01: TASK OUTCOMES
40 of 70 quadratic runs converged.
The baseline completed 0 of 70 runs under the same scenarios, seeds, mission limits, and safety filter. The quadratic policy uses a mock recovery agent; the baseline runs with the agent off, so this compares complete policies rather than isolated estimators.
Across the four nominal scenarios, 40 of 40 quadratic trials converged. The three stress scenarios contributed 0 of 30 successes; their infeasible routes and short time budget remain failures in the total above.
40 of 70 successes
0 of 70 successes
Seeded mock recovery runs
Quadratic terminations: 20 route infeasible, 10 timeout. Baseline terminations: 70 timeout.
02: BY SCENARIO
Failures remain visible by case.
| Scenario | Quadratic Success | Failure Reasons | Baseline Success |
|---|---|---|---|
| Open field | 10 of 10 | No failures | 0 of 10 |
| Noisy open field | 10 of 10 | No failures | 0 of 10 |
| Restricted corridor | 10 of 10 | No failures | 0 of 10 |
| Disturbed corridor | 10 of 10 | No failures | 0 of 10 |
| Near boundary | 0 of 10 | 10 route infeasible | 0 of 10 |
| Infeasible enclosure | 0 of 10 | 10 route infeasible | 0 of 10 |
| Far source | 0 of 10 | 10 timeout | 0 of 10 |
03: TRAJECTORY METRICS
Mean, spread, and median for each controller.
Values come from recorded trajectories. Obstacle clearance is unavailable in scenes without obstacles.
| Trajectory Metric | Quadratic Mean ± SD | Median | Baseline Mean ± SD | Median |
|---|---|---|---|---|
| Localization error (m) | 0.000 m ± 0.001 m | 0.000 m | 1.340 m ± 1.306 m | 0.466 m |
| Formation error (m) | 0.225 m ± 0.235 m | 0.046 m | 0.969 m ± 0.971 m | 0.416 m |
| Minimum separation (m) | 0.436 m ± 0.077 m | 0.413 m | 0.938 m ± 0.128 m | 0.885 m |
| Obstacle clearance (m) | 0.226 m ± 0.037 m | 0.203 m | 0.203 m ± 0.003 m | 0.203 m |
| Total path length (m) | 44.238 m ± 12.550 m | 46.831 m | 71.437 m ± 66.091 m | 46.491 m |
| Simulation duration (s) | 28.490 s ± 15.018 s | 27.200 s | 80.000 s ± 24.672 s | 90.000 s |
| Safety interventions | 65 ± 103 | 0 | 391 ± 417 | 1 |
| Robot collisions | 0 ± 0 | 0 | 0 ± 0 | 0 |
Obstacle clearance uses 30 samples per controller. Other metrics use 70 runs per controller.
04: BOUNDED AGENT RECOVERY
10 of 10 seeded corridor disturbances recovered.
05: SCOPE AND REPRODUCTION
This sample does not establish general reliability.
- Ten fixed seeds per scenario provide limited spread evidence, not broad robustness evidence.
- All agent decisions used the deterministic offline mock; Baseten was not called.
- Live inference latency was not measured because this evaluation used no live requests.
- The baseline and proposed controller share the same formation objective and safety filter.
Snapshot generated 2026-11-04T18:05:31+00:00. Regenerate it with scripts/run_evaluation.py.