Accelerate Autonomous Mobility in a Post-COVID-19 World with a Future-Proof Data Strategy

Autonomous mobility services, fueled by intensive R&D and improvements in autonomous driving technologies, have emerged as strong contenders to the future of automotive industry.

Pexels Truck

The COVID-19 pandemic has impacted human society and economic operations in ways like none other in recent history. All industry verticals and business models are being challenged to re-think and re-work on a strategy to emerge relatively unimpacted from the crisis. The automotive industry is no different. Global transportation and mobility needs are being pushed to re-structure themselves with the emergence of a New Normal in a post-COVID-19 world—with measures around social distancing, human safety and preventing the spread of this and other infectious diseases becoming a priority.

Autonomous mobility services, fueled by intensive R&D and improvements in autonomous driving technologies, have emerged as strong contenders to the future of automotive industry. These potential benefits don’t come without their own technological hurdles, however. A key technology problem that still needs to be addressed is the enhancement of a vehicles capability to perceive, think and respond like humans. This involves training computer algorithms to solve contextual vehicle awareness, perception of surroundings and path prediction challenge across multiple environments and covering the long tail of driving scenarios.

Delving deeper, these algorithms (commonly known as deep learning models) need an enormous amount of data to enhance a vehicle’s autonomous capabilities. The required datasets are sourced primarily through a fleet of OEM test vehicles (referred to as ego-vehicles) deploying multiple sensors—such as cameras, lidars, radars and ultrasound sensors—that generate terabytes of data per day for each individual test vehicle.

After data is gathered, it needs to be stored on in-vehicle recorders, transferred off-board, fused together to address spatial and temporal alignment, curated to identify objects (such as pedestrians, signals, surrounding traffic, road markings etc.) and then fed to deep learning algorithms for training and optimization.

As evident from above, solving the autonomous vehicle (AV) data challenge by building resilient AV data strategy will enable manufacturers to transform mobility in post-COVID-19 world. With that, there are four specific dimensions where automotive manufacturers will be able to make significant valuable impact.

Data pipeline: Improve efficiency in data generation, storage, transfer and management

Imagine a vehicle driving on a motorway, and a camera recording frames covering trees or sky images, capturing image frames which are of no use. Over time, perception and path planning algorithms will become smarter and would not need the recordings of commonly occurring scenarios on which algorithms have been already trained. Therefore, it is important that sensor devices or onboard loggers can intelligently filter the redundant data to optimize performance.

Organizations can transfer terabyte AV data from vehicle data loggers to off-board data processing platforms in one of the following two ways:

  • Online transfer. The in-vehicle recorders are transported to a local data center where the AV drive data is transferred to off-board processing platform using high-speed network connection.
  • Offline transfer. The in-vehicle recorders are transferred via offline logistic providers (courier services) to the off-board data processing platform.

Data management on target infrastructure should consider aspects around storage optimization for different kind of datasets, including the possibility of getting datasets ready for parallel high-performance compute read operations, enabling access to third parties (in case of data curation) or verification and validation tasks.

Data workflows: Orchestrate data-dependent tasks and improve data science team productivity

Data scientists and engineers delivering on the ‘brain’ behind AVs will not typically be very conversant with IT platforms and tools. One of the major problems faced by data scientists is the need to search for relevant datasets from the petabyte-scale AV data pool. To work around this, enabling custom metadata management and tagging drive scenes with weather and time-of-day metadata tags can improve searchability.

In addition to this, multiple data pipelines will need to exist including tasks ranging from deep learning training pipelines for perception and path planning, data labeling and ground truth generation, sensor data fusion and visualization, synthetic data management, verification and validation and more.

These pipelines will have different technology and operational requirements. These constraints are a direct reflection of the need for building a data platform foundation which can essentially decouple data, compute and deliver capabilities to address the requirements.

Operational costs: Optimal total cost of ownership 

As mentioned, data volumes required to manage an autonomous vehicle program exceeds multiple petabytes. With this kind of scale, costs can escalate if not managed properly—ultimately leading to program failures. For instance, a high-performance infrastructure with idle graphics processing units will build inefficiencies. Also, for a cloud-hosted model, egress costs to third parties or on-premise development centers can raise the operating costs as well. Running multiple sites, lack of process monitoring and ineffective continuous governance will amplify— constraints further adding up to program costs.

Data and process governance: Continuous monitoring, feedback and improvement

Data governance for an AV development program needs to widen the scope of a traditional governance model, with the key underlying principle being optimization and efficiency improvement. These should cover the measurement of variances in field-of-view sensor data and eliminating blind spots, the identification of driving sessions with limited or no value, the optimization of efficient driving dataset classification and automated meta-data label generation and more.

How do organizations adapt?

AVs have the potential to truly reach their disruptive potential in the post-COVID-19 world. Organizations with future-proof data strategies will be able to build a fully autonomous vehicle faster and create a formidable differentiator within the marketplace.

As a next step, automotive manufacturers who look to capitalize on this market opportunity should start by assessing their existing capabilities, preferably through an external industry partner, to identify capability white-spaces and develop a roadmap strategy to address them. They should look to contextualize the established capability blueprint and leverage ecosystem strengths and investments. Build an infrastructure with agility and resiliency at its core to scale alongside the rapid changes in available technology, including sensors, storage and computational power. Next, look to future-proof their operations by planning for the emerging regulatory standards and needs across different geographies and regions as they develop. Finally, optimize all operational efficiencies and work to reduce costs and justify business case investments.

It will be interesting to see how and when autonomous mobility becomes mainstream in the post-COVID-19 world. Industry leaders will take these necessary steps, but those that invest strategically and follow these specific guidelines will find themselves with a significant competitive advantage.