn-GAIL maintained a much higher performance floor at extreme sparsity levels compared to other methods.
Limitations and Future Work
- Longitudinal Focus: The paper primarily focuses on car-following behavior (acceleration/deceleration). While critical, a full simulator also needs to handle lateral behaviors like lane changing and complex turns, which were not the primary focus here.
- Computational Overhead: Training a dual-purpose Interpolator-Discriminator alongside a Reinforcement Learning generator is computationally more intensive than standard GAIL.
- Multi-Agent Complexity: While tested in multi-vehicle environments, the "credit assignment" problem in extremely dense, chaotic urban grids (where one car's interpolation depends heavily on five others) remains a frontier for future scaling.
Summary updated