Multi‑target on‑orbit servicing missions require planning tools that adapt to evolving operational conditions while accounting for the inherent outcome variability of safety‑critical spaceflight. This paper introduces an adaptive planning approach for low‑Earth‑orbit campaigns and applies it to inspection missions performed by a low‑thrust servicer. The approach relies on Distributional Reinforcement Learning, which learns the full distribution of mission returns and thus enables explicit risk‑sensitive decisions at each planning step. Specifically, a Quantile‑Regression Deep Q‑Network (QR‑DQN) agent selects the next target by jointly considering the predicted return’s expectation and variance through a mean‑variance criterion.\n\nThe mission environment models circular orbits with a spread in inclination and uses Edelbaum’s combined‑maneuver low‑thrust transfer estimates. Observations consist of orbital elements, visit history, the observed risk state of each target, and the remaining mission‑time budget. Each target’s risk state evolves stochastically during the campaign, making inspection outcomes uncertain and creating operationally meaningful distinctions between risk‑neutral and risk‑averse planning strategies. Because the time budget typically permits only a subset of candidate targets, the planner must also decide which targets to inspect.\n\nNumerical experiments with five target satellites show that a risk‑aversion weight moves the policy across a range of mean return and return variance while keeping the environment and reward unchanged. Compared with risk‑neutral agents, risk‑averse agents halve the return variance and markedly reduce the probability of mission loss, at the cost of a modest drop in mean return. To validate the learned policies, the efficient frontier—giving the highest mean return attainable at each variance level—is computed exactly by enumerating all admissible visit orders. At equal variance, the learned policies achieve mean returns close to this frontier, and the risk‑averse policies exhibit a lower mission‑loss probability than the frontier policies.\n\nReview: The work demonstrates the practicality of distributional reinforcement learning for space mission planning, especially the QR‑DQN‑based mean‑variance risk control, which provides quantifiable safety guarantees for low‑thrust on‑orbit inspections. Future work could extend the framework to richer orbital dynamics and multi‑mission coordination scenarios.