Chance-Constrained Inspection Viewpoint Sequencing for On-Orbit Servicing via Safe Reinforcement Learning

Multi-target on-orbit servicing missions require planning tools that can adapt to evolving operational conditions while accounting for the outcome variability inherent to safety-critical spaceflight. This paper presents an adaptive planning approach for multi-target servicing campaigns in low Earth orbit and applies it to inspection missions flown by a low-thrust servicer. The approach is based on Distributional Reinforcement Learning, which learns the full distribution of mission returns and thereby supports explicit risk-sensitive decision-making at each planning step. Specifically, a Quantile-Regression Deep Q-Network (QR-DQN) agent selects which target to inspect next by jointly considering the expectation and the variance of the predicted return through a mean-variance criterion. The mission environment models circular orbits with a spread in inclination and uses Edelbaum's combined-manoeuvre low-thrust transfer estimates. The observation comprises the orbital elements, the visit history, the observed risk state of each target and the remaining mission-time budget. The risk state of each target changes stochastically during the campaign, making the outcome of an inspection uncertain and creating operationally meaningful distinctions between risk-neutral and risk-averse planning strategies. Because the mission-time budget generally admits only a subset of the candidate targets, the planner must also select which targets to inspect. Numerical experiments on scenarios with five target satellites show that one risk-aversion weight moves the policy across a range of mean return and return variance, over an unchanged environment and reward. Risk-averse agents reduce the variance of the episodic return by approximately one half relative to risk-neutral agents and lower the probability of a mission loss, at a cost in mean return. To validate the learned policies, the efficient frontier, which gives the highest mean return attainable at each level of return variance, is computed exactly by enumerating the admissible visit orders. At equal variance, the learned policies lie close to this frontier in mean return, and the risk-averse policies have a lower probability of a mission loss than the frontier policies.

Autonomous inspection of a resident space object is usually posed as a contest between classical next-best-view planning and reinforcement learning. We argue that the choice of algorithm is the least informative of the design axes. An inspector spacecraft has no rest state: when it stops thrusting, orbital dynamics carry it along a natural trajectory that can be used to observe the target at no fuel cost. On a controlled ladder in which one design axis is varied at a time, fuel efficiency turns out to be set by three things: whether the agent is allowed to drift, whether its state records where drifting has taken it, and whether its planning horizon is long enough for the benefit to appear. Removing any one of them doubles the fuel bill or more. The algorithm still matters, but through what it can represent: a model-based agent whose learned world model ignores fuel spends several times more of it. Finally, no learned policy is safe by construction, and a control barrier function filter beneath the decision layer removed every keep-out-zone violation in two thousand evaluation episodes.

Harsh and rapidly varying orbital illumination introduces failure modes during spacecraft inspection. Existing planners rely on line-of-sight inspection metrics which perform occlusion checks against a known target and conflate geometric visibility with task-relevant information. These metrics overlook illumination effects including eclipse, self-shadowing, and specular reflection from spacecraft materials and tend to declare components as inspected despite the corresponding images lacking sufficient information. We propose an inspection-planning framework that replaces geometric visibility with a learned measure of camera information derived from a classifier variational autoencoder (VAE) that operates directly on the captured image. The encoder model is trained to classify spacecraft surface components from rendered patches, and its total latent variance is used to learn the varying illumination of said components, replacing binary visibility checks with a continuous, image-grounded quantity. We evaluate the approach by implementing a Guidance, Navigation, and Control architecture in a photorealistic physics simulator, and vary lighting and viewing geometries with multiple orbits. We show that, although all approaches successfully maximize their own objective, our approach results in better illumination: a mean brightness of $0.57$, with $81.6\%$ of observations above $0.25$, compared with $0.02$ ($3.0\%$) for geometric planning with an illumination band and $0.04$ ($6.0\%$) for purely geometric planning.