Product1 publisher3 min readPublished
Antioch raises $32m to put a robotics team's own sensors inside a calibrated simulator
The pitch is a software-style test loop for machines: model a customer's own hardware and sensors, calibrate against real runs, then try thousands of scenarios before booking rig time. Ring says the match held on scenarios it withheld.
The Product Desk · Product desk
What happened
- Antioch, a New York startup building simulation software for robots, drones and other autonomous machines, has raised a $32m Series A led by Greylock, without disclosing a valuation.
- The round was announced in a blog post on 8 September, with A*, Category Ventures, BoxGroup, Icehouse Ventures and angel investors joining, and Greylock partner Saam Motamedi taking a board seat.
- The platform builds a simulation of a customer's hardware, sensors, software and models, calibrates it against real-world data, then runs thousands of scenarios in parallel to find failures and check fixes.
- Ring's Jason Mitura, quoted in the announcement, said Antioch's simulations closely matched Ring's physical test results, including in scenarios Ring deliberately held out of calibration.
- The product sits on other companies' infrastructure, integrating Nvidia's Omniverse libraries, Isaac Sim and Isaac Lab, with Nebius supplying the compute to run simulations at scale.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- decision Once a supplier's simulation is trusted against held-out physical tests, the test site stops being a fixed cost of doing robotics and becomes a budget line an engineering manager has to defend.
- cost A buyer prices a licence and inherits an unpriced job, because keeping a calibrated twin honest across hardware revisions is recurring work that nobody in the announcement has costed.
- exposure Test coverage that migrates off the rig lands on someone else's GPUs and someone else's simulation libraries, so a release gate now depends on two suppliers the robotics team never evaluated.
- contradiction Antioch's argument against fully learned world models rests on a data requirement most teams cannot meet, while its own stack builds on the vendor promoting those open world models.
Robotics teams wait on rig time. The perception change is ready, the fleet still has to be driven by hand, and the answer arrives whenever the test site frees up. That queue is what Antioch is selling against: code teams get thousands of tries a day, and robotics teams get a handful because they depend on test sites, recorded datasets, rigs and hand-run fleets [6].
Teams often say a simulator's job is to prove the system works before it ships. In practice, on a given Tuesday afternoon, it gets used to decide whether a change is worth a slot on the rig. That second job is the one Antioch has named its product after, calling the thing a verifier for physical AI [8], and it is the job with an acceptance criterion a buyer can actually run. The hold-out detail is what gives Ring's statement weight [9], because a simulator tuned on the same data it is later scored against tells you nothing. The announcement does not size any of it: no count of withheld scenarios, and no price [22].
Calibration, then, is the product rather than a setup step. Antioch writes that "the destination is fully learned" and that the practical path is hybrid [13], which is candid about how much of the simulator is still built by hand from geometry and sensor layouts [12]. When a customer moves a sensor mount or swaps a lens, someone has to recalibrate against fresh real-world runs, and the announcement does not say who does that or how often it repeats [22].
Antioch raised $8.5m in April and announced $32m on 8 September, for $40.5m in total [4][2]. The Series A is 3.8 times the seed [20], roughly five months later [21]. How new the company is cannot be pinned down from the record: an April report is cited as putting the founding in May 2025, which sits after the seed [19][23]. Greylock's Saam Motamedi, now on the board [5], framed the round as software-speed innovation in the physical world [24]. The checkable part of the same post is smaller and more useful.
For a team weighing a pilot, the forcing function is two lists written before the trial rather than after. One names the failure classes a calibrated simulation would be allowed to clear with no physical run, ordered by what each of those failures costs to stage for real. The other names the classes that keep a physical run whatever the simulation says. An empty first list means the purchase is really synthetic data for rare events, which Antioch sells as a separate use of the same simulations [8], and rig time stays the constraint. An empty second list means nobody has run the hold-out procedure yet, and that procedure is what to demand in the pilot: withhold scenarios from calibration, then score the simulator against the rig on exactly those [9].
What to watch
- Whether any customer other than Ring publishes a number for physical test runs retired after moving coverage into simulation.
- Whether Antioch's learned components start replacing the hand-built geometry and sensor layouts, which changes who owns recalibration.
- Whether the Nebius compute bill for running thousands of parallel scenarios surfaces in Antioch's pricing as pilots scale.