Why AI-Based MaaS Is Poised To Be The Future Of Transportation
Shailesh Manjrekar. January 8, 2021
Shailesh Manjrekar is Global head of AI and strategic alliances at Weka.IO, driving AI strategy and business growth.
The mobility as a service (MaaS) market is at an inflection point and is set to expand from $4.7 billion in 2020 to a $70 billion market by 2030 at a 31.1% compound annual growth rate. This unprecedented growth is fueling the demand for a new set of requirements and new, modern architecture to support it.
MaaS is defined as integrating various forms of transport services into a single mobility service accessible on demand. It is a result of the disruption facing the transport sector. New technologies and customer expectations have shifted the mindset of industry players to one that aggregates various forms of transport, from public and private operators, and melds these into an aggregated system that offers the commuter a single application allowing them to select their preferred mode of transport.
In my role at Weka, I get to work with several ecosystem players, many of whom are our customers.
The MaaS Ecosystem
MaaS is more technology than it is a physical form of transport. It is an ecosystem of technologies that enable those players in the transport sector to adapt to this current disruption.
If we are to delve into this ecosystem, we will start to see that beyond the vehicle, there are onboard sensors, intelligent manufacturing, connected cars, autonomous vehicles, telematics such as in-car intelligence and assistance, safety, and security, as well as a layer of services that could extend to fleets, ride-sharing and the delivery of goods.
The common denominator here is that every aspect of this ecosystem is steeped in data or is itself data-driven, which increasingly leverages artificial intelligence (AI), machine learning (ML) and deep learning (DL) technologies and requires data center infrastructure often comprising of GPUs, high-speed networking and high-performance data lakes.
State Of The Union
A single company or service does not provide these solutions or services. It is instead an industry collective where different providers play at a different level of the MaaS stack. SAE International is a body that sets various standards and mandates the type of autonomy each of these manufacturers is trying to achieve, ranging from level 0 to level 5.
Most players are currently factored in at the SAE’s level 3, which speaks to conditional automation, where a driver is a necessity and must be ready to take control of a vehicle at any moment. The levels make an exception for no automation, driver assistance, partial automation, conditional automation, high automation and the nirvana, which is level 5 and pertains to full automation, where the vehicle can perform all functions and a “driver” merely has the option to control a vehicle when required. The reason for these standards and levels is to ensure that safety is always foremost.
But for MaaS to work, data is required to train the neural networks so the right inference is made at the edge. Again, at the heart of this is data, and the accuracy is only as good as the datasets you have on hand and the high-performance access to the data.
Edge To Core To Cloud Data Fabric — The Software-Defined Car
The building blocks to building a MaaS system is not as straightforward as it sounds. Current NAS solutions cannot act with the speed needed or even feed these “machines” with enough data to create the autonomy required.
Let us look at the software-defined car. Your edge, which is the autonomous vehicle itself, requires an edge aggregation, the core where neural net training is conducted, the cloud for economies of scale and a Kubernetes-orchestrated environment that supports the aggregation of containers. All of this relies on the data you capture, ingest into the system and label using pixel-level labeling called “semantic segmentation.”
Because of the parallelism created at the GPU layer, your traditional storage stack won’t support this. Now add data ingest, which can often be of the order of 2 petabytes per year from a single car, neural net training, and explainable AI. You will see that you need a next-generation storage layer that can support all these different pipeline requirements.
The training and edge inference are critical to the success of MaaS, which is where technology partnerships are crucial. This is why industry players actively working in this field are using SDKs and reference architectures to democratize these processes.
Challenges Of Building AI For Autonomous Vehicles
If we go back to the basics of building a MaaS model, we can summarize that safety remains nonnegotiable. So whatever system you create or deliver solutions off of, has to support:
- Scalable AI training, with several thousand images per seconds.
- Petabyte-scale AI testing.
- AI-based data selection and mining.
- Traceability of models (code and data).
- Seamless petabyte-scale data access.
- Workflow automation and integration for ingest, feature engineering, hyperparameter optimization, inference and life cycle management.
The Way Forward
Back to the data, which itself is just data if is not maintained. You have to be diligent in performing proper version control of datasets, models and experimentation. Only once all of these factors are put in place will you be able to say you have a production-grade functional AV platform.
MaaS requires not an evolution in thinking, but a revolution. It supports so many moving parts, but at its core is the ability to get actionable intelligence from data using AI/ML/DL, operationalization of your data at scale, and governance and explainability. If you are unable to rely on your existing systems to check these boxes and are unwilling to invest in modern applications and data center infrastructure to support it, chances are your autonomous vehicles will remain stuck in the parking bay — while your competitors are already moving to autonomous hovercrafts.
This article originally appeared on Forbes