EMATO: Energy-Model-Aware Trajectory Optimization for Autonomous Driving

TL;DR

This work addresses energy efficiency in autonomous driving trajectory planning by introducing EMATO, an online nonlinear programming framework that optimizes energy-aware trajectories using a differentiable energy model. The method jointly considers vehicle dynamics, road slope, and traffic predictions to guide Frenet-based polynomial trajectories toward lower fuel consumption, quantified through a calibrated fuel-rate surrogate and an energy-weighted objective $J$. Across Cruise Control, Pulse-and-Glide, Adaptive Cruise Control, and Frenet driving scenarios, EMATO demonstrates fuel-efficiency improvements ranging from $2.42\%$ to $49.97\%$ over traditional polynomial baselines, with online solve times around $0.01$–$0.04$ seconds on common hardware. By maintaining compatibility with existing planners and showing real-time feasibility, EMATO presents a practical path toward energy-aware autonomous driving and potential integration with Autoware and real-vehicle validation.

Abstract

Autonomous driving lacks strong proof of energy efficiency with the energy-model-agnostic trajectory planning. To achieve an energy consumption model-aware trajectory planning for autonomous driving, this study proposes an online nonlinear programming method that optimizes the polynomial trajectories generated by the Frenet polynomial method while considering both traffic trajectories and road slope prediction. This study further investigates how the energy model can be leveraged in different driving conditions to achieve higher energy efficiency. Case studies, quantitative studies, and ablation studies are conducted in a sedan and truck model to prove the effectiveness of the method.

Figures (6)

• Figure 1: EMATO framework in a Frenet highway system, it considers traffic and slope predictions and optimizes the sampled polynomial candidates with a differentiable energy model to achieve energy improvement.
• Figure 2: (A-B) Original engine power and fuel map. (C-D) Fitted differentiable energy model and optimized gear selection w.r.t. traction acceleration $a_t$ and vehicle speed $v$ for a light-duty 7-speed truck.
• Figure 3: Gray and light green lines denote overjerky and feasible quintic candidates. Green and red lines denote the "Energy" quintic candidate and the PnG trajectory produced by EMATO.
• Figure 4: (A) ACC simulation. (B) Simulated flat, rolling, and steep elevation profiles. (C) Case study results of quintic and EMATO methods in a HWFET cycle with a rolling road.
• Figure 5: Frenet highway simulation. (A-B) An example of applying EMATO to a "Speed" quintic trajectory (red), and the optimized trajectory waypoints are plotted in green dots.

Theorems & Definitions (6)

• Remark 1
• Remark 2
• Remark 3
• Remark 4
• Remark 5
• Remark 6