This paper addresses autonomous inspection of a target spacecraft as the first step of on‑orbit servicing. Selecting observation viewpoints is a combinatorial problem constrained by coverage, fuel, and safety. The authors model viewpoint sequencing as a Constrained Markov Decision Process (CMDP) and solve it with Proximal Policy Optimization (PPO) augmented by Lagrangian relaxation. The system consists of a high‑level planner that orders camera viewpoints on an observation sphere around the target, and a low‑level dynamics model that uses the Clohessy‑Wiltshire (CW) equations to evaluate each relative‑motion transfer. Deterministic constraints on single‑impulse reachability and the $\Delta v$ cap are enforced directly on the action set. Navigation and maneuver errors make the actual safety margin stochastic, so a chance constraint is imposed on the realized trajectory. The Lagrangian multiplier is adjusted online by a PID controller, allowing the penalty to adapt during training. Experiments use a Galileo‑class navigation satellite in medium Earth orbit, modeled as a point cloud where each viewpoint’s visibility accounts for field‑of‑view and line‑of‑sight occlusion. Across different risk budgets, the learned policy keeps the keep‑out violation rate within the prescribed limit while achieving near‑maximum coverage and respecting the $\Delta v$ budget. Compared with the conventional practice of adding a fixed safety margin to the keep‑out radius and planning deterministically, the chance‑constrained policy reduces the violation rate to less than half at comparable coverage, making collision risk an explicit design parameter and eliminating per‑case retuning. Moreover, the policy selects a viewpoint without the repeated Monte Carlo risk evaluations required by a greedy chance‑constrained planner at every decision point.
Review: The work explicitly incorporates safety constraints into a reinforcement‑learning framework, delivering an efficient viewpoint‑sequencing solution that is both more reliable and easier to tune for on‑orbit servicing missions.