How We Used Elman Networks for Path Planning
The project aims to simulate real-time decision-making in robotic agents using Recurrent Neural Networks. With a total of 182 input features (180 lidar + 2 position), the RNN predicts direction commands.
We used:
- Sigmoid and ReLU for activations
- Custom recurrent layer at third level
- Softmax-based output layer
The training data is generated using a function createTrainingSet that simulates various obstacle maps and paths. We used L_layer_model to implement training with backpropagation and Elman state tracking.
The model is evaluated using the testing function, which simulates robot movement, visualizes decisions, and ensures path feasibility. It performs well in most scenarios with the occasional loop or collision in high-complexity maps.
Overall, the model achieved 58% accuracy across 150 validation images, with category A and C performing well. Category B struggled due to overlapping characteristics with other categories.
This project demonstrates the potential of using Elman RNNs for real-time pathfinding in robotics, paving the way for more advanced AI applications in navigation.