PO-RRT*: Pareto-optimal, safety-aware, multi-objective path planning
Date
2026
DOI
Authors
Version
OA Version
Citation
Abstract
Trusting an autonomous agent requires an immense effort in transparency and an even greater effort in providing safety. Agents must balance the difficult task of completing predefined objectives, all-the-while avoiding hazards and obstacles. Many modern path planning algorithms and frameworks condense independent objectives into a single weighted sum, muddling the uniqueness of each and mixing safety with other goals. This thesis proposes PO‑RRT*, a sampling‑based framework that extends Bellman’s principle of optimality from scalar values to value sets, so that each node carries a Pareto-optimal set of cost vectors in a multi‑objective, multi-dimensional space. The two primary objectives we study are Euclidean path length and probability of failure computed from a discretized environment map. Rather than combining them into one scalar, we compute and propagate Pareto‑fronts over the course of generating the tree, yielding consistently optimal routes that provide unique solutions balanced between risk and efficiency. We implement human-robot cooperation in the form of pre‑processing and post processing, where a user may implement constraint filtering, path clustering, and objective scalarization for further analysis once tree generation is complete. These constraints resemble control barrier functions in spirit, restricting admissible paths by hard bounds on risk and cost while preserving Pareto optimality during planning. The uncertainty model is made consistent by treating risk as an additive log‑survival parameter accumulated along each edge of a path. We demonstrate PO‑RRT* in simulation, show the evolution of value sets via Pareto checks during expansion and rewire, and present human selection across the resulting Pareto front of complete paths. Compared to standard strategies such as re-running RRT* and weighted sum scalarization, PO‑RRT* preserves objective independence, stores dominant paths, and most importantly, the heart of the proposal, PO-RRT* presents safety as an independent, explicitly calculated, and observable metric to the user.
Description
2026