Autonomous driving controllers with neuromorphic spiking neural networks

Area of Science:

• Artificial Intelligence
• Robotics
• Computational Neuroscience

Background:

• Autonomous driving systems rely on sophisticated control algorithms.
• Neuromorphic computing, inspired by the brain, offers energy-efficient computational frameworks.
• Spiking neural networks (SNNs) are a key component of neuromorphic control.

Purpose of the Study:

• To explore neuromorphic implementations of four prominent autonomous driving controllers.
• To evaluate and compare the performance of neuromorphic controllers against conventional CPU-based implementations.
• To provide guidelines for neuromorphic architecture design in autonomous control.

Main Methods:

• Physics-aware simulation framework used to implement and test controllers.
• Neuromorphic implementations of Pure-Pursuit, Stanley, PID, and Model Predictive Control (MPC) explored.
• Extensive evaluation with varying intrinsic parameters and comparison with CPU-based models.

Main Results:

• Neuromorphic models demonstrated competitive performance against conventional counterparts.
• Optimal performance achieved with 100-1,000 neurons for most models.
• Hybrid conventional and neuromorphic designs (e.g., MPC) show promise.
• Performance degradation observed in neuromorphic models at speeds above 15 m/s.

Conclusions:

• Neuromorphic control is a viable and energy-efficient approach for autonomous driving.
• Careful tuning of parameters and neuronal resources is crucial for neuromorphic controller optimization.
• Hybrid architectures offer a path to overcome limitations and enhance neuromorphic control systems.