Trust-Aware Sequential Decision Making and Rollout Planning for Resilient Multi-Robot Systems
This paper addresses resilient multi-robot routing under GPS spoofing by introducing a trust-aware framework that combines a distance-constrained adversary model with a probabilistic monitor to detect and remove compromised agents, thereby restoring planner-execution consistency and enabling effective rollout-based decision-making.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a city where hundreds of self-driving cars and delivery drones work together to move people and packages. They rely on a central computer to decide who goes where, a process that requires every vehicle to report its location truthfully and then actually drive to the assigned destination. This system works beautifully when everyone cooperates, but it faces a unique vulnerability: what happens when a vehicle lies about where it is, or simply refuses to do the work it was told to do? In the real world, this isn't just a glitch; it is a strategic attack. A malicious actor could trick the system into sending a driver to a fake location, wasting time and leaving real customers waiting. The danger is that the central computer, believing the lie, might keep sending more drivers to that same fake spot, creating a traffic jam of wasted effort that spirals out of control.
Researchers at Harvard University and Johns Hopkins University have developed a new way to protect these systems, not by trying to stop every lie, but by changing how the computer thinks about the drivers it trusts. Their work focuses on a specific type of deception called "localization spoofing," where an attacker fakes their GPS signal. The team discovered that a clever attacker doesn't need to lie wildly to cause chaos; they only need to lie just enough to influence the system while staying close enough to their true location to avoid immediate detection. To solve this, the researchers created a "trust-aware" planning system. Instead of assuming every vehicle is honest, the system constantly evaluates two types of evidence: the reliability of the location signal itself, and the actual behavior of the vehicle on the road. If a vehicle claims to be at a pickup spot but never shows up, or if its location signal looks suspicious, the system flags it. Crucially, once a vehicle is flagged as untrustworthy, the central computer stops planning routes for it entirely, effectively removing it from the team's mental map so it can no longer disrupt the flow of traffic.
The researchers tested this idea using a massive simulation based on real taxi demand data from San Francisco. They built a digital city with over a thousand intersections and simulated a fleet of autonomous vehicles trying to pick up and drop off passengers. In their experiments, they introduced a small number of "adversarial" vehicles programmed to lie about their location and ignore their assigned tasks. They found that even a single dishonest vehicle could destabilize the entire system, causing a growing backlog of unfulfilled requests and canceling rides. The system failed not because it couldn't calculate a route, but because it was calculating routes based on a false reality. When the researchers added their trust-aware monitoring, the system began to detect the liars. It used a combination of checking the GPS signal quality and watching for failed pickups to identify the bad actors. Once identified, these vehicles were excluded from future planning.
The most significant finding emerged when they tested a sophisticated planning method called "rollout." This method works by simulating the future: the computer tries out different route choices in its mind to see which one leads to the best outcome hours from now. However, the researchers showed that this method breaks down if the simulation includes vehicles that don't actually exist in the real world. If the computer simulates a future where a liar is supposed to pick up a passenger, but that liar never actually moves, the simulation becomes useless, and the computer makes poor decisions. The study demonstrated that by removing the untrustworthy vehicles from the planning process, the system could restore the accuracy of its simulations. This allowed the advanced planning method to work again, recovering its ability to find better routes and clear backlogs.
The experiments revealed that the system needed more than just a quick check of location signals. The researchers found that attackers could hide by making their lies small and subtle, making the location data look almost normal. In these cases, the system relied on "behavioral trust," which looks at the history of what the vehicle actually did. Did it pick up the passenger? Did it show up on time? By combining the check of the signal with the check of the action, the system could catch liars that would have slipped past a simple location check. The study also showed that the speed of recovery depended on how far into the future the computer could plan. A short-sighted plan struggled to fix the mess left by the liars, but a plan that looked further ahead could successfully navigate around the damage and restore order.
This work does not claim to have solved every possible attack on autonomous systems, nor does it suggest that the problem is gone. The researchers noted that their tests were simulations using real-world data, and they acknowledged that real-world sensors might behave differently or that attackers could find new ways to hide. They also pointed out that removing a vehicle permanently is a conservative choice; in a real city, a temporary pause or a reduced role might be better than a permanent ban. However, the core insight is clear: for a team of robots to function under attack, the leader must not just watch for lies, but must actively reshape its understanding of the team based on what it sees. By filtering out the unreliable members before making decisions, the system can maintain a stable and efficient operation even when some of its parts are compromised. The study suggests that the future of resilient robotics lies in this dynamic relationship between trust and planning, where the map of the world is constantly updated to reflect not just where things are, but who can be counted on to get the job done.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.