Multi-Robot Networks Track Moving Targets via Decentralized Information-Driven Strategy

In what could reshape how search-and-rescue teams, security services, and environmental scientists deploy autonomous systems, researchers have unveiled a decentralized algorithm that allows a network of robots to track more moving targets than there are robots in the network — a problem long considered one of the thorniest computational challenges in multi-robot coordination. The work, published in the journal Autonomous Robots, demonstrates in both simulation and real-world experiments with ground robots tracking humans that a small team of machines can collaboratively divide up an overwhelming number of moving targets, assign them among themselves, and optimize their movement in real time, all without a central commander.

The core difficulty is deceptively simple to state. If you have, say, five robots and twelve people moving through a disaster zone, every target must be assigned to exactly one robot — no target can be ignored, and no two robots should waste effort duplicating each other’s work. Meanwhile, each robot must constantly adjust its own trajectory as both it and its targets move. Because each robot may end up responsible for multiple targets whose tracking objectives conflict — pursuing one person might mean turning away from another — the robot must somehow balance competing demands every single moment. The researchers formally proved that this combined problem of simultaneous target assignment and control is NP-hard, meaning that as the number of robots and targets grows, finding the perfect solution quickly becomes computationally intractable even for powerful computers.

“Real-world missions like urban search and rescue, underwater exploration, and disaster monitoring don’t wait for a centralized supercomputer to figure things out,” the research team notes in the paper. Communication between robots is often unreliable in these environments, which makes centralized approaches prone to single-point failures — if the central node goes down, the whole operation collapses. The team’s answer is a fully decentralized framework called bundle-based assignment and control, or BBAC, which runs concurrently on every robot in the network and relies only on local communication between neighboring machines.

The elegance of BBAC lies in how it dismantles the monstrous optimization problem into two digestible stages using a mathematical technique known as primal decomposition. The key insight is that the target assignment acts as a coupling variable: once you fix which robot is responsible for which targets, the giant problem splits cleanly into independent control problems that each robot can solve on its own. The first stage, called the Master Problem, has the network collaboratively decide who tracks whom. The second stage, the Subproblem, has each robot independently compute the movement commands that best serve its assigned targets.

For the Master Problem, the researchers drew inspiration from an unlikely source: economics. They developed an auction-inspired algorithm in which robots literally bid on targets. Each robot maintains a “bundle” — an ordered list of the targets it wants to track, built by sequentially adding the targets that offer the highest tracking value. Every target carries a price, and the robots iteratively place bids, broadcast those bids to their neighbors, and reach consensus on conflict-free assignments. When a robot is outbid on a target, it releases that target from its bundle and moves on. The result is a rapid, adaptive marketplace of tracking duties that adjusts instantaneously as robots and targets move — a critical advantage over earlier auction methods that assumed static targets and required complete replanning whenever anything moved.

Crucially, the researchers proved that this bundle-based assignment comes with a hard performance guarantee: in the worst case, it achieves at least half the performance of the theoretical optimum, a so-called one-half approximation guarantee. In the world of combinatorial optimization, where many problems offer no guarantees at all, this is a meaningful assurance that the decentralized auction won’t produce catastrophically poor assignments even in the most adversarial scenarios.

The second stage of BBAC tackles the subtle problem of a robot with multiple, conflicting objectives. The researchers formally demonstrated that when a robot is assigned several targets, no single movement can always maximize tracking quality for all of them simultaneously — imagine two targets placed on exactly opposite sides of the robot, where moving toward one necessarily moves away from the other. Mathematically, the best a robot can do is find a Pareto-optimal control: a movement that cannot be improved for any one target without degrading tracking of another. The team showed that a weighted-sum formulation of the problem provably produces such Pareto-optimal solutions, and they employ a genetic algorithm — a search method inspired by biological evolution that adaptively samples possible control actions and can escape local optima and handle non-convex, discontinuous objectives — to find these solutions fast enough for real-time operation.

Underpinning both stages is a novel tracking utility function that blends two competing philosophies of target tracking. Information-driven objectives, rooted in Shannon’s definition of information, quantify how much a future measurement would reduce uncertainty about a target’s estimated position and velocity. These functions push robots toward poorly tracked targets whose state estimates grow fuzzy over time, improving long-term performance — but they suffer from diminishing returns when a target is already being tracked well. Task-driven objectives, by contrast, reward keeping close to currently tracked targets for immediate gains. The new utility function is a hybrid: it combines the expected entropy reduction from a future sensor measurement (modeled through a recursive Kalman-filter covariance update and a Bernoulli detection model that captures the robot’s bounded, directional field of view) with a geometry-aware distance reward computed as an integral over the sensor’s field of view, plus a penalty term that discourages collisions with obstacles, other robots, and the targets themselves. The weights on these components can be tuned to the mission — for instance, a team operating in a cluttered warehouse would weight collision avoidance more heavily than one patrolling an open field.

The practical payoff was demonstrated on two fronts. In extensive simulations, BBAC consistently outperformed existing decentralized tracking algorithms and came remarkably close to the offline optimal solution — which requires far more computation — while running in a fraction of the time. More compellingly, the researchers deployed the system on a physical network of unmanned ground vehicles equipped with RGB-D cameras, tasked with tracking humans walking through an indoor workspace. The robots communicated over a limited range, negotiated their target assignments through the decentralized auction, and executed genetically optimized control commands in real time, successfully maintaining surveillance of moving people despite the targets outnumbering the robots. The framework is sensor-agnostic: while the experiments used depth cameras, the authors emphasize it applies equally to lidar, radar, or sonar, opening doors to underwater and aerial applications.

What makes the work especially significant is its robustness to the messy realities of field deployment. The robots share positions, target measurements, and bid information over multi-hop wireless links within a fixed communication range, meaning the network can fragment into smaller local clusters if robots drift apart — and the system keeps functioning. Earlier decentralized approaches either assumed static environments requiring constant replanning, restricted robots to a pre-approved menu of motion primitives, or offered no guarantees that assignments would remain mutually exclusive when robots carried limited field-of-view sensors. BBAC addresses all three shortcomings at once.

The implications stretch well beyond the laboratory. Search-and-rescue teams facing an unknown number of survivors in a collapsed structure cannot field a robot for every possible person; security operations monitoring crowds must do more with fewer assets; environmental scientists tracking animal populations face the same arithmetic imbalance. A proven, guaranteed, decentralized method that lets a handful of autonomous machines cover many more moving subjects — without a vulnerable command center and without prohibitive computation — addresses the fundamental economics of robotic surveillance. The researchers indicate that future work will tackle more sophisticated consensus strategies for merging target state estimates across the network, but for now, the days when a small robot team could only watch helplessly as targets outnumbered it may be drawing to a close.