Drodee

▶ Live interactive demo — choose a scenario below and press Play.

Area Defense — density predictor — the defended object is a perimeter, not a point. The classical density predictor projects noisy headings to the boundary, counts projected crossings per sector, and tasks perimeter interceptors toward attackers in the densest predicted breach sector.

① Left — noisy sensor picture of the area ② Middle — density projection to perimeter sectors ③ Right — perimeter interceptors and breach accounting

Scenario t = 0.0 s

ground-truth view — color breach contributors, interceptor tasking, and outcomes · uncheck for the engine's-eye view (uniform tracks; no truth/outcome colors)

① SENSOR PICTURE

radar · RF · EO/IR — noisy, ambiguous

primary threat secondary threat decoy / candidate clutter all hostile ✕radar EO/IR camera RF bearing

② K=5 THREAT FUTURES

structured discrete prior · single-shot

primary threat futures secondary threat futures candidate futures arrival hypotheses — engine keeps all 5 alive; it hasn't committed

③ CVaR ROBUST ALLOCATION

hedged over the K futures

ring size = share of sensor budget (%) color = green · amber · red alert

See the swarm

One air picture from radar, RF, and electro-optical feeds, reconciled under noise, dropouts, and disagreement. No single sensor sees the truth; Drodee fuses them in real time.

Hedge the futures

A swarm has many next moves. Drodee keeps several plausible futures alive instead of betting on one brittle guess — so the plan is robust to the attack that would hurt most, not just the likely one.

Allocate under risk

Human-approved recommendations that send scarce defenders to where the perimeter is most likely to fail — tracking worst-sector breach, defender depletion, and cost-exchange. Decision support, never autonomous engagement.

Methodology

A high-level map of the research behind the demo.

Structured priors

Structured Boltzmann priors for generative models, with sampler temperature handled as part of the training system.

Hypothesis hedging

A discrete prior that keeps K futures distinct but correlated, so alternatives stay plausible instead of becoming unrelated guesses.

The demo applies that idea to decision support: noisy tracks become K credible threat futures, then a robust allocation layer hedges attention.

1. SenseNoisy radar, RF, and EO/IR evidence.

2. EncodeScene context for the prior.

3. HedgeK correlated future hypotheses.

4. AllocateRobust attention and alert posture.

Built on our own structured-prior, multi-hypothesis research (classical sampling; no quantum-advantage claim). Method details are kept internal.