ML applied to AV systems
Developing highly sophisticated Deep Neural Networks (DNNs) with the ability to safely operate an AV is a highly complex technical challenge. Practitioners require PB of real-world sensor data, hundreds of thousands, if not millions, of virtual Central Processing Unit (vCPU) hours, and thousands of accelerator chips or Graphics Processing Unit (GPU) hours to train these DNNs (also called models or algorithms). The end goal is to ensure these models can operate a vehicle autonomously safer than a human driver.
In this section, we’ll talk about what is involved in developing models relevant to end-to-end AV/ADAS development workflows on AWS.
Model development
AVs typically operate through five key processes, each of which may involve ML to various degrees:
- Localization and mapping
- Perception
- Prediction
- Planning
- Control
Each of the steps also requires different supporting data and infrastructure to efficiently produce...
