Benchmarking LLM agents and classical controllers for safe spacecraft rendezvous
Paper in progress · 2026
Abstract
A planned benchmark, built on the ARCS simulator, that compares large language model (LLM) agents, classical controllers, and safety-filtered versions of them on spacecraft rendezvous with keep-out safety constraints.
Status
In progress. Working title. Nothing has been published yet. This page will link the preprint when one is posted.
Scope
The benchmark compares three kinds of controllers on the same rendezvous problems in ARCS:
- large language model (LLM) agents
- classical controllers
- safety-filtered versions of them
The starting point is the first ARCS result: a PD baseline violates the keep-out zone in 100% of 30 obstructed approaches.
Cite
A citation will be posted here with the preprint.