Tesla, Waymo, and Anduril each spend hundreds of millions per year on virtual testing infrastructure. Which used to mean: if you wanted to build competitive robot AI, you first needed that capital. Antioch wants to change that equation.
What is this?
Antioch is a cloud simulation platform that lets robotics and autonomous systems teams validate their software without physical testing. In physical AI — robots, self-driving vehicles, drones, industrial machinery — testing costs are extreme. Bad deployments can hurt real people.
Founded in New York in May 2025, the company was built by five co-founders: Harry Mellsop (Tesla Autopilot), plus alumni from Google DeepMind, Meta Reality Labs, and Chainalysis. The April 2026 seed round: $8.5M at a $60M valuation, led by A* and Category Ventures. Palantir CTO Shyam Sankar joined as an angel.
The core problem is the "sim-to-real gap." Robots that perform perfectly in virtual environments often fail the moment they enter the real world — because physics engines cannot fully replicate real-world friction, lighting, and sensor noise. Antioch addresses this by layering domain-specific libraries on top of top-tier physics engines like Nvidia Isaac Sim and World Labs models.
What changes?
Until now, validating a single robot behavior meant renting a warehouse, staging environments manually, and running each scenario one at a time. Mellsop described teams renting Airbnbs just to test household robots overnight. Dangerous edge cases — fire emergency protocols, for instance — were simply impossible to test physically.
| Physical testing | Antioch cloud simulation | |
|---|---|---|
| Environment setup | Weeks to months, millions | Instant, via code |
| Parallel runs | Limited by hardware count | Thousands simultaneously |
| Dangerous scenarios | Impossible or very risky | Fire, snow, fog — freely defined |
| Failure analysis | Manual video review | AI agent, frame-by-frame |
| Accessibility | Requires Tesla/Waymo-scale capital | Same tooling for any startup |
"What happened with software engineering and LLMs is just starting to happen with physical AI" — Category Ventures. The difference: failures in the physical world carry consequences far beyond a code bug, so the bar for verification tooling is dramatically higher.
In March 2026, Allen Institute for AI (Ai2) announced zero-shot sim-to-real transfer with their MolmoBot model — robots trained entirely in simulation successfully completed real-world tasks with no additional real-world data. Simulation quality has crossed a meaningful threshold.
How to get started
Antioch currently onboards via direct engagement. Request a demo at their site, and the workflow follows five stages.
- Onboard — Bring your existing code, unchanged
Whether ROS-based or custom, Antioch Ark containerizes your robot software and firmware, auto-generates digital twins, and connects virtual cameras, LiDAR, radar, and IMU sensors. No hardware changes needed. - Define — Design scenarios with the Python SDK
Define extreme conditions — rain, snow, fog, fire — as code. Use the curated sim-ready asset library or upload your own 3D assets. Set pass/fail criteria programmatically. - Simulate — Thousands of parallel runs in the cloud
Antioch Cloud runs thousands of simulations simultaneously. Integrates with CI/CD pipelines and tracks pass rates across versions. - Analyze — Frame-by-frame debugging
Replay any failure at the exact moment it happened. Native Foxglove and Rerun integrations for telemetry visualization, plus automatic cross-version comparison. - Accelerate — AI agent closes the loop
Built-in AI agents analyze failure causes and immediately validate whether proposed changes would have fixed them. MCP support lets external agents plug in too.
"We genuinely all think that anyone building an autonomous system for the real world is going to do so in software primarily in two to three years."
— Harry Mellsop, Co-founder, Antioch
