Tesla: Drive on cameras. Learn from the fleet.
Tesla removed radar and then ultrasonic sensors, betting that cameras and neural networks trained on its fleet could drive. Explore when fewer, cheaper inputs at scale beat more sensors, and what you give up in backup.
See the idea in one picture ↓- radar removed in May, North America
- 2021
- radar or ultrasonic sensors after 2022
- 0
- FSD (Supervised) miles, May 2026
- 10B
- paid robotaxi rides in Austin from June
- 2025
Dates from public reporting; mileage as reported by Tesla.
Independent educational analysis. Scenarios, people, numbers and illustrations on this page are original PMcademy practice material, not company research results or an endorsement.
From many sensors to cameras and a fleet
- 1Several sensors
Early cars combined radar, ultrasonic sensors and cameras.
- 2Cameras only
Tesla removed radar, then ultrasonic sensors, and relies on cameras and neural networks.
- 3The fleet teaches
Clips from many cars train the model, and the update returns to every car.
Animated illustration. Names, numbers and people are fictional teaching material.
More sensors feel safer. They also cost more on every unit.
Many sensors on a few cars
Cameras on every car, learning from the fleet
Read the analysis
Imagine you are building a machine that has to understand the world around it. Every extra sensor adds a way to see, and a cost on every unit you sell, and one more stream of data to combine. Fewer sensors make each unit cheaper, so more units ship and more of the world gets seen. The product question is whether you win on how much each unit can sense, or on how much the whole fleet can learn.
Cheaper units, more of them, more to learn from.
Illustrative data and simplified behaviour. This model does not represent a live company system or measured company results.
Read the analysis
Use the model below to compare three paths. Fit every car with many sensors, fit only cameras, or fit only cameras and learn from what the whole fleet sees. Watch the cost per car, how many cars you can equip, and how many rare situations the system has seen. The numbers in the model are illustrative.
Think about the person behind the wheel today.
“I want it to see what I see, especially when I am tired.”
Read the analysis
Our fictional learner, Marcus, drives forty minutes on the highway each way. He does not care which sensor sees the truck ahead. He cares that the car sees it every time, including in rain and at dusk, and that he knows when he must take over. Ask what your user needs to trust, and whether your design choice changes what they experience or only what it costs you.
Miles driven are not a safety rate.
Public reporting citing Tesla puts FSD (Supervised) at 10 billion miles by May 2026.
How often a person stepped in, or how it compares with human drivers on the same roads.
- Interventions per 1,000 miles
- Crashes per million miles by road type
- Performance in rain and at night
Read the analysis
Ten billion supervised miles shows how much the fleet has driven with the system on. It does not show how often a human stepped in, how it compares with human drivers on the same roads, or how it performs without supervision. In your project, separate how much your product is used from how well it does its job, and choose a rate you can compare fairly.
One kind of sensor means one kind of blind spot.
- Lower cost on every car
- One kind of sensor to master
- Data from the whole fleet
- No backup when cameras are blinded
- Harder to prove safety
- Doubt from regulators and the public
Read the analysis
Dropping radar and ultrasonic sensors removes cost and the work of combining different signals. It also removes a backup when cameras struggle, such as glare, fog or a dirty lens. Regulators, insurers and the public will ask for evidence that the system is safe without that backup. Decide what you will measure to show the gap is closed, and what the product does when it cannot see.
Add a sensor, add data or add a person?
Your fictional team builds a robot that restocks shelves in grocery stores. It sometimes misreads labels in poor light. You can add a second sensor to each robot, keep the camera and collect far more images from every store to improve the model, or keep the camera and route unclear cases to a remote person. Choose one and explain how you will know it worked.
Choose a next move
Leave with an inputs decision.
Your artifact
Write an inputs decision for one capability: inputs and their cost per unit, what you add or remove, the failure you fear most and its backup, and the rate you will track.
Critique my project artifact using only the evidence I provide. Write an inputs decision for one capability: inputs and their cost per unit, what you add or remove, the failure you fear most and its backup, and the rate you will track. Separate observed facts, assumptions, proposed tests and missing evidence. Identify a failure mode and a cheap check. Do not invent company results or user research. Ask for my artifact before starting.
Read the analysis
Pick one capability your product depends on. List the inputs it uses today and what each costs per user or per unit. Decide which to add, remove or replace with learning from usage. Write the failure case that worries you most, the backup for it, and the rate you will track to show the change is safe.