Data-centre GPUs
Perception and planning networks are trained on real terminal drives and synthetic scenes, then optimised for the vehicle with TensorRT.
Level 4 means no driver to fall back on. We get there the way the leaders in autonomy do: automotive-grade compute in the vehicle, a digital twin to test against, and a safety architecture that assumes things will fail.
No shift change and no lights-out. Camera, lidar and 77 GHz radar overlap so the vehicle keeps working in the sea fog, spray and darkness ports actually operate in.
Under SAE J3016, a Level 4 system performs the whole driving task and its own fallback within a defined operational design domain. Nobody is expected to take over; if conditions leave that domain, the vehicle brings itself to a safe state.1
| SAE level | Who drives | Fallback | Where it operates |
|---|---|---|---|
| Level 2 | Driver, supervised assistance | Driver | Anywhere the feature works |
| Level 3 | System, while engaged | Driver on request | Limited conditions |
| Level 4Our target | System | System reaches a safe state | A defined operating domain |
| Level 5 | System | System | Everywhere |
Simplified from SAE J3016_202104. Our operating domain is the mapped, geofenced area of each terminal we deploy in.
The industry's Level 4 programmes, from robotaxis to highway trucks, all start by drawing a tight operating domain.2 A container terminal offers one of the tightest there is.
A terminal is fenced, surveyed and controlled. The operating domain can be drawn precisely and enforced with geofences.
Quay-to-yard moves repeat on known lanes at low speed, far from the open-ended complexity of public roads.
Terminals segregate people and machines by design, and their operating systems already dispatch every container move.
Ports run 24/7. A vehicle that never changes shift turns autonomy directly into throughput and lower cost per move.
NVIDIA describes physical AI as three computers working in a loop: one to train the models, one to simulate the world they will face, and one to run them in the machine. Every mile and every edge case feeds the next round.3
Perception and planning networks are trained on real terminal drives and synthetic scenes, then optimised for the vehicle with TensorRT.
A digital twin of the terminal tests every software build in closed loop, across conditions too rare or dangerous to stage for real.
The validated software runs on the vehicle. What it encounters is logged and returned to training, closing the loop.
A Level 4 vehicle has to be right in situations it may meet once a year. Simulation is how those situations get tested thousands of times before deployment.4
Omniverse builds physically accurate digital twins on OpenUSD. Its AV simulation blueprint replays real sensor logs, produces ground truth and runs closed-loop tests, with physically based camera, lidar and radar models.
Cosmos world foundation models generate and transform driving scenes, turning one recorded move into the same move in fog, rain, glare or darkness, so rare and dangerous cases are trained on before they happen. 5
Isaac Sim, also built on Omniverse, simulates robots and generates synthetic training data, a fit for modelling container handoffs, docking alignment and the machines a vehicle works beside. 6
The simulation pipeline is on our roadmap; see the status table below.
The first is the platform: automotive-grade compute engineered toward functional-safety standards, with redundant power and a safety microcontroller alongside the main processor.
The second is ours: an isolated safety channel, outside the main computer, that checks every motion command against geofences and proximity limits and can stop the vehicle on its own, backed by a hardware emergency stop and contact edges.
Safety and compliance on the platform pageAn honest view of what is running, what we are building and what comes next.
| Component | Role | Status |
|---|---|---|
| Perception, planning and control stack | Autonomy software | In development |
| Independent safety monitor and e-stop | Fail-safe channel | In development |
| Fleet dispatch and monitoring | Operations software | In development |
| Omniverse terminal digital twin | Closed-loop simulation | Roadmap |
| Cosmos synthetic scenario generation | Training data | Roadmap |
| Isaac Sim handoff and docking models | Robotics simulation | Roadmap |
Pilot partners shape the operating domain, the first routes and the digital twin we validate against.