This paper, recently published in the arXiv e-print service, tackles a critical challenge for autonomous convoys: how to stay on track when the road itself is invisible.
TL;DR
The authors develop a "course-referenced" control strategy for vehicle convoys that follow GNSS trails rather than lane markings. They prove that following only the vehicle directly in front leads to unavoidable error accumulation and introduce an adaptive feedforward system that corrects for unknown passenger or cargo loads in real-time.
What Problem Does This Paper Address?
Imagine a convoy of autonomous trucks driving through a snowstorm. The lane markings are buried, so the lead vehicle uses a high-precision map or GPS to navigate. Every subsequent vehicle follows the "trail" (the sequence of GPS coordinates) left by the vehicles ahead.
This creates a classic control problem called lateral string instability. If the second vehicle swerves slightly, the third vehicle treats that swerve as part of the "correct" path and might swerve even more. In a long convoy, these tiny errors can amplify until the last vehicle is oscillating wildly or drives off the road.
Furthermore, real-world vehicles have varying loads. A truck carrying heavy cargo corners differently than an empty one. If the control system doesn't know the exact mass and tire stiffness of the vehicle, it will suffer from "steady-state offsets"—basically, it will consistently "under-steer" or "over-steer" on every curve.
Methods
The authors introduce three key innovations to solve these issues:
- Course-Referenced Modeling: Traditional models look at the vehicle's "yaw" (where the nose is pointing). This paper looks at the "course angle" (the direction the vehicle is actually moving). By formulating the math in terms of "arc length" (distance traveled along the road) instead of time, they simplify the complex 2x2 matrix math of error propagation into a single, manageable scalar transfer function.
- Composite Path Fitting: To prevent error growth, followers don't just look at the vehicle directly in front (the predecessor). They fit a path to a weighted combination of the predecessor’s trail and the original leader’s trail.
- Adaptive Curvature Feedforward: The researchers identified that all the "unknowns"—vehicle mass, center of gravity, and tire stiffness—can be condensed into a single coefficient ($\theta_2$) within the steering command. They designed an adaptive law that "learns" this coefficient whenever the vehicle is in a curve, allowing the car to adjust its steering to match its current load.
Datasets
The paper is theoretical and simulation-based. It does not use a standard external dataset but evaluates the proposed controller using a high-fidelity bicycle model simulation of an eleven-vehicle convoy. The simulation parameters (mass, inertia, tire stiffness) were based on previously identified real-world vehicle data.
Key Insights and Ideas
- The "Predecessor-Only" Trap: A major theoretical finding is that even if a controller is "non-amplifying" (meaning it doesn't make the predecessor's wiggle bigger), following only the predecessor still causes errors to accumulate linearly. On a road that shifts sideways, the error will grow with every vehicle in the chain. You must have information from the convoy leader to keep errors bounded.
- Sideslip is a Feature, Not a Bug: By using a course-referenced model, the controller naturally accounts for "sideslip" (when a car drifts slightly outward in a turn). A car cornering perfectly on the path has a non-zero yaw error but a zero course error; the authors' model recognizes this as "correct" behavior, whereas older models might have tried to "fix" it unnecessarily.
- The Power of One Coefficient: It is mathematically elegant that the complex physics of cargo loading and tire wear can be compensated for by estimating just one parameter. This makes the adaptive system robust and computationally efficient.
Key Results
- String Stability: Using a 50/50 mix of leader and predecessor information, the lateral error of the convoy leveled off by the fifth vehicle, staying under 20 cm. Without leader information, the error grew to 86 cm by the tenth vehicle.
- Load Compensation: In a "worst-case" scenario (snowy roads with a 480 kg load), the standard controller left the last vehicle 16 cm off the path during curves. The adaptive law reduced this offset to nearly zero (0.6 cm) within the first few hundred meters of driving.
- Decoupled Dynamics: The authors proved that each vehicle's adaptive estimator works independently. A mistake in the second vehicle's load estimation doesn't "poison" the learning process for the third vehicle.
Limitations and Future Work
- Constant Speed Assumption: The current analysis assumes the convoy maintains a constant speed. Real-world convoys speed up and slow down, which would change the dynamics.
- GNSS Noise: The model assumes relatively clean GPS data. In reality, GPS "drift" or "multipath" errors could introduce new types of string instability.
- Experimental Validation: While the simulations are rigorous, the authors note that the next step is testing this on instrumented physical vehicles to see how the adaptive law handles real-world sensor noise and actuator lag.
Summary updated