This site is the working journal of a doctoral research project — risk-aware planning and control for a free-flying space manipulator. Everything on it is generated from one analysis pipeline and version-controlled sources: every number traces to a logged simulation run, every figure to a committed specification, every derivation to a machine-checked proof or a step-by-step write-up. This page is the front door: what the project is, where to start reading, and how each phase connects.
Risk-Aware Active Inspection Planning and Control for Free-Flying Space Manipulators under State, Actuation, and Camera-Pose Uncertainty
Central question: how can a free-flying robotic inspection system make better inspection decisions under uncertainty? Two supporting questions follow: can risk-aware planning and control produce a more robust controller, and how must the guidance and control algorithms change to achieve it?
A free-flying space manipulator — a fully-actuated spacecraft base carrying a robotic arm with a camera in place of an end-effector — inspects a target satellite. The nominal guidance and control already fly the full inspection orbit. The thesis question is what happens when the system stops pretending it knows everything: when its own state estimate, its actuation, and its camera pose all carry uncertainty. The work formalizes that uncertainty explicitly, scores camera views by risk rather than by expectation alone, and makes the planner and controller behave more cautiously exactly when confidence is poor — then measures what that caution costs and buys.
| Phase | Question it answers | Status (July 2026) |
|---|---|---|
| Nominal lockdown | Is the deterministic baseline trustworthy enough to perturb? | active — the mission-clock case is closed; the along-track velocity bias is the sharpened open question |
| Phase 0 — foundation | What breaks first, and how do we measure “worse”? | active — three of four risk metrics implemented and verified |
| Phase A — CVaR view scoring | Does risk-aware view scoring beat nominal and expected-value scoring under camera-pose uncertainty? | derivation drafted, under review |
| Phase B — state uncertainty | How should guidance behave when the state estimate itself is uncertain? | upcoming, detailed after Phase A’s gate |
| Phase C — actuation uncertainty | Can the controller derate itself risk-awarely? | upcoming, after Phase B’s gate |
| Phase D/E — objective extensions | Reconstruction confidence and probabilistic margins as objectives | upcoming, research-first |
Two tracks run in parallel and never gate the main line: the seven-degree-of-freedom extension, and the Lean 4 formalization of the stability mathematics.
Every phase is pre-registered before it runs: the question, the prediction, and the pass/fail criteria are written down first, and the measurement follows. Every simulation result comes from one pinned pipeline — a run specification in version control, executed by one orchestrator, measured by one analysis module. Every derivation is machine-checked symbolically before it is trusted, and the stability layer carries a second, stricter seal: proofs compiled by the Lean 4 kernel, each accompanied by its human derivation. Where a result disagrees with a prediction, the disagreement is reported as the finding.