Provably Safe Decentralized Contingency MPC under State-Only Information and Limited Sensing for Nonlinear Multi-agent Systems

Imagine a busy intersection where dozens of cars arrive from every direction, each driver knowing only their own destination and the position of the vehicles immediately around them. They cannot talk to one another, cannot see the intentions of drivers further away, and must rely solely on what their own sensors can detect in the moment. In this high-stakes environment, the goal is simple: get everyone to their exit without crashing. Yet, achieving this without communication is one of the hardest problems in robotics and autonomous driving. If a car cannot predict what others will do, it must be prepared for the worst-case scenario at every turn. This is the realm of decentralized control, where each agent acts independently, and the challenge lies in balancing the desire to move efficiently with the absolute necessity of remaining safe.

For years, researchers have tried to solve this by having vehicles share their future paths or by using simple reaction-based rules that avoid immediate collisions but often lead to gridlock or erratic behavior. The problem is that without a shared plan, a car might make a safe move that accidentally blocks a neighbor, creating a chain reaction of stops. To move forward with confidence, a vehicle needs more than just a reaction; it needs a guaranteed backup plan that works no matter what the neighbors do. This new research introduces a method that allows a group of nonlinear, complex vehicles to navigate such a chaotic intersection safely, using only local sensing and without ever needing to exchange data about their future moves.

The core of this work is a strategy called contingency model predictive control. In plain terms, this means that at every single moment, each vehicle calculates two things simultaneously. First, it plans its best possible route to reach its destination, optimizing for speed and smoothness. Second, and just as importantly, it calculates a guaranteed "safety net" maneuver. This safety net is a pre-computed path that leads the vehicle to a complete stop or a safe waiting spot, ensuring that even if everything else goes wrong, the vehicle can always retreat to safety without hitting anyone. The brilliance of this approach is that it separates the goal of moving well from the goal of staying safe, allowing the vehicle to pursue its mission while keeping a verified escape route ready at all times.

The researchers, working with a team at the University of Luebeck in Germany, developed a system where these safety nets are not static but adapt dynamically based on what the vehicle sees. They created a concept called a "safe set," which is essentially a protective bubble around the vehicle. Inside this bubble, the vehicle knows it can always execute its emergency stop. The innovation here is how these bubbles interact. Instead of needing to know the entire history of a neighbor's movements or reconstructing their exact path, each vehicle only needs to know the current position of its neighbors. From this single snapshot, the system generates a new, slightly larger protective bubble that represents the worst-case reach of that neighbor.

When two vehicles approach each other, their protective bubbles might overlap. In the past, this overlap would require complex negotiations or memory of past interactions to resolve. This new method solves that by creating a simple, invisible wall between the two overlapping bubbles. This wall is calculated instantly based on the current positions, effectively splitting the available space so that each vehicle has its own guaranteed zone to operate in. If a vehicle enters a new area and sees a neighbor for the first time, it immediately establishes this boundary. If the neighbor leaves, the boundary disappears. This allows vehicles to join and leave the traffic flow at any time without disrupting the safety of the others, a feature known as "plug-and-play" operation.

To test if this theory holds up in the real world, the team ran extensive computer simulations involving a four-way intersection. They modeled fifty different vehicles, each with the complex physics of a real car, including steering angles and acceleration limits. These vehicles entered the intersection from all sides, choosing to go straight, turn left, or turn right, with no prior agreement on who would go first. The simulation introduced new cars every two seconds and removed them as they finished their journey, creating a constantly changing and unpredictable traffic pattern. The vehicles were equipped with sensors that could only see up to a specific distance, mimicking the limitations of real-world hardware.

The results were striking. Despite the lack of communication and the constant influx of new, unpredictable drivers, every single vehicle successfully navigated the intersection and exited without a collision. The closest any two vehicles came to each other was 1.026 meters, which remained safely above the physical limit of 0.8 meters required to avoid a crash. Throughout the entire experiment, the vehicles maintained their lanes and followed their intended routes, only slowing down or adjusting their paths when the safety system dictated it was necessary. The data showed that the "safety cost" for each vehicle decreased over time, proving that the system was not just avoiding crashes but was also converging toward a stable, efficient state.

This work demonstrates that it is possible to achieve rigorous safety guarantees in complex, nonlinear systems without relying on communication networks that can fail or be jammed. By relying on a memory-free interaction where vehicles react only to the current state of their neighbors, the researchers have created a framework that is robust to plug-and-play changes in the environment. While the current study relies on perfect mathematical models and simulated data, the results provide a strong foundation for future real-world applications. It suggests a future where autonomous vehicles can share a road safely, trusting not in a central controller or a network of radios, but in a shared, mathematically proven understanding of space and safety.