Developing an AI.SDK

How do you safely deploy complex AI into safety-critical automotive systems? It requires separating cloud development from in-car execution.

#1about 3 minutes

The complexity of AI in safety-critical automotive software

AI introduces a new dimension of complexity to automotive software, requiring both an AI.SDK for development and an AI Runtime for in-car execution.

#2about 5 minutes

Overview of the data-driven development lifecycle for cars

The end-to-end machine learning loop involves cloud-based data processing, training, and optimization, followed by in-car deployment, inference, and monitoring.

#3about 7 minutes

The role of the AI runtime in the VW.OS

The AI Runtime Environment abstracts hardware and manages optimized model inference within the centralized Volkswagen Operating System (VW.OS).

#4about 1 minute

Comparing platform-dependent and independent model deployment strategies

Models can be deployed as a standard format like ONNX for on-device compilation or as a pre-compiled binary from the cloud for direct execution.

#5about 3 minutes

Why a unified AI.SDK is essential for automotive development

An AI.SDK provides a standardized and abstracted way to develop applications, tackling the challenges of AI safety and a heterogeneous hardware landscape.

#6about 6 minutes

Standardizing data preparation and management in the AI.SDK

The data preparation component of the SDK standardizes pre-processing, ensures data consistency, and enriches metadata to enable traceability and active learning.

#7about 4 minutes

Evaluating model performance and robustness with dedicated libraries

The AI.SDK includes components for performance evaluation and adversarial robustness checks, using a dedicated DNN test metric library for standardization.

#8about 2 minutes

Productionizing models through compression and hardware-aware optimization

The productionization step uses techniques like compression, quantization, and neural architecture search to reduce model size and improve inference time on target hardware.

#9about 6 minutes

Skills and challenges of working with automotive AI

Working in automotive AI requires a mix of software, hardware, and statistics skills to tackle challenges like massive data volumes and embedded system constraints.

#10about 2 minutes