Accelerating physical AI development: How Antioch built the simulation platform for robotics

Building reliable autonomous systems like robots and drones requires a staggering volume of real-world testing. Every iteration cycle demands hardware in the loop. Teams spend weeks staging environments, running controlled trials, and collecting just enough data to make their next change. At the same time, edge cases are usually too expensive, dangerous or impractical to test in the real world. The result is a slow development cycle that produces robots with poor policy performance.

Antioch was founded to bridge the gap between simulation and reality. The company’s platform replaces hardware-dependent iteration with closed-loop simulation. It is a cloud-hosted environment where AI agents and engineers convert their real systems into digital twins, run massive suites of test scenarios in parallel, and validate autonomous behavior. The result is more predictable robots that are safer in the real world.

“Today, the development of reliable and safe physical AI systems is bottlenecked by expensive, slow, and incomplete real-world testing. The edge cases that degrade real-world performance or create unpredictable behavior are precisely the ones that are prohibitively expensive or impractical to test in the field”

— Harry Mellsop, Co-Founder, Antioch

Antioch’s platform is industry-agnostic and cloud-hosted, with no setup or configuration required to spin up a simulation environment. Teams simulate various robots — humanoids, quadrupeds, UAVs, autonomous mobile robots, industrial workcells — and Antioch provides digital twin creation pipelines, sensor and hardware model libraries, state-of-the-art scene generation tools, and evaluation frameworks to run those systems through thousands of simulated scenarios.

Antioch is compatible with leading simulation frameworks, like NVIDIA Isaac Sim and Isaac Lab, meaning that teams who have already built simulation capabilities internally can onboard in minutes and benefit from dramatically improved simulation fidelity and cloud scalability.

The platform includes an agent layer that pairs the simulation infrastructure with agents that have native expertise in autonomous systems. When a simulation run surfaces a failure mode, the Antioch agent learns from the replayable telemetry, autonomously proposes improvements to the robotic stack, and retests in parallel simulation to validate the change. The result is a closed feedback loop that compounds accuracy: each simulation run feeds the next, and each agent iteration tightens the system’s performance envelope.

Because Antioch’s customers need simulation environments in seconds, Antioch needed the ability to dynamically provision GPU compute exactly when customers need it. Behind the scenes, that means provisioning the latest-generation NVIDIA GPUs at the precise moment demand arrives and scaling those resources down just as efficiently when a session ends. When usage spikes, whether from a single customer running hundreds of parallel simulations or from multiple onboarding accounts going live simultaneously, the entire experience depends on whether the infrastructure can keep pace.

Among competing options that included hyperscalers, Nebius stood out for infrastructure capability and depth of technical expertise. From a rapid procurement process to sharp architectural consultation and ongoing white-glove support, Nebius approached the engagement as a long-term partner investing in a shared future.

“Nebius clearly understands the step change that physical AI will bring for the cloud compute landscape. Accordingly, it is optimizing to support the rapidly scaling companies, like Antioch, that will define the physical AI era”

— Collin Schlager, Co-Founder & Head of Product, Antioch

In addition, Nebius’ orientation toward the physical AI ecosystem, not just toward GPU capacity, was a key differentiator. Antioch was not looking for a commodity cloud but a partner cloud for robotics. Most importantly, the Nebius Physical AI Workbench provides an open-source framework that connects best-in-class tools across the physical AI development pipeline, from data curation (for example, with Voxel51's FiftyOne), to synthetic scene generation (for example, with NVIDIA Cosmos 3), through to simulation and policy training. For Antioch’s customers, this means the simulation environments they run on Antioch can more easily draw from a richer upstream data pipeline. For Nebius’ customers, it opens a seamless path to integrate with Antioch, delivering a step-change in simulation fidelity and scale compared to alternatives. As a result, curated and synthetically generated training data can flow directly into the simulation layer through pre-wired configurations, not glue code.

Antioch’s proof of concept on Nebius began with a clear set of requirements: latest-generation NVIDIA GPUs, dynamic scaling, and complex lifecycle orchestration. Within a week, Antioch and Nebius had a rollout plan ready. The engineering team ported the Antioch stack to Nebius instances and immediately began running internal development workloads: hundreds of parallelized simulations using both classical and differentiable physics solvers, world models, robotic perception pipelines, and photorealistic rendering.

On Nebius infrastructure, Antioch streams optimized simulation scenes to the browser at 100+ frames per second, with physics stepping at 240Hz. Even running heavy perception pipelines and multi-robot fleet simulations simultaneously, Antioch achieves 60+ fps with the full visual livestream active. This level of performance lets engineers watch their autonomous systems operate in real time and intervene when needed. Further, headless simulations, running without a visual output for pure evaluation purposes, step faster than real time to enable rapid sweeps across thousands of scenarios.

The practical impact on Antioch’s customers is significant. For one customer developing a quadrupedal warehouse-surveying robot, the platform reduced time to results from weeks to hours while improving model accuracy. Traditionally, adjusting control algorithm hyperparameters or sensor configuration requires staging a warehouse environment and running scenario testing over the course of weeks. Even then, most edge cases go untested because staging every variant is impractical. With Antioch running on Nebius, the same customer runs, in hours, a far more robust suite of scenarios across numerous environments and conditions.

Nebius has improved efficiency for Antioch’s cloud operations. Contrary to conventional thinking, the Antioch team found that simulation speed compounds as utilization scales. When a customer runs multiple heavy simulations headlessly in parallel, speed and resource utilization both improve. Further, every simulation run is catalogued with replayable telemetry for the Antioch agent to learn from. The agent proposes improvements, retests in parallel, and validates the change before a human engineer reviews it. Every marginal gain in simulation throughput becomes a multiplier across the entire development loop.

To Antioch, manufacturing, logistics, construction, energy, and agriculture collectively represent a $50 trillion global opportunity. As the company scales to meet upcoming demand, Nebius is positioned as a critical long-term partner. The customer base will expand rapidly as a new generation of robotics companies emerges, and simulation will become the central node in every physical AI development cycle.

Beyond raw compute, the Nebius and Antioch partnership reaches into how the physical AI development stack gets built. The Nebius Physical AI Workbench connects the Antioch simulation platform to the broader pipeline that physical AI teams run: from synthetic data generation with NVIDIA Cosmos, through Isaac Sim-based training environments, to policy evaluation — all composable, agent-driven, and running on the same Nebius infrastructure. As Antioch’s customers deepen their development cycles, the combination of Antioch’s simulation platform and the Workbench’s pipeline infrastructure means the full loop from failure detection to validated policy can run as a single workflow, without the orcustom integration work glue code.

“Together, we are positioned to become a fundamental layer of the physical AI stack. Nebius provides the cloud infrastructure, and Antioch provides the development platform”

— Harry Mellsop, Co-Founder, Antioch