An Efficient Pick-and-Place Pipeline based on Quantized Segmentation

Pick-and-place tasks in unconstrained environments use two or three deep learning methods to execute an object grasp, thus requiring highly capable computation devices. We propose a new pipeline that solves the pick-and-place tasks in unconstrained environments. Our pipeline only uses one deep learning method to make object grasping possible. It leverages quantization methods to speed up its inference time and improve its memory efficiency. We provide multiple experiments that compare the different methods and provide results for using this pipeline in the real-world. A new concept named Hybrid-QaT is introduced. Hybrid-QaT uses GPUs' power to speed up the neural networks' converging initially. Then in the last training epochs, it starts the quantization-aware training process to fine-tune the neural network's weights in an 8-bit integer representation. After the training, the conversion of the method to quantize can be done without losing accuracy.