Generative Simulation Benchmarking for Autonomous Urban Air Mobility Routing for Low‑Power Autonomous Deployments

Description and use

Generative scenario synthesis creates realistic, adversarial tests for low‑power autonomous agents, enabling compute‑aware, energy‑accounted benchmarking of routing policies.

Problem

Low‑power autonomous agents (drones, ground robots, embedded AI) fail when deployed because simulators are too clean and do not reflect real‑world constraints such as wind, sensor dropout, compute limits, and regulatory changes.

Outcome

The generative benchmarking loop produces realistic, adversarial scenarios that expose routing policy weaknesses early, ensures policies are evaluated under deployment‑level precision (e.g., INT8), and quantifies energy consumption per decision, guiding design toward the optimal trade‑off.

Workflow

• 1. Train a conditional VAE to learn the manifold of realistic UAM scenarios (wind fields, traffic density, sensor noise, regulatory masks).
• 2. Sample scenarios conditioned on difficulty from the VAE.
• 3. Run the routing policy on each sampled scenario on hardware‑mirrored compute and energy models.
• 4. Measure success rate, latency, and joules per decision.
• 5. Aggregate results to assess Pareto frontier between capability and consumption.
• 6. Iterate: refine scenario generator and policy based on continuous benchmarking feedback.

Evidence

The simulator I'd used was too clean, too generous, and utterly disconnected from the constraints of low‑power edge hardware.

Source

Every benchmark run must report joules per decision, not just success rate.

Source