Evidence status: local Robotarium Python Simulator only. No physical Robotarium run, hardware validation, account submission, or peer-to-peer communication claim is made anywhere in this repository.
This repository contains a fixed-seed eight-robot coverage-control pilot. Each robot follows the density-weighted centroid of its discrete Voronoi cell; the official Robotarium single-integrator-to-unicycle transform and boundary-aware barrier certificate filter the resulting commands. A final common actuator scale leaves margin below the platform limits documented by Georgia Tech.
The frozen local simulation reduced the density-weighted coverage objective by 37.96%, from 0.133043 m² to 0.082543 m², while the simulator reported no errors or warnings. These are simulator measurements, not physical robot results.
Only the original experiment, tests, configuration, and generated evidence are
stored here. The Georgia Tech simulator is not copied or vendored. Setup
fetches the official Robotarium Python Simulator at commit
307269846b2761528586e9c3f47d0a8bec21692f.
That exact dependency is the base on which the pilot was developed and the
version used for the frozen evidence.
The publication candidate was extracted without modifying experiment.py
from reviewed local pilot commit
a233f3ddb350322654f800b8884571a198d9bf0b, an unpublished descendant of the
official base. The source hash is recorded in
PROVENANCE.md.
The operational region is a 2.84 x 1.64 m rectangle inset by 18 cm from the published 3.2 x 2.0 m Robotarium boundary. A deterministic 45 x 27 quadrature grid represents a nonuniform workload with a uniform floor, three unequal Gaussian peaks, and a low ridge. Eight robots start in a fixed, well-separated 2 x 4 arrangement.
At each 0.033 s control step, the experiment:
- obtains the eight poses through the public
get_poses()API; - assigns every workload sample to its nearest robot, using lowest robot ID for deterministic tie-breaking;
- computes each cell's density-weighted centroid;
- caps nominal single-integrator feedback at 0.11 m/s and maps it to unicycle commands with the official Robotarium transform;
- filters commands with the official boundary-aware unicycle barrier certificate, configured with a 0.15 m safety radius and 0.14 m/s magnitude limit; and
- applies one common scale so requested wheel speed remains at or below 12.0 rad/s and requested body rotation remains at or below 3.5 rad/s.
Each nominal command is defined by one robot's own Voronoi cell, but this implementation vectorizes the partition from globally tracked poses and uses the simulator's centralized barrier QP. It is a distributed-by-agent coverage law implemented on a centralized testbed, not a peer-to-peer system.
The 3,600-step control phase represents 118.8 s. Including ordinary simulator
initialization, the pinned simulator reported 185.30 s of simulated execution.
The random seed is 20260817.
| Metric | Initial | Final / worst local-simulator value |
|---|---|---|
| Density-weighted coverage cost | 0.133043 m² | 0.082543 m² (-37.96%) |
| Workload within 0.35 m of a robot | 56.54% | 75.91% (+19.36 points) |
| Workload within 0.50 m of a robot | 84.34% | 98.63% (+14.30 points) |
| Weighted RMS distance | 0.3648 m | 0.2873 m |
| Minimum offset collision-center spacing | - | 0.5078 m |
| Minimum reference-pose boundary clearance | - | 0.3752 m |
| Minimum conservative disk-edge boundary clearance | - | 0.3202 m |
| Maximum commanded linear speed | - | 0.1097 m/s |
| Maximum commanded angular speed | - | 3.4909 rad/s |
| Local control compute time | - | 2.99 ms p99; 4.83 ms max |
| Local simulator validation | - | no errors or warnings |
The spacing metric follows the pinned simulator's collision check: pose references are shifted 2.5 cm forward, then measured center-to-center. It is not an edge-to-edge body gap. The conservative boundary metric subtracts a 5.5 cm disk radius from the reference-pose clearance, based on the published 11 cm robot diameter.
The timing values above describe the machine that generated the frozen local run; they are not hardware latency measurements and are expected to vary on other computers. On that same frozen host, barrier filtering changed the nominal command on 20.28% of control steps, and the common actuator scale activated on 1.42%. Those two fractions count threshold crossings near equilibrium; they are environment-sensitive diagnostics, not cross-environment outcome gates.
The first GitHub-hosted full reproduction, Actions run
31993185231,
used the same controller, seed, pinned simulator commit, public direct-package
pins, and 3,600 control steps. It reproduced the coverage reduction within
8.81e-12, had a maximum sampled XY-coordinate difference of 2.361e-5 m
(23.61 micrometres) and an XY RMS difference of 1.863e-6 m. Simulator
validation remained available with zero errors or warnings.
Its minimum offset collision-center spacing was 0.5021326 m, compared with
0.5077624 m on the frozen host: a 0.0056298 m (1.109%) difference. The
hosted value still exceeded the published 0.15 m recommendation by 0.3521 m.
The 0.5078 m value in the table remains the exact frozen-host measurement; it
is not relabelled as a byte-exact cross-environment result.
The frozen PNG embeds Matplotlib 3.11.1, while the public pin and hosted run use 3.10.8. The original frozen run did not record NumPy or CVXOPT versions, so the frozen environment is only partially known. The five committed frozen files pass their SHA-256 manifest; that integrity check does not claim a later run regenerated timing-bearing metrics, summary, log, or plot bytes.
A fresh Fedora control under the exact public direct pins (NumPy 2.2.6, Matplotlib 3.10.8, CVXOPT 1.3.2) reproduced the frozen trajectory byte for byte and all gated summary fields exactly. The hosted Ubuntu run used the same public direct pins, and both install logs list the same Python package version strings. However, transitive packages are resolver-selected, wheel builds and system math libraries are not locked, and the operating system, BLAS runtime, and CPU environment differ. The audit therefore establishes bounded cross-environment variability, not a single isolated cause.
The first hosted XY delta above 1e-12 m occurred at 33.66 s, well after the
3.63 s sustained-convergence marker. The pinned official barrier utility solves
through CVXOPT with relative and feasibility tolerances of 1e-2, and the
larger trace differences are in angular/heading behavior near equilibrium.
This pattern is consistent with environment-sensitive barrier-QP numerics, but
it does not isolate CVXOPT, BLAS, package artifacts, the operating system, or
another environment component as the sole cause. Threshold diagnostics changed
more visibly (barrier intervention: 20.28% frozen versus 42.72% hosted;
actuator scaling: 1.42% versus 3.61%) while XY motion, safety margin, and
coverage were stable.
The cross-environment comparison contract was introduced transparently after that first hosted run and corrected after the clean public-pin control established the evidence above. It retains exact controller configuration, simulator commit, evidence scope, validation availability, and zero-error requirements. It also enforces configured command limits, positive boundary clearance, at least 0.15 m collision-center spacing, no more than 0.01 m spacing drift from the frozen run, field-specific outcome tolerances, and trajectory limits of 50 micrometres per XY coordinate and 5 micrometres RMS. Tests demonstrate that unsafe spacing, simulator errors, changed guards, outcome regressions, actuator regressions, and meaningful trajectory drift all fail the comparison.
Prerequisites are Git and Python 3.10-3.12. The direct package versions and
official simulator commit are pinned in pyproject.toml;
transitive packages retain the constraints selected by those distributions.
python3.12 -m venv .venv
.venv/bin/python -m pip install --upgrade pip
.venv/bin/python -m pip install -e '.[test]'
.venv/bin/python -m pip freeze --all
.venv/bin/python scripts/verify_frozen_evidence.py
.venv/bin/python -m pytest -q
mkdir -p results/reproduction
ROBOTARIUM_HEADLESS=1 ROBOTARIUM_FAST=1 \
ROBOTARIUM_OUTPUT_DIR=results/reproduction \
.venv/bin/python experiment.py | tee results/reproduction/simulator.log
.venv/bin/python scripts/compare_reproduction.py \
evidence/simulator-3072698/robotarium_coverage_summary.json \
results/reproduction/robotarium_coverage_summary.txt \
evidence/simulator-3072698/robotarium_coverage_trajectory.npy \
results/reproduction/robotarium_coverage_trajectory.npyThe comparison deliberately excludes wall-clock compute-time fields, plot bytes, and the two environment-sensitive threshold diagnostics. It reports those diagnostic deltas while enforcing the documented semantic outcome, safety, actuator, validation, and full sampled-trajectory gates above. The test suite also includes a short end-to-end simulator run. Continuous integration performs the same integrity, test, and full-reproduction checks.
evidence/simulator-3072698 contains the frozen
local run:
- comma-delimited time-series metrics;
- sampled poses as a NumPy array with shape
361 x 2 x 8; - machine-readable summary JSON;
- summary figure; and
- captured simulator output.
All hashes are pinned in
evidence/simulator-3072698/MANIFEST.sha256.
Run the integrity checker before quoting any result.
experiment.py uses only the public get_poses, set_velocities, step, and
debug methods for control. One isolated helper optionally reads the local
simulator's private _errors dictionary for evidence reporting. If a hardware
runtime omits it, validation is marked unavailable rather than falsely reported
as zero; control and file generation do not depend on it.
The pinned 2024-11-20 guide uses the older call_at_scripts_end() name, while
the pinned simulator source exposes debug() and its examples use debug().
This project follows the tested pinned source. Confirm the live runtime before
any future upload.
A physical run requires a separately approved Robotarium account and submission. After a real run, preserve the returned script, settings, logs, data, and video; verify them independently before adding any hardware claim. Until that happens, the accurate description is Robotarium-ready local simulation.
See NOTICE.md for dependency attribution and license boundaries.
This project is independent and is not endorsed by or affiliated with Georgia
Tech or the Robotarium organization.
Original project files are released under the MIT License; see LICENSE.
No third-party package is redistributed here. Installed dependencies retain
their own licenses, including CVXOPT's GPL-3.0-or-later terms; see
NOTICE.md.