Zack Novak, now the CEO at AI-enabled simulation and synthetic data software company Parallel Domain, previously ran the managed solutions division at Quantix, which operates one of the largest dry bulk trucking networks in North America.
Novak said each of Quantix's bulk trucks hauled close to 50,000 pounds of plastic pellets, which puts the rig near 80,000 pounds going down the road.
“With that much weight, you just can't have safety misses. Simulation is the only safe way to train for that,” Novak said. “We believe every physical AI system should go through simulation before it reaches the real world. You can't put a loaded Class 8 into a near-miss on purpose to teach a model how to recover. In simulation you can, thousands of times, before the truck ever touches a public road."
Parallel Domain (PD) tested and validated its technology with an autonomous trucking developer this year, and now the company is making a big commercial push with the launch of its new software PD Forma, which transforms raw capture data into validated, simulation-ready environments in hours.
The AV truck developer worked with an early version of PD Forma prior to its formal launch to create thousands of replicas of real-world driving environments built straight from its own logs and used those replicas to train and test its system in the exact places it operates.
It used simulation for two jobs: first, testing every new model against the scenarios that matter before it goes on a truck, and second, taking the hard moments from their logs — such as a hard braking instance or a close call — and turning each one into thousands of variations to train on.
“The feedback was that replica quality held-up and the automation was easy to build into their pipeline,” Novak said.
How PD Forma works
PD leverages its technology products PD Replica and PD Sim to create highly realistic virtual worlds in which companies developing self-driving vehicles, drones, airplanes and robots can test, train and improve their AI without having to physically put their machines into every possible situation in the real world.
PD Replica converts raw video, image and sensor data captured by an autonomous machine into a photorealistic, geometrically accurate 3D/4D digital twin of the real-world environments in which it operates. The replica depicts the machine’s static surroundings, including buildings, roads, traffic signs and trees as well as all moving actors, such as other vehicles and pedestrians.
Parallel Domain
PD Sim connects directly with autonomy software stacks and simulates synchronized multi-sensor output from camera, LiDAR and radar to test how the autonomous machine reacts within the PD Replica.
PD Forma unifies them.
"With PD Forma, teams can move from real-world logs to validated, simulation-ready environments in just hours without sacrificing fidelity," Novak said. "We're helping customers accelerate their time to market by making program scale simulation more accessible and efficient. And we've seen demand explode; customer usage has increased 19X over the past six months."
PD Forma is the software suite built around PD Replica and PD Sim that orchestrates the workflow pipeline, from building the digital twin and assessing its quality to fixing its issues and testing scenarios in simulation.
Before PD Forma, a team would send its raw data to Parallel Domain for processing and reconstruction, resulting in a PD Replica. The new software product gives teams the full workflow to build and test replicas on their own through a web app, CLI (command line interface) or API (application programming interface) so they can scale their simulation workflows faster.
On the front end
Teams collect and package logs with the PD Forma SDK (software development kit) or CLI tools, which run automated validation checks locally and provide pass, warn or fail results with explanations and fixes.
Once the team fixes any issues, they upload the data into PD Forma using presets based on compatibility with driving, drone and indoor scenes. They can process uploads in batches, supporting stage logs that track progress as the AI — powered by Parallel Domain’s 4D Geometric Foundation Model — builds the replica. This enables teams to identify points of failure as they monitor the build from the web app or API and allows them to restart the build from the last good stage instead of starting from scratch.
The software then generates a quality assessment that shows how closely the virtual reconstruction matches the original real-world setting. Teams review that report and inspect the replica in 3D before publishing it to PD Sim.
On the back end
PD Forma pushes the PD Replica scene into the PD Sim sensor simulation platform, where it is fully converted into a permanent test location. Sim then renders synchronized, multi-sensor data streams (camera, LiDAR, radar) from the developer’s autonomy stack to simulate what the vehicle's sensors would observe inside that reconstructed environment.
PD Forma preserves the scene's geometry while varying the environmental conditions, allowing teams to reuse the same replica across different times of day such as dawn or dusk and weather conditions like rain or snow.
"By reconstructing the scenario with variations, PD Forma is designed to help users generate simulation experiences that pressure test how things may change as the environment does," Novak said. "That more effectively supports AI training, enables the investigation of system behavior and validates performance across conditions that would be difficult to capture consistently in the real world."
Teams can also incorporate multiple capture logs into larger replicas, work with mixed-camera inputs and dynamically reconstruct a scenario, enabling teams to replay a captured scenario or vary it in simulation.
