Robo-Taxi Fleet Controller Nears Oracle Profit and Adapts to New Cities in 10 Episodes

Key Findings · At a Glance

• ▸On a 16-station New York network, the Graph-RL policy came within 1.6% of the oracle benchmark's performance.
• ▸Graph-RL cut rebalancing trip costs by more than 36% against learning approaches using conventional neural network architectures.
• ▸Meta-RL achieved similar performance to single-city training with 10 episodes instead of 10,000, a thousandfold gain in data efficiency.
• ▸In a calibrated Luxembourg traffic simulation, Graph-RL delivered profit within 2.3% of the MPC oracle and beat all heuristics.
• ▸Trained only in the coarse simulator, the policy lost 1.1% profit versus high-fidelity retraining, cutting training from 5 days to 17 hours.
• ▸Starting from an offline-trained policy avoided over 12,000 online steps that cost $98,265 in lost revenue.
• ▸Under a sudden demand shift, meta-RL beat single-city training by 50.3% and a forecast-driven MPC controller by 4.6%.

Opportunity

The case for robo-taxi fleets rests on utilisation. A centrally coordinated fleet can match passengers to vehicles and move idle cars toward where demand is about to appear. The authors argue that poor coordination squanders that advantage, producing longer passenger waits and congestion from empty vehicles driving around the network. Large-scale deployment, in their framing, depends on solving fleet coordination efficiently and at city scale.

The team, led from Politecnico di Torino with collaborators at the Technical University of Denmark, Google DeepMind, MIT and Stanford University, builds a controller that splits the problem in three. A linear program assigns idle vehicles to waiting passengers. A graph neural network, trained by reinforcement learning, decides what share of available vehicles each region should hold next. A second linear program converts that target into the cheapest set of rebalancing trips. They test it across three cities in a coarse macroscopic simulator and in a calibrated mesoscopic simulation of Luxembourg built on 24 hours of real traffic data, and release the code, datasets and simulators openly.

Why Fleet Coordination Has Not Yet Delivered

Each established approach fails in a different place. Rule-based heuristics are fast but rarely near-optimal. Model-predictive control can be optimal under perfectly predictable conditions, yet the full problem grows with the square of the number of stations multiplied by the planning horizon, which becomes intractable for real-time use, and it typically ignores uncertainty in demand and travel times. Pure reinforcement learning scales better but is sensitive to distribution shifts, needs heavy training data and struggles to enforce hard constraints.

Two gaps matter commercially. Most learned controllers are trained and tested within a single city, and the authors note that training a new controller from scratch for each city is costly and time-consuming. And most evaluations rely on simplified, low-fidelity traffic models, leaving open whether results survive realistic congestion.

Results

In the macroscopic tests, the graph-based policy came within 1.6% of the perfect-foresight oracle on a 16-station New York network and cut rebalancing costs by more than 36% relative to learning approaches built on multilayer perceptron or convolutional architectures. Across networks of 16 to 400 stations it computed decisions far faster than optimisation-based control, because each decision reduces to one pass through a neural network plus two convex solves.

Transfer is where the numbers become commercially interesting. Trained across several cities and tested on unseen ones, the meta-learning variant finished 4.3%, 1.8% and 5.6% short of the oracle in San Francisco, Brooklyn and Shenzhen West. Single-city training needs 10,000 episodes per city; meta-RL reached similar performance in 10, which the authors describe as "a 1000× improvement in data efficiency". When demand shifted suddenly in part of Queens, meta-RL outperformed single-city training by 50.3%, zero-shot and fine-tuned policies by 22.1% and 20.2%, and a forecast-driven MPC controller by 4.6%.

The offline results address a practical cost. Offline RL methods trained on static datasets generated by three behaviour policies outperformed those policies, whereas behaviour cloning was capped by them. Training from scratch needed over 12,000 online steps to catch a pretrained agent, incurring $98,265 in lost revenue and $1,029,162 in excess rebalancing costs in simulation.

In the Luxembourg simulation, with the city aggregated into ten regions during the 4–6pm peak, Graph-RL earned profit within 2.3% of the MPC oracle, served the highest demand, matched by the equal-distribution heuristic and MPC, and cut rebalancing costs by 29% against an equal-distribution heuristic. It kept 75% of the fleet active, against 69% for MPC, and delivered the lowest average passenger wait; the equal-distribution heuristic ran 80% utilisation yet produced waits 7.52% longer. A policy trained only in the coarse simulator lost just 1.1% profit against one retrained in the high-fidelity environment, with training time falling from 5 days to 17 hours.

The Practical Applications

Assessment

The authors are clear about what remains. Demand and fares are treated as exogenous, drawn from historical data, and the macroscopic tests assume travel times unaffected by the fleet. Generalisation also has limits: a policy trained on the afternoon peak lost up to 18% against the oracle at night, narrowing to 5.8% once night and morning data were added to training. Future work targets multi-agent settings, real-time data streams, equity and energy objectives, and battery and charger constraints for electric fleets.

For buyers, the central caution is that every result comes from simulation, including the calibrated Luxembourg scenario, and the high-fidelity tests cover one city, mainly at a ten-region aggregation. The oracle and heuristic baselines are well chosen, but no comparison is made against production dispatch systems. The open codebase and benchmarks are arguably as valuable as the headline numbers, since they give operators and vendors a shared basis for checking claims before committing to field trials.