Satellites can now rehearse how to dodge a collision before they make a move, thanks to a new AI planner that imagines the future instead of just reacting to it.
Orbit-Planner, described in a new arXiv paper, is a two-stage world model that learns how a spacecraft's actions change its trajectory, then uses that knowledge to roll out imagined future paths inside a compressed digital representation rather than a full physics simulation. A companion module called the Physics Probe converts those imagined paths back into real physical state information, so the system can compare each rollout against where it expects obstacles to be. If a rollout's decoded trajectory points toward a collision, the planner discards it and imagines another, repeating that check-and-replan cycle at every control step - what the researchers call closed-loop navigation. Tested in the Isaac Sim simulator, the approach avoided collisions successfully 91.7% of the time.
Most on-orbit collision-avoidance systems lean on pre-built maps and fixed assumptions about the environment - fine for known debris fields, brittle when something unexpected shows up. Orbit-Planner's pitch is adaptability: because it learns dynamics rather than memorizing a map, it should, in theory, generalize to situations its designers never explicitly programmed for.
That 91.7% is a simulation number, not a flight record, and Isaac Sim is a far more forgiving environment than the real unpredictability of orbital debris - the gap between the two is exactly where research like this usually gets tested hardest.