You can no longer rely on traditional software validation approaches.
As autonomous vehicles grow more capable, the challenge of ensuring their safety has migrated from the realm of code to the realm of data — a quieter but more fundamental frontier. On June 1st, Foretellix announced a reference solution built within NVIDIA's Alpamayo ecosystem, offering developers a structured path through the ungoverned terrain of AI training data: cleaning it, organizing it, and stress-testing it through synthetic scenarios. The partnership reflects a maturing recognition that the question of whether a self-driving system is ready for the world is, at its core, a question of whether the data that shaped it was honest, complete, and wide enough to hold the full complexity of human roads.
- Traditional software validation can no longer keep pace with AI-driven autonomy — developers are navigating a paradigm shift where data quality, not just code quality, determines whether a vehicle is safe.
- Raw autonomous vehicle logs are noisy and unreliable, creating a fragile foundation for training; without clean ground truth, the entire validation chain is compromised.
- Foretellix's workflow attacks the problem systematically — denoising logs, curating high-value driving segments, and organizing them into structured warehouses that give engineers real visibility into what their systems have and haven't encountered.
- Coverage gaps in operational design domains remain a stubborn danger; the toolchain identifies untested scenarios and generates synthetic variations at scale to fill them before deployment.
- The solution was demonstrated publicly at CVPR on June 5th, signaling industry-wide momentum toward treating data infrastructure as the central safety discipline for next-generation autonomous vehicles.
Foretellix, a safety infrastructure company for autonomous vehicles, announced on June 1st that it has built a reference solution integrated with NVIDIA's Alpamayo ecosystem. The goal is to help self-driving developers move more confidently through the stages that matter most: preparing data, generating test scenarios, and validating that AI systems are genuinely ready for real roads.
According to CEO Ziv Binyamini, the rise of AI-powered autonomy has made traditional software validation obsolete. The new challenge is data itself — gathering it, cleaning it, and ensuring it is comprehensive enough to train systems that will operate in unpredictable environments. Foretellix's solution begins with denoising raw drive logs to extract reliable ground truth, then branches into two directions: labeling real scenarios across time, and designing synthetic ones that can be controlled, varied, and replicated for testing.
From there, engineers move into structured data curation — transforming disorganized data lakes into warehouses that reveal exactly which conditions and environments a vehicle has been trained or tested on. This visibility is critical for identifying coverage gaps: scenarios the autonomous stack has never adequately encountered. Using Foretellix's Foretify designer alongside NVIDIA's Omniverse NuRec and Cosmos platforms, engineers can reconstruct scenes, modify actor behavior, and generate diverse synthetic situations at scale to close those gaps.
Foretellix presented the solution at CVPR on June 5th, appearing at both the NVIDIA Expo Theater and exhibition booth 826. The partnership marks a broader shift in how the industry is thinking about autonomous vehicle safety — less as a software engineering problem, and more as a question of whether the data used to build and test these systems is clean, complete, and representative enough to earn a place on public roads.
Foretellix, a company focused on safety infrastructure for autonomous vehicles, announced on June 1st that it has built a reference solution designed to work within NVIDIA's Alpamayo ecosystem. The move is meant to help developers working on self-driving systems move faster through the critical stages of preparing their AI models for real-world use: gathering and organizing data, creating synthetic scenarios for testing, and validating that their systems work as intended.
The shift toward AI-powered autonomy has fundamentally changed how autonomous vehicles get built and tested, according to Ziv Binyamini, Foretellix's CEO and co-founder. Traditional software validation methods no longer suffice. Developers need new approaches—ones that treat data itself as the central problem to solve, not just the code. Foretellix's solution, built for the Alpamayo platform, provides the methodology and tools to train and validate autonomous driving stacks with greater confidence and at larger scale.
The workflow starts with a step called denoising: taking raw drive logs from autonomous vehicles and extracting clean, reliable ground truth from them. This foundation enables two parallel paths. One allows engineers to label specific scenarios across time. The other lets them design synthetic scenarios—controlled variations of real-world driving situations that can be replicated and modified for testing. From there, developers move into data curation and warehouse exploration. Instead of drowning in an unorganized data lake, they identify high-value driving segments and organize them into a structured warehouse. This gives validation engineers immediate visibility into which parts of the operational design domain—the specific conditions and environments a vehicle is designed to handle—the system has actually encountered or been tested on, whether in simulation or on real roads.
The heart of the solution addresses a persistent challenge: gaps in operational design domain coverage. Foretellix's toolchain analyzes which scenarios the autonomous vehicle has not yet seen or tested adequately, then allows engineers to design new synthetic scenarios that directly fill those gaps. Using the Foretify scenario designer, integrated with NVIDIA's Omniverse NuRec platform, engineers can modify how actors behave in reconstructed scenes or introduce new actors entirely. They can further enhance these scenes with NVIDIA Cosmos. The ability to generate diverse scenarios at scale is essential—it lets teams validate that their autonomous driving stack is genuinely ready for safe deployment in the messy, unpredictable complexity of real-world driving.
Foretellix demonstrated the solution at CVPR on June 5th, presenting it both at the NVIDIA Expo Theater and at booth 826 during the exhibition. The partnership signals a broader shift in how autonomous vehicle development is approaching the problem of safety: not as a question of code quality alone, but as a question of whether the data used to train and test AI systems is comprehensive, clean, and representative enough to justify putting that system on public roads.
Citações Notáveis
Developers can no longer rely on traditional software validation approaches. Our NVIDIA Alpamayo-based reference solution demonstrates the Foretellix data-centric infrastructure required to train, validate, and safely scale next-generation AI driving systems.— Ziv Binyamini, CEO and Co-Founder of Foretellix
A Conversa do Hearth Outra perspectiva sobre a história
Why does the shift to AI-driven autonomy require such a different approach to validation?
Because traditional software testing assumes you can enumerate all the paths your code might take. With AI, the system learns patterns from data. You can't test every possible scenario the network might encounter. You have to ensure the data it learned from was representative enough.
What does denoising actually do?
Raw drive logs are messy—sensor noise, labeling errors, inconsistencies. Denoising extracts the true signal. It's the difference between learning from a blurry photograph and a clear one.
And synthetic data generation fills the gaps?
Exactly. Once you know which scenarios your system hasn't seen—rare weather, specific intersection types, edge cases—you can create controlled versions of those scenarios in simulation. You're not guessing anymore; you're targeting the holes.
Why does organization matter so much? Isn't data just data?
A data lake is chaos. You can't find what you need, can't see patterns, can't measure coverage. A structured warehouse lets you ask: Have we tested this vehicle in rain at night with pedestrians? The answer should be yes or no, not maybe.
What's the real bottleneck this solves?
Time and confidence. Without this infrastructure, teams spend months manually curating data and guessing whether they've tested enough. With it, they can see exactly what they're missing and generate it. That's the difference between launching in two years and five.