<![CDATA[Robotic vision enables machines to "see" and interpret the world around them. It can help with tasks such as obstacle avoidance, navigation, and object detection. This kind of work is well suited to modern deep learning but requires a lot of memory, power, and processing speed. Most of the small robots do not have a powerful computer. However, they have a low-cost microcontroller. This is a smaller computer. These devices are very resource-constrained. It becomes difficult to run regular deep learning models. This research work proposes to address the above issue with model quantization. When a model number is not a nice integer. It is quantized, i.e., rounded to a simpler integer with 8 bits. Such a reduction reduces memory consumption and increases the computing speed. This work is a comparison of two popular approaches. They are Post-Training Quantization (PTQ) and Quantization-Aware Training (QAT). Researchers construct a small and simple ConvNet that has only 28,069 parameters. It is examined on two popular microcontrollers: ARM Cortex M7 and ESP32. Both methods have been found to be effective. By reducing the use of the flash memory by approximately 71% and RAM by 60%, quantization cuts down memory usage. It doubles the processing speed and reduces power consumption by 35 to 40 percent. QAT retains 99% of the original model accuracy. The advantages of PTQ are that it is faster to set up, but with slightly less accuracy. The model performs well in real time on both platforms, with no assistance from the clouds. This work shows deep learning can run well on low-power devices. It helps make smart, low-cost robots more widely available.]]>
