Lightweight Real-Time Navigation for Autonomous Driving Using TinyML and Few-Shot Learning.

Area of Science:

• Artificial Intelligence
• Robotics
• Computer Vision

Background:

• Autonomous vehicle navigation demands low-latency, energy-efficient machine learning models for dynamic, resource-constrained edge devices.
• Traditional deep learning models often exceed the computational and memory limits of embedded systems.
• Real-time adaptation to novel driving scenarios with minimal data is a significant challenge.

Purpose of the Study:

• To develop a unified TinyML-optimized navigation framework for real-time autonomous decision-making on edge devices.
• To enable rapid adaptation to unseen driving scenarios using metric-based few-shot learning with minimal data.
• To significantly reduce memory footprint and inference latency through edge-aware optimization techniques.

Main Methods:

• Integration of a lightweight convolutional feature extractor (MobileNetV2) with a metric-based few-shot learning classifier.
• Implementation of an end-to-end pipeline combining feature extraction, few-shot generalization, and edge-aware optimization.
• Application of post-training quantization and structured pruning for model compression and efficiency enhancement.

Main Results:

• The proposed framework achieved 93.4% accuracy on unseen road conditions.
• Demonstrated an average inference latency of 68 ms and memory usage of 18 MB.
• Outperformed conventional Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models in efficiency while maintaining competitive performance.

Conclusions:

• The TinyML-optimized navigation framework effectively enables scalable, real-time autonomous decision-making on low-power edge devices.
• The approach successfully balances predictive performance with significant reductions in latency and memory requirements.
• This work presents a viable solution for deploying advanced AI navigation capabilities in resource-limited autonomous systems.