Our Journey & Milestones

Our Founding Bet on End-to-End Architecture

Since our founding, Turing has committed to a single strategic direction: developing full autonomous driving through End-to-End (E2E) architecture. Rather than combining sensors, high-definition maps, and rule-based systems to achieve conditional automation, we’ve chosen a fundamentally different approach. We take raw camera footage as input and deploy a single neural network to output driving behavior. This reflects our conviction that in a road environment where long-tail edge cases are infinite, scaling is ultimately impossible without this architecture.

E2E is a difficult technology to bootstrap and slow to gain initial momentum. Because of this, our early phase focused less on the model itself and more on building the foundational infrastructure needed to sustain the entire cycle: data collection, on-vehicle systems, real-world validation, and continuous improvement. These early investments created the prerequisite conditions for everything that followed.

Building a Real-World Improvement Cycle

Full autonomous driving is not development that ends once a model is built. The process requires driving, observing failures, relearning, and driving again—a cycle that must never stop.

Our fundamental development cycle has always been: collect driving data → train models → deploy to real vehicles → validate on public roads → feed results back into training. We learned early that evaluation can’t be completed on a desk. The real world reveals gaps constantly—not just “it works” or “it doesn’t work,” but many cases of “it works, but poorly.”

Once this improvement cycle is running, the central question shifts from “what can we do?” to “how fast can we improve?” The body of work Turing has accumulated is less a list of outcomes and more the construction of a structure capable of sustaining this cycle.

Data-Centric Development

Turing has consistently prioritized data-centric development. Driving data exists in enormous volume, but raw data is unusable for training. We must remove missing values and inconsistencies, restructure the distribution to match what learning benefits from, and create differentiated datasets aligned with our hypotheses.

The critical insight is that data quality depends not on volume alone, but on distribution design. Which scenarios are underrepresented? Which types of difficulty is our learning missing? By identifying these gaps, deliberately collecting the right data, and connecting it to training, we build cumulative capability against long-tail scenarios.

This work of data design and iterative refinement is unglamorous—but it is a foundation that compounds over time. We have continued building and operating this infrastructure as an indispensable element of scaling autonomous driving.

MLOps and Compute Infrastructure

Full autonomous driving development depends directly on how much compute we can run. E2E autonomous driving advances by training large neural networks, evaluating them, and continuously improving them. This makes GPUs not merely tools, but a resource that determines the speed of technical progress and organizational growth.

Turing has built the compute infrastructure needed to run the training–evaluation–improvement cycle, and alongside it, an MLOps pipeline—a “factory” connecting data collection, dataset generation, training, evaluation, and real-world validation. Running the improvement cycle at speed requires not just individual effort, but a system that operates as a mechanism.

Infrastructure and MLOps are not visible from the outside, but we treat them as the single most critical determinant of development velocity, and have made sustained investment and improvement in both.

Tokyo30: A Milestone

In March 2024, we announced the Tokyo30 project—a goal to drive continuously through central Tokyo for over 30 minutes without human intervention. At the time, no clear path existed. Through months of painstaking work across model development, data collection, and control system construction, we achieved Tokyo30 in November 2025.

This milestone represents a meaningful transition: E2E development has moved from a stage where “making it work at all” was the challenge, to a stage where we can accumulate steady improvement.

In May 2026, we are targeting the construction of a model trained on over one million scenes, pursuing an autonomous driving system with higher generalization performance.

¥27 Billion in Total Funding

Underpinned by our technical choices and accumulated execution, Turing has raised a total of ¥27 billion to date. In November 2025, we completed the Series A 1st close at ¥15.3 billion, marking our advancement to the next stage of capital formation.

Our bet on E2E from Day 1, the real-world improvement cycle we’ve built, and the concrete milestone of Tokyo30—the technical foundation and development capacity we’ve built through this process have been recognized as a long-term bet, and capital has concentrated accordingly.

Full autonomy is not a short-cycle project. It requires simultaneously scaling compute resources, data, organization, and business sustainability. Turing will continue to pursue the necessary options—including additional funding rounds—as our development and business phases advance.

Building Toward Full Autonomy

What we’ve built so far is not a finished product. Full autonomous driving still holds many unsolved challenges. At the same time, our foundational choices—the E2E bet from Day 1, the real-world improvement cycle, data-centric development, MLOps and compute infrastructure, and the Tokyo30 milestone—have together established the platform from which the next stage can begin.

Turing aims to make full autonomous driving viable as social infrastructure, through the continued accumulation of real-world improvement.