The problem of in-orbit collision is becoming a significant concern for the space community. On one hand, more and more objects are placed in Low Earth Orbits (LEOs) each year because of the decreasing cost to launch satellites. On the other hand, as handling these objects at their end-of-mission is challenging, thousands of space debris exist and will remain in orbit for decades to come. Therefore, the risk of collision in very populated regions of space is becoming very likely and must be addressed.

To properly dispose of a satellite at its end-of-mission, significant maneuvers must be performed to reduce its altitude and let it burn in the Earth's atmosphere. Traditionally, this process puts a significant burden on human operators who have to design a de-orbit plan and send commands to the satellite. The reliance on the ground segment also limits the overall reactivity during de-orbiting. In fact, when lowering its altitude, the satellite will cross very crowded regions, in particular around the orbits of mega-constellations. This translates into significant collision risks to monitor and avoid if necessary. However, due to visibility constraints between the satellite and ground stations, trajectory updates can be significantly delayed.

To improve the reactivity of this collision avoidance process, autonomous systems have raised increasing interest. On-board risk assessment and maneuver computation present multiple benefits. First, the reliance on the ground is reduced, which improves resilience in case of missed visibility periods or communication failures. Moreover, the maneuver planning can benefit from a better knowledge of the satellite's current status using on-board sensors. To this end, on-board algorithms have been designed to compute station-keeping maneuvers with collision avoidance considerations.

We propose an extension of these methods in the context of a de-orbiting trajectory. In this mission phase, the satellite is not constrained by an operational orbit to maintain but has to deal with additional limitations, namely the varying altitude, atmospheric uncertainties and degraded performance of thrusters or batteries at end-of-life. If electrical propulsion is used - like many New Space missions - the duration to reenter the atmosphere is extended, hence a higher risk of collision. The presented methodology relies on a short-term horizon, where collision risks are well identified through Conjunction Data Messages (CDM) and a long-term horizon, in which the collisions are regarded from a statistical point of view. The trajectory optimization problem encapsulates both of these aspects to design a safe maneuver plan while being compliant with on-board computing capabilities.