We're introducing Helios Engine, SimForge's own renderer for driving simulation. It renders the same scenario at a fraction of the cost of CARLA, keeps every sensor on one clock, and runs the places we've rebuilt as digital twins.
Helios is built for closed-loop work. A policy drives, the world reacts, and the cameras, LiDAR and radar it sees are rendered on the fly, deterministically, from a scenario you can read and edit. The same engine that renders a hero clip at 1080p renders 512×384 training frames in about 9 ms a step.
Performance. A single 1080p camera frame takes 16 ms in Helios against 74 ms in CARLA Epic on the same RTX 5080. At training resolution, one 512×384 camera, Helios simulates 5.3 driving-hours per GPU-hour against 1.3 for CARLA at Low. The worker image is 0.5 GB; maps stream in.
Determinism. Repeat renders on the same GPU are byte-identical. The GPU LiDAR is bit-identical to its CPU reference over 3.8 million rays on two different cards. Over a 400-frame parking scene, 22,933 compared poses differed by at most 10.5 µm.
Scope. Digital twins of real places. 164 catalogued actors and props, from sedans and buses to pedestrians, sidewalk robots, drones and animals. Up to 64 sensors on one rig. Rain, fog, seven times of day, and a night lit by street lights placed from the survey.
The cases you can't collect.
A child steps out between parked cars at a dark school crossing. A cut-in under low sun, glare across the lens. These are the frames a fleet records once a year and a policy needs to see a thousand times. Helios renders them on demand, with the scene, actors and sensors described in a scenario file you can diff.
Not just cars.
The actor catalogue is one list, and everything on it drives, walks, rolls or flies through the same physics and the same sensors. Sidewalk robots share a crosswalk with pedestrians; a drone crosses a camera's field of view at the wrong moment; a flagger waves traffic through a work zone.
| Cars, vans, trucks, buses, tram | 37 vehicle IDs |
|---|---|
| Motorcycles, bicycles, scooters | two-wheelers |
| Pedestrians | 38 animated, adult and child |
| Emergency vehicles | ambulance, police, fire |
| Sidewalk robots | rover, quadruped, humanoid |
| Drones | delivery, camera, emergency |
| Animals | dog, cat, deer, raccoon, goose |
| Work zones | 15 props, flagger to excavator |
Every sensor, one clock.
Cameras, LiDAR, radar and the ego state are rendered from the same scene at the same tick, up to 64 sensors per rig. The GPU LiDAR is bit-identical to its CPU reference. If two sensors disagree, that is the scenario, not the renderer.
Real places.
Ten maps rebuilt from survey and imagery: Belmont, Richmond Field Station, Easterbrook, El Camino, Yale Street, Di Rosa, Garching and three phases of San Ramon. Lane geometry, signals, street lights and parking are data, so a scenario authored on one map runs on another.
Recast: from log to scenario.
Recast turns a recorded drive into a scenario: actors, ego path and signals extracted from the log, then re-driven in Helios so the policy can take over at any tick. Our internal library holds about 159,000 of them, each one passed by automatic gates.
Available now, on your machine and in the cloud.
Helios Engine ships in the SimForge runtime, free for non-commercial use: install the CLI, sign in and render locally. SimForge Cloud is open in beta, so it also renders on our GPUs: sign up and start with $10 in free credits. Dedicated capacity is available on request.