ARCS: Autonomous Rendezvous and Control Suite

Hammoudeh, Jamil

Abstract

ARCS is an open-source, headless simulator for testing AI controllers, including large language model (LLM) agents, and classical controllers on spacecraft rendezvous with keep-out safety constraints. It models relative motion with the Clohessy-Wiltshire equations, validated against the closed-form solution to sub-millimeter accuracy over one orbit, and simulates one hour of flight in under 0.1 s. A provider-neutral LLM agent interface logs every prompt, response, parse failure, and latency, and a safety-filter wrapper counts its interventions. In the first result, a PD baseline violates the keep-out zone in 100% of 30 obstructed approaches, which motivates safety-filtered autonomy.

What ARCS is

ARCS (Autonomous Rendezvous and Control Suite) is a simulator for one question: can an AI controller fly a spacecraft rendezvous without breaking a safety constraint? It runs headless, so experiments are scripts, not screen sessions. It is open source under the MIT license.

ARCS tests two kinds of controllers on the same problem:

AI controllers, including large language model (LLM) agents.
Classical controllers, such as a PD baseline.

Every approach has a keep-out zone around the target. Entering it counts as a violation.

Dynamics and validation

Relative motion follows the Clohessy-Wiltshire equations. I validated the propagator against the closed-form solution. The error stays under one millimeter over one full orbit. One hour of simulated flight runs in under 0.1 s, so large batches of approaches are cheap.

LLM agent interface

The agent interface is provider-neutral. Any LLM can sit behind it. For every step, ARCS logs:

the prompt sent to the model
the response
any parse failure
the latency

These logs make agent runs auditable after the fact.

Safety filter

A safety-filter wrapper can sit between any controller and the spacecraft. When a command would break the constraint, the filter steps in. ARCS counts every intervention, so a result shows both whether the approach stayed safe and how often the filter had to act.

First result

A PD baseline violates the keep-out zone in 100% of 30 obstructed approaches. That result is the case for safety-filtered autonomy, and it sets up the papers below.

What comes next

Two papers built on ARCS are in progress. Neither is published.

The code, tests, and roadmap are in the ARCS repository on GitHub.