An autonomous vehicle generates roughly four terabytes of data a day. A test vehicle in active development can produce more than two hundred terabytes in a single eight-hour shift, and a wave of ten of them can put down close to two petabytes before the engineers go home for the night. That is the strange truth of autonomous vehicle data collection: the cars are extraordinarily good at collecting data, and the industry is still working out how to move, store, and make sense of it. This guide covers what self-driving cars collect, how much, how the pipeline works, and why the collecting was never the hard part.
What is autonomous vehicle data collection?
Autonomous vehicle data collection is the process of capturing sensor data from a self-driving car, from cameras, LiDAR, radar, and other sensors, to let the vehicle perceive its surroundings in real time and to build the datasets that train its machine learning models. It is the fuel of autonomy: every driving decision and every model improvement traces back to data the vehicle gathered.
What data do autonomous vehicles collect?
A self-driving car carries a stack of sensors that each see the world differently. Cameras capture high-resolution images for reading signs, lane markings, and traffic lights. LiDAR sweeps the surroundings into a dense point cloud used for object detection, high-definition maps, and precise localization. Radar measures the speed and distance of other objects and works when cameras struggle, in rain, fog, or darkness. Ultrasonic sensors handle close-range tasks like parking, and GPS and inertial measurement units track position and motion. The vehicle fuses all of it, hundreds of megabytes per second at peak, into a single moving picture of the road.
How much data does an autonomous vehicle generate?
The widely cited figure is around four terabytes per day for a self-driving car in normal operation, equal to the mobile data of roughly three thousand people. That is the modest end. A development vehicle bristling with high-resolution sensors for testing can generate over two hundred terabytes in an eight-hour shift, and a small test fleet pushes into petabytes a day. LiDAR alone can stream tens of megabits per second, and cameras produce tens of megabytes per frame, so the totals climb fast. The number that matters is not the headline volume but the fact that almost none of it can be moved off the vehicle in real time.
How autonomous vehicle data collection works: the data pipeline
The data feeds a loop that the industry calls the data engine. The vehicle collects raw sensor data during operation. That data is ingested, often physically, because development fleets still offload drives onto SSDs and carry them to an ingest station rather than upload them. Engineers then curate it, selecting the clips worth keeping, the edge cases and failures, and discarding the endless hours of uneventful highway. The kept data is labeled, with objects, lanes, and behaviors annotated so a model can learn from it. The labeled data trains a new model, the model is validated and deployed back to the fleet, and the fleet collects again. The quality of that loop, not the size of it, is what separates the programs that improve from the ones that drown.
The challenges of autonomous vehicle data collection
The challenges all stem from volume. Storage is the obvious one: a development car needs a server’s worth of flash storage in the trunk, and a program needs a data center to hold what the fleet brings back. Transmission is worse, because four terabytes a day, let alone two hundred, cannot travel over a cellular link, which is why physical drives are still the dominant ingest method in 2026. Curation is the quiet bottleneck: most of what a car records is uneventful and worthless for training, so the real skill is finding the rare clip that teaches the model something, and doing that across petabytes is its own engineering problem.
Privacy is the challenge that does not shrink with better hardware. Autonomous vehicle data is dense with personal information, faces, license plates, and precise locations of everyone the car drove past, which makes governance a legal requirement rather than a nicety. Masking and anonymizing that data, and enforcing where it can be stored, has to happen early, ideally before the raw footage ever leaves the vehicle.
The answer the industry already reached: reduce at the source
The most useful conclusion in autonomous vehicle data collection is one the industry came to years ago: collecting everything and sorting it out later does not scale. It is tolerable for a handful of research cars and impossible for a fleet, let alone the millions of vehicles a production program implies. The fix, as analysts at DXC and others have put it, is to do smart data reduction in the car, deciding what is worth keeping at the moment of capture rather than after a petabyte has piled up.
That is a data-infrastructure problem, and it is the one Expanso solves. Running a programmable layer at the edge, inside the vehicle or the ingest gateway, lets a team filter, select, and govern sensor data at the source, keeping the edge cases and dropping the noise, masking sensitive footage before it moves, and turning a firehose into a curated stream that the labeling and training pipeline can actually use.
The principle is already proven at fleet scale, even short of full autonomy. One automaker processing telemetry across 2.3 million connected vehicles cut the data it transmitted by 94% by filtering at the source, and autonomous vehicles, which produce orders of magnitude more, are simply the extreme case of the same problem. Expanso is not an autonomy stack, a perception system, or a labeling tool; it is the automotive data layer that decides what is worth moving before storage and bandwidth bills are paid on data no one will ever train on.
Frequently asked questions
How much data does an autonomous vehicle generate per day? A self-driving car in normal operation generates around four terabytes per day. Development and test vehicles, which carry more and higher-resolution sensors, can produce over two hundred terabytes in a single eight-hour shift, and small test fleets reach petabytes per day.
What sensors do autonomous vehicles use to collect data? The main sensors are cameras for visual detail, LiDAR for 3D point clouds and mapping, radar for speed and distance in poor visibility, ultrasonic sensors for close range, and GPS and inertial measurement units for position and motion. The vehicle fuses all of them into one model of its surroundings.
Why is autonomous vehicle data collection so difficult? The difficulty is volume. The data is too large to transmit over normal networks, expensive to store, and mostly uneventful, so the hard work is selecting the small fraction worth keeping and labeling. Privacy adds another layer, since the data captures faces, plates, and locations that must be governed.
What is a data engine in autonomous driving? A data engine is the loop that powers AV development: collect sensor data, curate the valuable parts, label them, train a new model, validate and deploy it to the fleet, and collect again. The strength of that loop determines how fast an autonomous system improves.
How do companies reduce autonomous vehicle data volumes? By processing at the source. Filtering, deduplicating, and selecting data inside the vehicle or at an ingest gateway, rather than uploading everything, keeps the valuable clips and discards the noise, which cuts storage and bandwidth cost and speeds up the training pipeline.
The car was never the bottleneck
Autonomous vehicle data collection is a solved problem in the sense that the sensors work and the cars record everything around them in exquisite detail. The unsolved problem is what to do with the result, because a fleet that records four terabytes a day per car cannot move, store, or label all of it, and trying to is how programs stall. The teams that get to scale will be the ones that decide what is worth keeping in the vehicle, at the moment of capture, instead of paying to haul the whole road home.
Drowning in sensor data from a connected or autonomous fleet? See how Expanso processes automotive data at the edge. The car collects everything. The advantage is in keeping only what matters. Or book a data consultation to talk through your data pipeline.
