Transportation is responsible for a large slice of global carbon emissions, and most of that energy is wasted in the small decisions: braking a little too late, accelerating toward a light that is about to turn red, holding a speed that the road ahead does not reward. A connected and automated vehicle has information a human driver does not — the state of the traffic signal, the speed of the car three vehicles ahead, the grade of the hill coming up. The question my group has worked on is deceptively simple: given all that foresight, what is the most fuel-efficient way to drive?
The idea
We frame eco-driving as a stochastic optimal control problem and solve it with reinforcement learning tailored to the powertrain. In A deep reinforcement learning framework for eco-driving in connected and automated hybrid electric vehicles (2023), we learn a velocity and power-split policy that exploits vehicle-to-vehicle and vehicle-to-infrastructure signals. Because a mistake in a real vehicle is expensive — and unsafe — we followed up with Safe Model-based Off-policy Reinforcement Learning for Eco-Driving (2022), which keeps the learned policy inside hard safety and powertrain constraints. And in Real-Time Ecodriving Control … Using Approximate Dynamic Programming (2022) we showed the whole thing can run on-board, in real time, rather than as an offline planner. The same instinct shows up even at the routing level in Traffic-Aware Adaptive Routing for Minimizing Fuel Consumption (2019): sometimes the greenest route is not the fastest one.
Why it matters
The impact here is unusually concrete. These are not marginal academic gains — fuel and emissions are saved on every trip, with no new hardware, just better use of information the vehicle already has. Three things make the work matter beyond a single demo:
- It respects reality. Safety and actuator constraints are treated as hard limits, not soft penalties. That is the difference between a paper and something an automaker can ship.
- It runs in real time. A controller that needs a data center is useless in a car; one that runs on an embedded chip is deployable.
- It generalizes. Hybrid powertrains today, but the pattern — turn foresight into an optimized control policy — applies to trucks, drones, and buildings.
What it means going forward
We are entering a world of embodied AI: systems that don’t just predict but act in the physical world under energy and safety budgets. Eco-driving is a clean microcosm of that world. The lesson that transfers is that raw predictive power is not enough — the value comes from converting predictions into decisions that honor physical constraints and a real-time compute budget. As electric and autonomous fleets scale, control policies like these compound: a few percent per vehicle, across millions of vehicles, is a meaningful dent in the transport sector’s emissions.
Papers behind this post: A deep reinforcement learning framework for eco-driving in connected and automated hybrid electric vehicles (2023) · Safe Model-based Off-policy Reinforcement Learning for Eco-Driving (2022) · Real-Time Ecodriving Control in Electrified Connected and Autonomous Vehicles Using Approximate Dynamic Programming (2022) · Traffic-Aware Adaptive Routing for Minimizing Fuel Consumption (2019). See them on the Publications page, or the eco-driving project.