Abstract: In this paper, we extend a framework that we developed earlier for coordination of connected and automated vehicles (CAVs) at a signal-free intersection to incorporate uncertainty. Using the possibly noisy observations of actual time trajectories and leveraging Gaussian process regression, we learn the bounded confidence intervals for deviations from the nominal trajectories of CAVs online. Incorporating these confidence intervals, we reformulate the trajectory planning as a robust coordination problem, the solution of which guarantees that constraints in the system are satisfied in the presence of bounded deviations from the nominal trajectories. We demonstrate the effectiveness of our extended framework through a numerical simulation.
Introduction
Coordinating Connected and Automated Vehicles (CAVs) at signal-free intersections promises to eliminate stop-and-go traffic, significantly reducing fuel consumption and travel time. Most existing coordination algorithms operate under the assumption of "perfect" execution—where a vehicle is expected to adhere precisely to its planned trajectory.
In real-world scenarios, however, low-level controller errors, sensor noise, and environmental factors cause deviations from these "nominal" trajectories. If two vehicles are scheduled to pass through a conflict point with a narrow safety margin, even a small deviation can lead to a collision. This research addresses the critical gap between idealized deterministic planning and the uncertain reality of physical vehicle dynamics by introducing a robust, learning-based coordination framework.
Methodology
The proposed framework extends a decentralized "single-level" coordination approach by incorporating real-time uncertainty quantification and robust optimization. The process follows three primary stages:
- Learning Uncertainty via GP Regression: As a CAV approaches the intersection, it collects noisy observations of its actual position relative to its planned path. It employs Gaussian Process (GP) regression with a Matern 3/2 kernel to model the error function online.
- Error Propagation and Bounding: The framework maps errors in "time-at-position" to errors in speed and position. Recognizing that speed deviations typically follow a unimodal distribution, the authors utilize the Vysochanskii-Petunin inequality. This allows for the creation of tighter, more reliable 95% confidence intervals than standard methods like Chebyshev’s inequality.
- Robust Reformulation: The coordination problem is reformulated to minimize exit time while satisfying safety constraints for the entire "tube" of possible trajectories defined by the learned confidence intervals.
Simulation and Results
The framework was validated through numerical simulations of a signal-free intersection with 24 CAVs entering from six different paths at a traffic flow of 3600 vehicles/hour. To test the GP's learning capabilities, actual deviations were simulated using a logarithmic function combined with Gaussian noise.
Key Insights
- The "Safety Tube" Concept: Instead of planning a single line, the algorithm plans a trajectory "tube." If the tubes of two vehicles do not intersect, the physical vehicles are statistically guaranteed (at a 95% confidence level) to remain safe.
- Online Replanning: The system includes specific "uncertainty characterization points." When a vehicle reaches such a point, it updates its GP model and broadcasts a refined, more accurate trajectory. Following vehicles then adjust their plans accordingly to maintain safety margins.
- Efficiency vs. Robustness: Simulations demonstrated that the robust framework successfully prevented lateral and rear-end collisions that occurred when using standard deterministic models. Crucially, the system maintained high throughput, proving that safety robustness does not necessitate significant performance degradation.
Conclusion and Future Work
This research provides a viable path for deploying CAVs in uncertain real-world environments. By learning and accounting for trajectory deviations online, the system ensures safety without being overly conservative. Future work will focus on reducing the computational complexity of the non-linear robust optimization solvers and validating the framework on physical scale-model CAV testbeds to account for real-world communication latencies and hardware constraints.
Summary updated