NPC: Neural Predictive Control for Fuel-Efficient Autonomous Trucks

TL;DR

The paper tackles the challenge of achieving fuel efficiency in long-haul autonomous trucking without relying on precise physical vehicle models. It introduces Neural Predictive Control (NPC), a data-driven framework built around the NVFormer transformer to predict future engine states and fuel consumption from a combination of online past data primitives and knowledge-derived future data, guided by a fuel-saving optimizer. NVFormer surpasses LSTM, Transformer, and conventional PCC baselines in both offline prediction and closed-loop evaluations, achieving up to 3.45% fuel savings in open-road tests and 2.41% in simulations. The approach leverages Frenet-coordinate representations, anchor-based future sampling, and a hybrid encoder–decoder architecture to enable fuel-aware control for autonomous freight with practical deployment potential.

Abstract

Fuel efficiency is a crucial aspect of long-distance cargo transportation by oil-powered trucks that economize on costs and decrease carbon emissions. Current predictive control methods depend on an accurate model of vehicle dynamics and engine, including weight, drag coefficient, and the Brake-specific Fuel Consumption (BSFC) map of the engine. We propose a pure data-driven method, Neural Predictive Control (NPC), which does not use any physical model for the vehicle. After training with over 20,000 km of historical data, the novel proposed NVFormer implicitly models the relationship between vehicle dynamics, road slope, fuel consumption, and control commands using the attention mechanism. Based on the online sampled primitives from the past of the current freight trip and anchor-based future data synthesis, the NVFormer can infer optimal control command for reasonable fuel consumption. The physical model-free NPC outperforms the base PCC method with 2.41% and 3.45% more significant fuel saving in simulation and open-road highway testing, respectively.

Figures (5)

• Figure 1: The Future Data Sampler. The s-value of each anchor point refers to the position of the local maximum or minimum of the altitude curvature. The start point, anchor points, and end point together comprise the key points. Velocity curves are sampled within the speed bound of each anchor point.
• Figure 2: Two of our datasets routes. Route (a) is located in East China, Route (b) is in Southeast China, and (c) represents their altitude variations from north to south.
• Figure 3: The altitude curves of the 15 different simulation scenarios generated from the real-world map.
• Figure 4: An example of interpolating the fuel consumption at one scenario. We estimate the fuel consumption for each method by interpolating the simulated results at various target speeds from 19.44 m/s to 23.61 m/s.
• Figure 5: Result of NPC and PCC in part of open-road highway testing. This 8-km long clip covers 2 complete uphill and downhill processes, and one of these three descents is relatively more gradual.