LiDAR-Based Steering Angle Prediction
A deep learning project utilizing PyTorch to predict autonomous vehicle steering angles by processing raw LiDAR sensor data from the MIT-Intel dataset.
Overview
This project focuses on a fundamental task in autonomous navigation: predicting vehicle control commands based on environmental perception. By utilizing the MIT-Intel dataset, which contains synchronized odometry and laser scan information from a mobile robot, I built a deep neural network capable of predicting the steering angle solely from raw 2D LiDAR scans.
Technical Highlights
- Sensor Data Synchronization: Developed a custom parser to handle the raw
intel.logfile, successfully aligning asynchronous time-series data streams (Odometry and LiDAR) to match sensor inputs with correct steering labels. - Data Preprocessing: Implemented a pipeline to normalize distances to a [0, 1] range and downsample LiDAR scans from 180 to 90 points, creating a computationally efficient input vector.
- Deep Learning Architecture: Designed a
RegularizedLiDARNetusing PyTorch. The architecture features multiple linear layers with ReLU activation and Dropouts to prevent overfitting on the noisy sensor data. - Hyperparameter Optimization: Conducted a systematic grid search to determine the optimal learning rate, batch size, and dropout probability, achieving a final test loss of 2.9374.
- Interoperability: The trained model was exported to ONNX format, ensuring cross-platform compatibility and optimized inference capabilities.
Tech Stack
- Core: Python, NumPy
- ML Framework: PyTorch (Training), ONNX (Inference/Export)
- Data Science: Scikit-learn, Matplotlib
- Dataset: MIT-Intel Mobile Robot dataset