Tesla Full Self-Driving: Technology, Versions, and the Path to Autonomy
Tesla's FSD software powers the Robotaxi fleet and millions of supervised vehicles. Here's how the vision-first, end-to-end neural net approach works — and where it's headed.
What is Tesla Full Self-Driving?
Tesla Full Self-Driving (FSD) is Tesla's autonomous driving software, designed to navigate urban streets, highways, intersections, and complex traffic scenarios without human input. It is the software stack that powers both Tesla's supervised FSD feature (available to consumer Tesla owners) and the unsupervised Robotaxi service operating across U.S. cities.
As of September 2026, the current production version is FSD v15, described across multiple analysis channels as a major upgrade over previous versions. Supervised FSD has accumulated 8–10+ billion miles of real-world driving data. The unsupervised Robotaxi fleet has reached 1 million miles with no safety driver.
The Vision-First Approach
Tesla's approach to autonomy is fundamentally different from competitors like Waymo. Tesla uses cameras only — no LiDAR, no radar, no ultrasonic sensors on the Cybercab. The rationale is both economic and philosophical: cameras are cheap, and if a vision system can match or exceed human perception (which relies on two cameras — our eyes), then adding expensive sensors creates unnecessary cost that prevents scaling.
Critically, Tesla does not use HD maps — pre-built, centimeter-accurate 3D maps of every road. HD-map-based systems require expensive mapping campaigns and struggle when road conditions change. Tesla's FSD navigates entirely from real-time camera input, making it immediately deployable anywhere the neural networks have enough training data for.
This vision-first philosophy is why Tesla can deploy FSD to millions of consumer vehicles simultaneously and why the Cybercab can operate without the sensor costs that make competitor vehicles dramatically more expensive.
End-to-End Neural Networks
Tesla's FSD software uses end-to-end neural networks — a paradigm where the system learns to map raw camera input directly to steering, braking, and acceleration outputs, rather than relying on hand-coded rules for each driving scenario. The network is trained on massive datasets of real-world driving, including millions of edge cases and disengagement events.
This approach improves continuously. Every mile driven by a Tesla — whether supervised or unsupervised — generates training data that feeds back into the neural network. The system learns from corrections, from human driving patterns, and from the Robotaxi fleet's own unsupervised operations. Each new FSD version incorporates this accumulated learning.
Key distinction: Traditional autonomous driving systems use modular pipelines (perception → prediction → planning → control) with hand-tuned rules at each stage. Tesla's end-to-end approach replaces this with a single learned system that maps pixels to actions — closer to how a human learns to drive.
FSD Version 15
FSD v15 is the current production version deployed across both the supervised consumer fleet and the unsupervised Robotaxi fleet. Across analysis channels, it is consistently described as a major upgrade — with noticeably smoother behavior in complex urban scenarios, better pedestrian interactions, and more naturalistic lane changes and turns.
Riders in the Robotaxi service report that unsupervised FSD v15 feels "smoother than supervised FSD" — the system's behavior is calm, patient, and predictable. The software handles urban complexity, hills, night driving, and traffic with a level of competence that multiple experienced Tesla drivers have described as surpassing their own driving.
The system demonstrates patient behavior in common scenarios — waiting for pedestrians at crosswalks, yielding at intersections, and navigating construction zones. Examples of this behavior have been documented across numerous ride-along videos from Austin, Dallas, and Houston.
Supervised vs. Unsupervised FSD
Tesla operates two tiers of FSD deployment:
Available to consumer Tesla owners. The vehicle drives itself, but a licensed human must be in the driver's seat, attentive, and ready to take over. Has accumulated 8–10+ billion miles.
Deployed in the Robotaxi fleet and Cybercab. No human driver, no safety monitor. The vehicle operates entirely autonomously. Has accumulated 1 million miles as of September 2026.
The supervised fleet serves as Tesla's data engine — billions of miles of real-world driving feed the neural network training pipeline. The unsupervised fleet validates that the trained system can operate safely without human backup. Both contribute to the continuous improvement cycle.
Safety Philosophy
Tesla's safety case for FSD rests on three pillars: vision + AI that processes the environment faster and more comprehensively than a human; passive safety designed into the vehicle from the start (the Cybercab was designed with disability groups and first responders); and the data flywheel — the more miles the system drives, the safer it gets.
The system is presented as already safer than average human drivers in its operating domain. Documented examples include patient behavior at crosswalks, yielding at four-way stops, and navigating complex construction zones. Remote operations exist as backup infrastructure, but the goal is fully unsupervised operation.
Critics note that Tesla's fleet size is still small compared to Waymo's paid-ride volume, and that the gap between registered and active vehicles raises questions about deployment pace. The consensus across analysis channels is that the technology is converging but scaling to volume remains the key challenge.
The Data Flywheel
Tesla's competitive advantage is the data flywheel: millions of Tesla vehicles on the road generate billions of miles of driving data. This data trains the neural networks. Better networks improve the driving experience, which increases adoption, which generates more data. The cycle compounds.
With 8–10+ billion supervised miles and 1 million unsupervised miles, Tesla has an order-of-magnitude more real-world driving data than any competitor. This is why Tesla believes a camera-only, software-centric approach will ultimately win — the more vehicles running the software, the faster it improves, and the cheaper the hardware, the more vehicles you can deploy.
For a comparison of how this approach stacks up against Waymo's sensor-heavy strategy, see our Robotaxi vs. Waymo analysis.
PUT THIS INTELLIGENCE TO WORK.
Use our fleet tools to model real economics, estimate location revenue, and plan your Cybercab deployment.