New AI Fixes Electric Carpooling So Fewer Rides Get Left Behind

The paper tackles a simple-sounding but messy problem: running a shared, electric taxi service. Each car must pick up several strangers, drop them off in the right order, keep enough battery, and find a charger that isn't already taken. Fix one thing — say, move a car closer to its next pickup — and you can quietly ruin something else later. To study this, the researchers built a computer simulation of a city where electric ride-pool cars serve passengers and line up for plugs on a first-come, first-served basis.

Their idea is a 'repair layer,' best pictured as a helpful supervisor watching the dispatcher. When a passenger ends up unassigned — the ride-pool version of being left standing at the curb — the supervisor moves them to a different car. Every move has to obey the rules: the route must still work, the battery must hold, the time windows must fit, and the trip must get cheaper right away. A second learning step, trained by trial and error, guesses which few moves are worth a closer look, so the system doesn't waste time checking thousands of bad options.

In tests on 12 Manhattan-style scenarios with 100 ride requests and 30 electric cars, the repair layer cut operating costs by 6.57% and completed 74.92 rides instead of 72.75 — wins in all 12 runs. Being smart about charger queues cut charging wait time by 41.0%, and electricity used per completed ride fell 1.73%. The method also held up at much bigger scale: 8,000 requests and 2,400 cars.

The catch: this is a simulation, not a real city. Traffic jams, bad weather, broken chargers, drivers quitting mid-shift and passengers who never show up are simplified or missing entirely. So the gains are a strong hint, not a promise — the next step would be testing this with an actual ride-share operator.