01
Resource-constrained pickup-and-delivery
Each customer has a delivery node, a pickup node, and a processing window. Pickup cannot start until delivery completes — even across vehicles.
public apiready
v2.0 · ALNS + BRKGA · ISC license
Route optimization,
reimagined for
real fleets.
FleetPilot is a two-stage metaheuristic (ALNS + BRKGA) solver for resource-constrained pickup-and-delivery — built for the chaos of Indian logistics, tuned for the discipline of production systems. Time windows, multi-depot, traffic-aware, fully deterministic.
- ≥ 30k
- Population / solve
- BRKGA parallel islands
- 500
- ALNS iterations
- adaptive destroy / repair
- < 1 ms
- Incremental checks
- O(1) RouteLoad decoder
Solve preview20 customers · 3 vehicles · 2 routes shown
Vehicle 01Vehicle 02
iter 1,247 / 500 · best 453.15moptimizing
Feasible✓
Customers20
Runtime185 ms
synthetic-20c · makespan 453.15m · feasiblelive render
— Capabilities
Engineered for the parts of
routing that don’t fit on a slide.
Every feature is in the public API — no flag-flipping, no proprietary modules, no “enterprise edition”. The same solver that ships on npm is what your CI runs.
02
Time windows
Earliest / latest delivery and pickup per customer. Soft penalties keep the search honest; feasibility flags surface violations.
public apiready
03
Multi-depot
Vehicles may start and end at different depots. Auto-detected from your problem JSON — no flag-flipping required.
public apiready
04
Traffic-aware routing
Time-dependent travel times via a pluggable traffic model. Tune rush-hour multipliers per segment, simulate realistic congestion.
public apiready
05
Inter-vehicle transfers
Exchange resources at hub nodes. Concurrency limits, directional permissions, transfer durations — all configurable.
public apiready
06
Multi-objective
Pareto-optimal fronts across makespan, distance, cost, and CO₂. Compare solutions with `SolutionComparator`.
public apiready
07
Parallel solvers
Run ALNS and BRKGA concurrently via `worker_threads`. For evolutionary scale, enable island-model BRKGA with elite migration.
public apiready
08
Analytics & GIS export
Vehicle utilization, wait times, load profiles. GeoJSON, KML, and CSV — drop into QGIS, Google Earth, or Excel.
public apiready
09
Deterministic by default
Seeded mulberry32 RNG. Same problem, same solution — across runs, machines, and CI.
public apiready
— Workflow
From a map full of pins
to a dispatch-ready plan
in three moves.
The web console lives at frontend/ — but the solver behind it is the same npm package you can call from a backend in a single line.
- step 01
Define the problem
Drop depots and customer stops on a map. Add customers with delivery + pickup nodes, processing times, optional time windows. Configure vehicles with capacity, cost, and CO₂ per km.
- step 02
Solve with the metaheuristic
Call `solve()` once. ALNS adapts destroy / repair operators across the search; BRKGA evolves a 4n-gene chromosome, warm-started from the ALNS solution. Parallel islands via worker threads.
- step 03
Simulate and ship
Replay the solved routes on a map. Watch vehicles move, ETAs tick, deliveries flow into pickups. Export to GeoJSON, KML, or CSV — drop straight into QGIS or Google Earth.
— Under the hood
Two stages, one winner.
FleetPilot pairs Adaptive Large Neighborhood Search with a Biased Random-Key Genetic Algorithm — each stage sharpening the other.
Stage 01
ALNSAdaptive Large Neighborhood Search
Destroy / repair metaheuristic. Six destroy operators (Shaw, random, worst, route, proximity, string) and four repair operators (greedy, regret-2, regret-3, sequential). Operator weights adapt every segment via reinforcement learning.
- Simulated annealingWorsening moves accepted with exp(Δ/T); T decays geometrically.
- Multi-restartUp to 3 restarts on stagnation; temperature halved, weights zeroed.
- Adaptive sizingRemoval fraction grows 10% → 45% as stagnation deepens.
Stage 02
BRKGABiased Random-Key Genetic Algorithm
Evolutionary search over a 4n-gene chromosome (priorities, assignments, dependencies, transfers). Each child inherits each gene from the elite parent with probability 0.7.
- Warm-startALNS seeds 15% of the BRKGA initial population.
- Island modelMulti-population parallel BRKGA with elite migration via worker_threads.
- Stagnation resistanceElite mutation, adaptive mutant injection, immigrant replacement.
ALNS acceptance
P = exp((c − c′) / T)
Simulated annealing
ALNS weights
wᵢ ← (1−λ)wᵢ + λ·(sᵢ/uᵢ)
λ = 0.1, segment size 50
BRKGA crossover
child[i] = elite[i] w.p. 0.7
Biased toward elite
Decoder
O(1) RouteLoad
Incremental capacity check
— Performance
Numbers from a real
CI smoke run.
Reduced config (alns 50 · pop 100 · gen 50) keeps the smoke under 30 seconds. Production defaults run 30k × 20k and converge deeper.
8
Benchmark families
350+
Tests · 85% coverage
< 10s
50-cust synthetic
100%
Smoke feasibility
benchmarks/results/smoke-results.json
8 rows| Family | Instance | Customers | Vehicles | Makespan | Runtime | Feasible |
|---|---|---|---|---|---|---|
| synthetic | synth-10c-small.json | 10 | 2 | 379.24m | 41 ms | feasible |
| synthetic | synth-20c-medium.json | 20 | 3 | 453.15m | 185 ms | feasible |
| cordeau | mdvrp-2d-16c.json | 16 | 4 | 153.99m | 105 ms | feasible |
| cordeau | mdvrp-3d-24c.json | 24 | 6 | 233.93m | 274 ms | feasible |
| cordeau | mdvrp-3d-48c.json | 48 | 6 | 497.79m | 1.36 s | feasible |
| darp | darp-12req-4veh.json | 12 | 4 | 250.64m | 68 ms | feasible |
| salhi-nagy | vrpb-30c.json | 30 | 3 | 357.25m | 864 ms | feasible |
| lilim | lc1_2_1.txt | 106 | 50 | 631.22m | 8.89 s | feasible |
Reproduce locally with
npm run test:coverage — full suite in benchmarks/.— API
One solver.
Three surfaces.
Call it from TypeScript, run it from the CLI, or feed it a JSON file. Same engine, deterministic seed, predictable behavior across Node and the browser.
- Node.jsPrimary runtime. worker_threads for parallelism.
- BrowserWeb Worker bundle at fleetpilot/worker.
- CLIfleetpilot binary in PATH after global install.
typescript
~23 linesvar(--accent)]">import {
FleetPilotSolver,
Problem,
LocationNode,
Customer,
Vehicle,
} var(--accent)]">from var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">'fleetpilot';
var(--accent)]">const nodes = {
var(--color-iris)]">0: var(--accent)]">new LocationNode(var(--color-iris)]">0, var(--color-iris)]">28.61, var(--color-iris)]">77.23, var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">'Delhi Depot'),
var(--color-iris)]">1: var(--accent)]">new LocationNode(var(--color-iris)]">1, var(--color-iris)]">28.54, var(--color-iris)]">77.20, var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">'Customer A · Drop'),
var(--color-iris)]">2: var(--accent)]">new LocationNode(var(--color-iris)]">2, var(--color-iris)]">28.56, var(--color-iris)]">77.25, var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">'Customer A · Pick'),
};
var(--accent)]">const customers = [var(--accent)]">new Customer(var(--color-iris)]">1, var(--color-iris)]">1, var(--color-iris)]">2, var(--color-iris)]">50)];
var(--accent)]">const vehicles = [var(--accent)]">new Vehicle(var(--color-iris)]">1, var(--color-iris)]">5)];
var(--accent)]">const problem = var(--accent)]">new Problem(nodes, customers, vehicles, var(--color-iris)]">0);
var(--accent)]">const solver = var(--accent)]">new FleetPilotSolver(problem);
var(--accent)]">const solution = var(--accent)]">await solver.var(--color-aurora)]">solve({ maxTimeMs: var(--color-iris)]">30_000 });
console.var(--color-aurora)]">log(var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">`Best makespan: ${solution.var(--color-aurora)]">makespan.var(--color-aurora)]">toFixed(var(--color-iris)]">2)} min`);
console.var(--color-aurora)]">log(var(--accent)]">class="text-[var(--accent)]">var(--color-amber-glow)]">`Feasible: ${solution.var(--color-aurora)]">isFeasible()}`);— Ship it
Stop guessing routes.
Start solving them.
One npm install puts a production-grade metaheuristic in your backend. Bring your problem JSON, get a feasible plan back in seconds — with full GIS export and analytics out of the box.