Antioch Raises $32M—Simulation Still Has a Reality Gap

New York physical-AI startup Antioch raised a $32 million Series A on September 8, 2026, led by Greylock with A*, Category Ventures, BoxGroup and Icehouse Ventures participating. The financing backs a platform designed to test changes in simulation before code reaches robots, drones and other physical systems.

Antioch’s account of the round states that an earlier $8.5 million seed brought its total funding to $40.5 million, describes a hybrid real-to-sim-to-real evaluation loop and names Amazon and Launchpad Build AI as users alongside NVIDIA and Nebius in ecosystem roles. The business proposition is not that physical testing vanishes, but that continuously calibrated simulation can screen more changes before scarce hardware and test facilities are used.

Antioch wants simulation to become a release gate

The platform sits between a proposed system change and deployment to a physical machine. It incorporates a customer’s hardware, sensors, software, models, environments and operating constraints, then uses physical outcomes to recalibrate the virtual environment.

That setup lets teams evaluate changes to perception models, planners, controllers, sensor configurations or mechanical designs across parallel scenarios. A regression found virtually can stop a weak build from advancing immediately to a test rig or field fleet, while engineers can reproduce failures without repeatedly arranging the original physical conditions.

A credible real-to-sim-to-real release loop has four distinct stages:

1. Collect measurements and failures from the physical system, while reserving representative cases for validation rather than calibration.
2. Model known geometry, kinematics, sensor layouts and operating constraints, and learn effects that are difficult to specify explicitly.
3. Evaluate proposed changes against routine conditions, rare events and previously observed failures, comparing the result with an accepted baseline.
4. Run selected builds on hardware and return the observed outcomes to the next calibration cycle.

The crucial distinction is the authority assigned to each result. A simulated failure is useful evidence that a change needs investigation; a simulated pass makes the stronger claim that the virtual test did not overlook a field-relevant failure.

Synthetic coverage is not the same as predictive fidelity

Antioch combines explicit simulation with components learned from real-world data. Explicit models provide control over known system properties, while learned components can approximate sensor behavior, contact effects and other interactions that conventional models struggle to reproduce.

This hybrid approach can generate labeled examples of rare failures, unusual environments, changed hardware configurations and dangerous situations that would be costly or irresponsible to stage repeatedly. Synthetic data has clear value when it expands training coverage or exposes a weakness absent from ordinary field logs.

Neither outcome proves that simulation predicts field performance. A simulator can create many varied scenarios while misrepresenting the frequency, severity or causes of real failures. Establishing predictive value requires comparison against physical results that were not used to tune the simulator.

The fidelity evidence remains too limited for broad conclusions

In a Forbes interview and account of the Ring deployment, Amazon and Ring executive Jason Mitura said Antioch’s simulations closely matched physical tests, including scenarios withheld from calibration, while co-founder Harry Mellsop described real-world data as the continuing standard and simulation as a way to improve sample efficiency. Held-out scenarios are more informative than agreement on calibration data, but the disclosed result is not a comparable benchmark.

The public material does not provide simulation error rates, confidence intervals, the number or difficulty of Ring’s held-out cases, or the exact systems and operating conditions covered. It also omits a false-negative rate showing how often a virtual evaluation accepts a change that later fails on hardware.

For safety-critical deployment, minimum useful fidelity evidence would identify the supported hardware and operating envelope, measure error on decision-relevant outputs, separate calibration data from validation data and show whether known failures are detected. Those comparisons would also need to be repeated after material changes to hardware, software or the environment.

Calibration can drift as sensors age, components vary, motors heat up, payloads change or uncontrolled conditions enter the system. A dependable release gate therefore needs a way to detect deteriorating agreement and narrow its claims when physical behavior no longer matches the virtual model.

Customers will determine how much authority the gate earns

Ring supplies Antioch with a named deployment context, while Launchpad Build AI represents an industrial-manufacturing use case. NVIDIA’s Omniverse libraries, Isaac Sim and Isaac Lab support the simulation stack, and Nebius provides infrastructure for running evaluation workloads at scale.

Those relationships answer different questions. Infrastructure integrations can make scenarios easier to construct and parallelize; customer comparisons between simulated and physical outcomes determine whether results are trustworthy enough to delay, advance or redesign a release.

The financing gives Antioch resources to expand its platform and team. The next material evidence will be validation across multiple hardware classes, measurable failure coverage and explicit rules for when a virtual pass still requires physical confirmation. At present, the supportable claim is that simulation can reduce and focus hardware testing—not that it can replace the real-world tests needed to prove field performance.

Also read:

• QuantHealth Raises $45M—Virtual Trials Still Depend on Real-World Proof
• API Security’s Hidden Gap: Valid Logins Can Still Expose Business Data
• What Is a Startup? The EU Has Rules, but Business Reality Is Broader