V-BAR: Verified Benchmark for Autonomous Rendezvous

vbar is a fast, headless, cross-platform benchmark for testing AI and classical spacecraft controllers, with and without safety filters.

Status: v0 (early development). One scenario, validated close-range dynamics, classical and LLM agent interfaces.

Why

Testing AI controllers for spacecraft rendezvous today usually means running a full game engine (KSPDG, built on Kerbal Space Program): a paid game, GUI-only, macOS/Linux, real-time. That makes large, statistically meaningful experiments slow. V-BAR runs anywhere Python runs, with no game, and simulates an hour of flight in a fraction of a second.

Goals:

• Fair comparisons: the same scenarios, metrics, and number of runs for every controller: classical, reinforcement learning, or LLM.
• Safety as a first-class metric: keep-out violations are checked continuously (including between time steps), and safety filters can be switched on and off.
• Reproducible: seeds, logged prompts and replies, and validated physics.

Install

pip install -e ".[dev]"   # from a clone of this repo

Quickstart

from vbar import ApproachConfig, ApproachEnv, run_episode
from vbar.agents import PDAgent

env = ApproachEnv(ApproachConfig(start_state=(0, 400, 0, 0, 0, 0)))
result = run_episode(env, PDAgent())
print(result["captured"], result["keep_out_violated"], result["dv_used_mps"])

Run the demo (saves a trajectory plot to outputs/):

python examples/v0_demo.py

Plug in any LLM

from vbar.agents import LLMAgent

def complete(prompt: str) -> str:
    ...  # call any model and return its text reply

agent = LLMAgent(complete, name="my-model")
result = run_episode(env, agent)
agent.log  # every prompt, reply, parsed action, parse failure, latency

Scenario: approach to hold point

The chaser must reach a hold point 100 m behind the target on the V-bar and stop, without entering a 50 m keep-out sphere around the target.

Dynamics: Clohessy-Wiltshire relative motion (circular target orbit at 500 km, close range, no perturbations). Integrated numerically (DOP853) and validated against the exact closed-form solution to sub-millimeter accuracy over a full orbit (tests/test_cw.py).

Metrics (in info and the result dict): capture, keep-out violation (checked between steps), minimum range to target, delta-v used, time, filter interventions.

Validation

pytest -q

Tests check the dynamics against known exact results: the closed-form solution, V-bar equilibrium, closed relative ellipses, and the 12π·|x0| along-track drift per orbit.

Citation

See CITATION.cff.

License

MIT