Connected and automated vehicles can see far beyond their own bumper. By fusing vehicle-to-vehicle (V2V) and vehicle-to-infrastructure (V2I) signals — the speed of the car ahead, the timing of the next traffic light — a controller can plan a velocity profile that reaches the destination on time while burning far less fuel.

We develop real-time approximate dynamic programming and safe model-based reinforcement learning methods for this problem in hybrid-electric powertrains. The core challenge is computational: the value function lives over continuous state and action spaces with hard constraints, so we design rollout and warm-start schemes that re-solve the perturbed program fast enough to run on-board.

Highlights

• A deep reinforcement learning framework for eco-driving in connected and automated hybrid-electric vehicles.
• Safe, model-based off-policy RL that respects powertrain and safety constraints.
• Approximate dynamic programming that warm-starts from a previously solved program to meet real-time budgets.

See the related papers on the Publications page (filter by Journal).