Air Compressor Control Optimization in Commercial Vehicles Using Reinforcement Learning
Fabian Schuppert,,∗, Raul Gorek,, Bodo Rosenhahn and Timo von Marcard
AI summary
Problem
Conventional rule-based air compressor controllers in commercial vehicles lack predictive capabilities, causing inefficient fuel consumption and excessive mechanical wear due to an inability to anticipate future driving conditions or leverage multivariate vehicle state correlations.
Approach
The authors propose a reinforcement learning controller trained on realistic synthetic CAN driving profiles generated by a tunable Hidden Semi-Markov Model, enabling predictive optimization of compressor operation during energetically favorable phases while minimizing switch frequency.
Key results
- Configurable HSMM generator produces scalable, realistic CAN driving profiles without full vehicle simulation
- PPO-trained RL agent outperforms rule-based baselines across free-air utilization, energy demand, and switch frequency
- Accurate reproduction of real-world braking dynamics and driving transitions for reliable training
- Demonstrates feasibility of RL for subsystem-level control and future contextual signal integration
Why it matters
This approach offers a scalable, data-efficient pathway to reduce fuel consumption and extend component lifespan in heavy-duty fleets, establishing a foundation for AI-driven vehicle energy management.
Abstract
Air compressors are critical for braking, suspen- sion, and auxiliary systems in heavy-duty commercial vehicles, but their operation increases fuel consumption and contributes to mechanical wear. Existing rule-based controllers cannot an- ticipate future driving conditions or leverage correlations across vehicle states, limiting efficiency. We propose a Reinforcement Learning (RL) approach for predictive compressor control, supported by a tunable Hidden Semi-Markov Model (HSMM) generator that produces realistic Controller Area Network (CAN) driving profiles. The generator reproduces braking dynamics that govern air consumption, enabling scalable RL training without reliance on extensive real-world data. Using Proximal Policy Optimization (PPO), the RL agent outperforms rule-based baselines by balancing two objectives: maximizing compressor operation during free-air phases—driving states such as downhill or deceleration where kinetic energy can power the compressor without fuel—and limiting switch frequency to ensure mechanical longevity. These results demonstrate the feasibility of RL for subsystem-level control and its potential as a foundation for future efficiency-oriented vehicle strategies.