REAL
SELF-DRIVING
Building something different
The self-driving industry has raised billions of dollars without delivering real autonomy. Investment followed perception, not performance. Credentials and staged demos were easier to evaluate than real software capability, so companies were rewarded for looking credible rather than for shipping systems that actually drive. The field optimized for fundraising signals instead of engineering reality. Behind polished videos, most robotaxi systems remain far more fragile than they appear.
Mars Auto ships production autonomy.
With 10 engineers and $13M raised, we operate revenue-generating self-driving trucks in Korea and the United States. Our vehicles haul commercial freight every day in normal operations. We are the first autonomous trucking company in Korea, with more than 1.2M autonomous miles and cash-positive trucking operations.
We built a vision-first, end-to-end driving system in 2019, and it now runs in production. We began commercial operations in Korea in 2022. A model trained only on Korean driving data worked in the United States immediately, allowing us to launch U.S. freight operations in 2024.
Mars Auto will deploy unmanned trucks at scale.
Data-driven autonomy
HD maps and hand-written logic are a dead end. They encode assumptions about the world that break as soon as conditions change, forcing teams to patch failures instead of improving the system. Mars Auto builds autonomy by training a single, vision-first, end-to-end driving model directly from human driving data, without relying on maps or manually engineered rules. The system learns how humans actually drive rather than memorizing routes or edge cases. Learning replaces patching. The system operates beyond curated environments and improves automatically as data accumulates. Progress comes from scale and iteration, not from added complexity.
Data flywheel
Our data pipeline is built and scaled in Korea, where we operate under real four-season driving conditions and the fastest wireless network infrastructure in the world. This allows high-bandwidth data collection and rapid iteration without artificial bottlenecks. All data flows into a fully automated training pipeline. Collection, filtering, labeling, synthesis, training, validation, and deployment run continuously without manual intervention. Every production mile feeds directly back into the same driving model.
Trucking from day one
Trucking is fundamentally different from consumer driving. Trucks are heavier, slower to stop, operate for long hours, and carry real payloads with real consequences. These conditions expose weak autonomy systems quickly, while consumer vehicles allow failures to hide behind short trips and constrained use cases. Mars Auto has focused exclusively on Class-8 trucks from the beginning. Every design decision, dataset, and model iteration is built around trucking constraints, not consumer convenience or demo conditions.
