Anantha Kancherla — Building Level 5 Autonomous Vehicles

Infrastructure

Anantha Kancherla outlines the complex infrastructure required for developing autonomous vehicles at Lyft. The process involves high-performance computing systems, which can vary from multiple smaller computers to a single large one, each running different operating systems depending on their role 1. Anantha emphasizes the importance of balancing cloud costs, likening the cloud's potential financial pitfalls to handling "sharp knives" 2.

You can imagine multiple smaller computers or one large computer and one small computer. There's different configurations possible, and you just have to figure out how you're going to break down your workload.

--- Anantha Kancherla

Teams must also consider data and compute infrastructure, ensuring that testing and deployment are seamlessly integrated with cloud-based and on-road systems 1.

   

Testing

Testing and validation are critical in the development of autonomous vehicles, as Anantha Kancherla explains. The process involves rigorous integration testing to ensure models perform accurately in real-world scenarios, with a focus on metrics like safety and comfort 3. Anantha notes that improvements often stem from acquiring the right data rather than altering model architecture 3.

You have to run through the entire gamut. If you do something brand new.

--- Anantha Kancherla

The evolution of processes at Lyft has led to more structured frameworks, enabling teams to manage complex testing and validation tasks efficiently 4.

   

Process Evolution

As Lyft's autonomous vehicle team has grown, so have their processes, becoming more sophisticated and standardized. Anantha Kancherla describes the transition from informal methods to structured frameworks that accommodate larger teams and complex projects 4. This evolution includes the development of custom tools for tracking data and experiments, which are crucial for maintaining efficiency and accuracy.

One very underrated aspect of machine learning is that it has to be coupled with domain knowledge.

--- Anantha Kancherla

Anantha stresses the importance of domain knowledge in machine learning, highlighting that understanding the problem and data is more valuable than sheer data volume 5.