Bearbeiten

Freigeben über


Autonomous Vehicle Operations (AVOps) design guide

This article provides an overview of the stages, architecture, and challenges involved with creating a back end to enable an autonomous vehicle solution at scale. To learn more about the information here, technology recommendations, and partner and open-source solutions for specific areas like simulation and data models, see the Automated Vehicle Operations solution idea.

Autonomous Vehicle Operations (AVOps) typically require a substantial amount of storage and compute to:

  • Capture and process data and scenes from test vehicles, as learning material for the perception models that vehicles need to drive autonomously.
  • Train perception models to recognize an environment, as the base functionality for driving autonomously.
  • Perform safety validation based on open-loop and closed-loop simulations.

Key stages

The development of an autonomous driving (AD) solution typically involves three key stages:

  • Ingest and curate data. Collection and refinement of carefully chosen data sets for advanced driver-assistance system / autonomous vehicle (ADAS/AV) development.
  • Iteratively test, train, and simulate. Simulation and training of ADAS/AV models across numerous ground truth scenarios.
  • Build and validate. In-vehicle software verification and validation with connected vehicles.

AVOps implements an autonomous driving development lifecycle:

Diagram that shows the autonomous driving development lifecycle.

Elements of the architecture

The AVOps architecture consists of the following four main elements. The next article in this series describes these elements in more detail.

  • DataOps. Ingest measurement data (videos, images, lidar, and radar), curate and extract data, and label data.
  • MLOps. Train algorithms, like perception models and post-perception models.
  • ValOps. Validate autonomous driving functions based on trained models and raw ground truth data.
  • AVOps centralized functions. Provide overarching functionalities like metadata search, data catalog, overall orchestration, platform governance, and standardized infrastructure templates.

Diagram that shows the elements of an AVOps architecture.

Challenges

  • Data collection. Collecting and analyzing large amounts of data to identify patterns and improve the vehicle's performance over time. Most of the costs of autonomous vehicle development are spent on data management and testing.
  • Data management. Handling the large amounts of data generated by vehicle sensors and systems, and determining which data is useful.
  • Scenario coverage. Ensuring that the OEM has tested the vehicle in a range of scenarios, including different weather, lighting, and road conditions.
  • Complexity. Managing the large and diverse set of algorithms and systems that are required for autonomous operation.
  • Verification and validation. Thoroughly testing the software to ensure that it behaves as expected in a range of scenarios and environments.
  • Data availability. Sharing data. Globally dispersed teams and third parties make sharing a challenge.

AVOps allows organizations to take advantage of the scalability, flexibility, and cost-effectiveness of cloud-based infrastructure and reduces the time to market for automated vehicles.

Contributors

This article is maintained by Microsoft. It was originally written by the following contributors.

Principal authors:

Other contributors:

To see nonpublic LinkedIn profiles, sign in to LinkedIn.

Next steps

To learn more about the information here, technology recommendations, and partner and open-source solutions for specific areas like simulation and data models, see the solution idea:

You might also find the following resources useful: