Identifying Robot Collisions using Convolution Neural Network
Completion: May 2024
A Convolutional Neural Network allows for images to be processed in 3 dimensions (depth, height, width) instead of vectorized forms.
Dataset:
The pretrained ResNet-18 model was used as a CNN. This model was fed 1000 images of the robot in various collision states along with their collision status (image, label) pairs.
Training:
The model's classifier is trained on the 900 sample images generated earlier to update the weights and finetune it for this task.
Evaluation:
Since the model is now trained, the remaining 100 images are used to validate the accuracy of the model. The Learning Rate and Number of Epochs hyperparameters are modified during training maximize the accuracry of the model.
Tags: Robotics Computer Vision Finetuning Transfer Learning Pybullet Python PyTorch Columbia