Antioch Robotics raises $32M to turn robot testing into a software loop

Why it matters

Robotics developers cannot iterate at software speed while every meaningful change requires a physical test. Antioch is betting that a continuously calibrated simulation layer can become the default gate between new code and real machines, creating a valuable position in the physical AI stack if its virtual results reliably predict hardware behavior.

Harry Mellsop (@HarryMellsop) and his four co-founders at Antioch Robotics announced a $32 million Series A on September 8th, giving the New York simulation developer fresh capital to make testing robots look less like field work and more like running software.

Antioch Robotics on X

Greylock led the round, with A*, Category Ventures, BoxGroup, Icehouse Ventures and angel investors participating, according to Antioch's announcement on X. Greylock partner Saam Motamedi (@saammotamedi) is joining Antioch's board.

The financing brings Antioch's disclosed funding to $40.5 million, including an $8.5 million seed reported on April 16. That seed, led by A* and Category Ventures, valued Antioch at $60 million. Antioch says the Series A will fund product development, hiring and deeper simulation capabilities.

Mellsop built Antioch with Alex Langshur, Michael Calvey (@michaeljcalvey), Colton Swingle and Collin Schlager in May 2025. Mellsop previously worked on Tesla's Autopilot program and served as CTO of Transpose, a blockchain data infrastructure developer he founded with Langshur and Calvey. Chainalysis acquired Transpose in May 2023 for an undisclosed sum.

The other two founders brought experience closer to Antioch's current problem. Swingle worked on frontier model development and infrastructure at Google DeepMind, while Schlager worked at Meta Reality Labs and conducted machine learning and computational neuroscience research at Stanford.

That combination matters to Antioch's pitch. Mellsop, Langshur and Calvey have already built and sold developer infrastructure. Swingle and Schlager have worked on models and physical computing systems. Their second act applies the software tooling playbook to machines whose mistakes happen in warehouses, roads and factories.

Building a verifier for machines

Antioch describes its product as a "verifier" for physical AI: a development environment that predicts whether a change will improve an autonomous system before engineers put it on real hardware.

Robotics developers still rely heavily on physical test sites, recorded datasets, hardware-in-the-loop systems and manually operated fleets. Each method consumes equipment and engineering time. Failures can be difficult to reproduce, while rare combinations of weather, lighting, terrain, sensor degradation and hardware behavior may never appear during a conventional test program.

Antioch builds customer-specific digital twins that combine hardware, sensors, software, models and operating conditions. Developers can run scenarios in parallel, inspect regressions and generate synthetic data for situations that would be costly or dangerous to recreate physically. Antioch's website says developers can dispatch thousands of cloud simulations and connect those evaluations to continuous integration and deployment workflows.

The founders are pursuing the feedback loop that transformed software development: make a change, run a test, inspect the failure and repeat. In a reply to the funding thread, Antioch wrote that the "physical AI dev flywheel" needs to move as quickly as pure software development.

Antioch sits above and alongside underlying simulation technology rather than attempting to replace every physics engine. Antioch says its platform integrates with NVIDIA Isaac Sim, Isaac Lab and Omniverse libraries, then adds customer-specific calibration, scenario management, cloud execution and evaluation. Antioch is also working with Nebius on the computing infrastructure required to run those simulations at scale.

The fidelity problem does not disappear in the cloud

Simulation only becomes a useful engineering system when its results predict what hardware will do. A model that misses contact physics, sensor noise, latency or an unexpected environmental disturbance can give developers confidence in a machine that still fails outside the simulator.

Antioch's answer is a hybrid architecture. The platform explicitly models details such as geometry, kinematics, sensor placement and physical constraints, then uses real-world data to learn behavior that is harder to specify. Antioch says customer telemetry continuously recalibrates the simulation, creating a real-to-sim-to-real cycle that should improve as deployments produce more data.

Antioch says its simulations closely matched physical test results, including scenarios withheld from calibration. The central technical test remains measurable transfer from simulation to hardware across different machines and environments. The commercial test is whether Antioch can turn intensive customer calibration into a repeatable software product instead of accumulating bespoke integration work with every deployment.

Greylock backs the layer between models and machines

The Series A arrives as investors spread capital across the physical AI development stack. Applied Intuition raised $600 million at a $15 billion valuation in June 2025 for vehicle intelligence and autonomy software spanning automotive, defense and industrial markets. Data infrastructure developers such as Encord are pursuing the collection, curation and evaluation layer, while NVIDIA supplies widely used simulation and robot-learning frameworks.

Antioch's narrower wager is that evaluation infrastructure becomes a distinct control point. If robotics teams run every proposed software, sensor or mechanical change through Antioch before touching hardware, the simulator becomes embedded in daily engineering decisions. Each test can also produce data that improves the next one.

Greylock is funding that path less than 16 months after Antioch was founded and less than five months after its seed announcement. The money gives Mellsop and his co-founders room to recruit across simulation, machine learning, infrastructure, graphics and sales. It also raises the burden of proof: Antioch now has to show that software-speed iteration survives contact with the physical world.