A Prediction Trick Lets Fewer Satellites Catch a Runaway One

Three spacecraft close in on a fourth that refuses to be caught, each one just far enough away that a single burn of the target’s thrusters could throw the whole chase off course. A team working out of Harbin Institute of Technology built a pursuit algorithm that guesses that burn before it happens, and in simulation, the foresight lets three chasers corner an evasive spacecraft about as reliably as a much larger fleet manages without it. The trick sounds almost too simple: teach the hunters to predict where their target is about to go, then hand that guess to the same learning system that decides how they maneuver. It works because, in orbit, knowing a rival’s next move turns out to matter more than having another ship on the chase.

Orbital pursuit games are not new; researchers have modeled spacecraft chasing each other since 1964. What has changed is how crowded the game board has gotten: roughly 46,860 objects are now regularly tracked in Earth orbit, and every one of them is a thing a maneuvering spacecraft has to reckon with.

Every maneuver in this game is a single sharp burn rather than a continuous push, the kind of impulsive delta-v that shows up across the wider orbital pursuit-evasion literature because it is what fuel-limited spacecraft actually use. The pursuers and their target both move under the Clohessy-Wiltshire equations, the standard shorthand physicists use for how one spacecraft drifts relative to another nearby, and both sides are boxed in by how much they can burn at once and how much fuel they have left overall. A chase counts as a win the moment any pursuer gets within 5 kilometers of the target; anything short of that, and the evader has slipped away. Pursuers can only compare notes with the two nearest teammates inside a 100 kilometer communication bubble, so no single ship ever sees the whole board.

The obvious fix for a fast, evasive target is more hunters, and that gets expensive fast, in fuel and in satellites nobody wants to spare for one chase. So the team asked a different question: what if the pursuers got better at guessing, instead of more numerous?

Teaching the Hunters to Guess

Their answer pairs two neural networks: one is a standard reinforcement-learning system built for continuous control that every pursuer runs identically, reacting only to what it can observe locally, and the other is new, a lightweight add-on trained only after the main policy already exists. That second network watches the relative positions and velocities of the nearest pursuers and forecasts the evader’s next burn before it fires. Because it trains on data the main policy already generated during its own runs, it costs almost nothing extra to build.

There is a nice bit of irony in an evasion-prediction system built by watching an evader try to escape a machine that is, itself, still learning how to hunt. The prediction network never sees the evader’s control code; it only ever sees the traces the evader leaves in its own past maneuvers.

Inside the Two Networks

Training runs under what the field calls centralized training with decentralized execution: during training, every pursuer sees the full state of the game, including the evader’s real maneuvers, information no pursuer gets once the system actually flies. Once training ends, each of the identical policy networks acts on nothing but its own 29-number picture of the world, its distance and closing speed on the evader, the positions of its two nearest teammates, and now, the predicted next move of the target. Both the decision-making policy and the prediction network are compact, four-layer systems, the kind that could plausibly run on modest onboard hardware rather than a data center. The reward that shapes the pursuers’ behavior is simple by design, a bonus for closing distance, a larger one for an outright capture, and a penalty for burning fuel, since a more elaborate scoring scheme risked becoming as hard to tune as the pursuit strategy itself. Across 200,000 joint training episodes, the reward curves for pursuer and evader move in opposite directions and then flatten out after roughly 125,000 episodes, the simulation’s version of two rivals reaching a rough truce.

The prediction network’s forecasts held remarkably tight through an actual chase, a positional error typically under 0.3 kilometers and averaging about 140 meters across the whole pursuit. Two older tricks, Kalman filtering and simply assuming the evader repeats its last move, tracked almost as well early on but drifted toward errors approaching 0.6 kilometers exactly as the chase tightened and precision mattered most.

When the Target Stops Playing Along

None of this holds against an evader playing a different game than the one it trained for: swap in a target that moves randomly or on an unfamiliar repeating pattern, and both policies lose ground, the learned one included. A system trained to anticipate one style of evasion carries no guarantee it can anticipate another, which is exactly the bottleneck the authors flag as unresolved.

Measured across 100 randomized simulation runs per configuration, the payoff still shows up hardest exactly where fuel is scarcest: with three pursuers, the prediction-equipped policy reached a 95 percent capture rate in simulation, roughly 30 percentage points above the conventional approach at the same pursuer count. With only two pursuers it still closed the target down 65 percent of the time, a gap wide enough to matter when a mission planner is deciding how many spacecraft a chase is worth. Against a target that never maneuvers at all, the prediction-equipped policy’s success rate climbs to 99 percent, because a motionless target is trivial to forecast, while the conventional policy manages only 38 percent of its chases. None of the study’s chasers worked with just one pursuer against one evader, though; a single ship cannot herd a target free to dodge in any of three dimensions, no matter how well it guesses.

The team is already looking past the specific game modeled here, toward pursuit strategies that need no advance data on the evader at all, and toward the messier problem of many pursuers chasing many targets at once rather than the tidier version studied so far. Whatever comes of that, the underlying wager is a modest but real one: a fleet does not have to be bigger to be better at cornering something that does not want to be found, so long as it can spend some of its intelligence guessing what happens next.

FAQ