Autopilot: Reduce Cost through Data Storage

A long and costly process of achieving autonomous driving is needed.

On the one hand, the difficulty of making a profit has prolonged the expected time to achieve full autonomous driving, while on the other hand, research and development investments are consuming a large amount of funds. General Motors’ autonomous driving company Cruise lost $1.2 billion in 2021, had a pre-tax loss of $1.9 billion in 2022, and lost $560 million in the first quarter of this year, equivalent to an average daily loss of $6.22 million.

One of the reasons for the continuous burning of money is the increasing development costs associated with commercialization. Take data storage as an example, in 2016, Intel estimated that each autonomous vehicle generates 4,000GB of data per day, the volume of which would cost  about $350,000 for their storage for one year according to the current charging standards of the Amazon cloud platform. As the number of test/operating vehicles increases, storage costs are rising accordingly.

The high cost of data storage has become a pain point for self-driving car companies. In order to achieve cost reduction and efficiency, many of the companies are targeting data storage as one of the ways to reduce costs.

Google’s Waymo company claims that at this stage, it focuses more on the quality of self-driving data rather than its quantity. It sets strict data quotas for the computing infrastructure team responsible for data collection and processing, retaining only data with storage value and only saving newly collected data. Andrew Chatham, an engineer responsible for Waymo’s computing infrastructure, said, “If we treat storage as unlimited, the cost will be astronomical.”

“It (Deciding which data to keep) turns out to be a business decision,” said Chatham. To save on data storage, the team needs to consider which is more valuable to the study, data collected in rainy or snowy weather conditions? At last, the team decided that data collected in snowy days was more valuable for storage. In addition, they found that the data related to rainy days included parking data, which the team deemed unnecessary.

Waymo is not alone.

Balajee Kannan, vice president of driverless technology provider Motional, said that data related to very incidental objects, like bicycles with surfboards, would not be of much value.

Cruise says it does not store all its data in the cloud because very little is useful. Less than 1% of the data collected in San Francisco, for example, is valuable. As the number of vehicles in operation continues to grow, Cruise is working to improve its data storage system to make it easier and cheaper to get autonomous driving off the ground.

Screening data is not the only solution to reducing data storage costs, in addition to which, tiered storage of data, i.e. storing less frequently accessed data on a suitably performing and less costly infrastructure, is also the way to go.

Self-driving startup Aurora captures and classifies data generated by self-driving trucks, and engineers tagging critical data (e.g. recent dangerous accidents) for regular storage and less frequently used data being moved to less frequently accessed storage areas, and the less frequently used data being sorted and deleted after three months. Only about 15% of the data is said to be located in the most frequently accessed storage areas.

No one knows how much more money will be needed for the game of autonomous driving. The increasing scale of test/operational vehicles, the growing number of more advanced sensors on board and potentially tighter budgets are going to impose self-driving car developers to be even more “picky” about the data on the servers.

Translator：Wei Xiong                                            

Reviser：Yan Luo