# Tesla Robotaxi Fleet Runs Early FSD V15 as Expansion Awaits Safety Validation

Tesla's Robotaxi fleet is running an early version of FSD V15 that has incorporated about 40% of planned improvements, while expansion remains limited by strict safety validation, Elon Musk said on the company's Q1 2026 earnings call\.

By China EV Signal
Published: 2026-10-15T13:01:57.357Z
Updated: 2026-10-15T13:01:57.359Z
Canonical article: https://chinaevsignal.com/news/tesla-robotaxi-fsd-v15-expansion-safety-validation

## Key takeaways

- FSD V15 expands neural network parameters roughly 10-fold over the previous version, and the Robotaxi fleet version incorporates about 40% of planned improvements\.
- Nevada has approved Tesla to deploy up to 5,000 Robotaxis over the next 12 months\.
- On September 3, the production Cybercab opened for paid public rides in Austin, when the Robotaxi fleet had accumulated over 1 million miles of unsupervised operation, up from about 380,000 miles six weeks earlier\.

## Expansion Hinges on Safety Validation

Elon Musk said on Tesla's Q1 2026 earnings call that the real constraint on Robotaxi expansion is rigorous safety validation\. On the Q2 2026 earnings call, Tesla AI chief Ashok Elluswamy said the work required to launch a new city is decreasing, and Tesla wants to keep compressing that time, eventually moving from city-by-city rollouts to covering an entire state\.

The Cybercab has no steering wheel or pedals, so it cannot keep a safety monitor ready to take over as other Robotaxis can\.

Nevada has approved Tesla to deploy up to 5,000 Robotaxis over the next 12 months\.

## FSD V15 in the Robotaxi Fleet

Tesla AI chief Ashok Elluswamy said the Robotaxi fleet is already running an early version of FSD V15\. Seven planned major improvements are still in parallel development, and the fleet version has incorporated about 40% of the expected improvements\.

Elluswamy said on X that V15 focuses on earlier hazard prediction, faster reaction times, and overall better safety and collision avoidance\.

FSD V15 expands neural network parameters roughly 10-fold over the previous version\. FSD V14, by comparison, extended the context window from 8 seconds to 24 seconds and raised total parameters 4\.5 to 10 times over V13\.

## Preceding Versions and Architecture Shift

When FSD V12 was released in early 2024, Tesla compressed roughly 300,000 lines of autonomous-driving code to about 3,000 lines, replacing a modular perception-prediction-planning-control architecture with a single end-to-end neural network\.

Musk has predicted that a fully re-architected version could reach a safety level ten times that of a human driver\.

On the input side, seven cameras continuously capture high-resolution images, with an input context of about 2 billion tokens\. The output has only two degrees of freedom: the next steering and acceleration\.

Tesla uses multiple auxiliary tasks in joint training, with the model learning more than ten tasks including object detection, object tracking, scene geometry, and traffic sign understanding\.

## Cybercab Opens for Paid Rides in Austin

On September 3, the production Cybercab opened to paid public rides in Austin\. At that time, the Robotaxi fleet had accumulated more than 1 million miles of unsupervised operation, compared with about 380,000 miles six weeks earlier\.

In its FSD safety report, Tesla says FSD Supervised covers 15 countries across North America, Europe and Asia-Pacific and has accumulated more than 15\.5 billion miles\.

Tesla AI engineering director Phil Duan said that training specifically on clips where model output differs from human driver actions is far more efficient\.

## Data, Compute and Hardware

Tesla engineers' typical dataset contains more than 120 years of driving data and reaches hundreds of petabytes in scale\.

The Cortex 2 AI training cluster at the Texas Gigafactory is being expanded, with data center power capacity already exceeding 115 megawatts\. On-site compute more than doubled in the first half of the year\.

AI4 is sufficient to run V15 and unsupervised FSD, and Tesla is also advancing an interim compute hardware upgrade codenamed AI4\+, with roughly 10% more compute and about double the memory, reserving room for model iteration before AI5 launches\.

Former Tesla AI software engineering director Srihari Sampathkumar said in an interview on Silicon Valley 101 that the challenge of scaling lies in who has the right infrastructure to expand deployment faster\.

## End-to-End Model Architecture

Tesla describes an end-to-end foundation model for robotics that is trained with large amounts of data, has a large context, runs at 36 Hz, and directly produces control actions\.

The minimum requirement for driving is an extremely long context: an input context of 2 billion tokens made up of 7 cameras at 36 FPS and 5 megapixels with 30 seconds of history, divided into 5x5 pixel patches, plus navigation maps and route for the next few miles, 100 Hz kinematic data such as speed, IMU and odometry, and 48 KHz audio data, with the output being the next steering and acceleration\.

The model needs to learn the correct causal mapping from 2 billion tokens to 2 tokens\.

Auxiliary tasks usable for improving training efficiency include 3D geometry, 2D and 3D object detection, object tracking, semantic segmentation, instance segmentation, dense depth, scene and optical flow, OCR and sign understanding, image classification, scene understanding, visual question answering, localization and mapping, future video frame prediction, and image reconstruction\.

A neural network closed-loop simulator can be trained with cheap-to-collect state-action data\. Inputs at time t include camera videos, navigation maps, vehicle kinematics and audio, processed by a World Sim NN to produce the same categories of outputs at time t\+1\.

Tesla frames its recipe as more data and bigger models equaling better generalization\.

A chart of collected video clips shows a current cumulative clip count of 60\.7 million and 54\.5 years of video, against a maximum of 105 million clips and 148 years, across dates from 01/01 to 03/27\.

## Company coverage

- [Tesla news](https://chinaevsignal.com/manufacturers/tesla)

## Original publisher

汽车之心. Published: 2026-10-15T10:08:00.000Z