Onboard sensors travel with the vehicle. Cosmic explores what changes when the environment contributes a view of its own.
RF spatial view — geometry and observation lines.
Implementation · Automotive
The road does not end at the edge of sight.
At a junction, a building can hide an approaching road user. A sensing position elsewhere may observe that same space differently.
The idea is to bring those environmental observations into a shared spatial model—giving a vehicle additional evidence to evaluate alongside its own sensors and eventually steer the vehicle autonomous through the infrastructure itself with only a fail safe mechanism on board.
Research / A different allocation of intelligence
Change where perception lives. Not what powers the car.
Cosmic’s vision is to make world perception a shared city service, rather than a reconstruction workload repeated inside every vehicle.
The context
Compute is a real load. An EV is not a prerequisite.
Onboard sensing and AI processing require electrical power and thermal management. NVIDIA’s automotive platforms are explicitly designed for complex AI workloads.
EV platforms provide an electrical supply and electronically controlled propulsion—useful engineering characteristics for integration. But a large traction battery is not a requirement for autonomy: Waymo launched its service with the Chrysler Pacifica Hybrid before moving to an electric fleet. Its stated EV partnerships also emphasize sustainability.
Autonomy and electrification are separate engineering decisions. Battery size alone does not determine autonomous-driving capability.
The Cosmic proposition
Reconstruct once. Share the spatial state.
Distributed RF nodes↓Regional AI + physics reconstruction↓World-model stream to vehicles
The proposed infrastructure performs RF reconstruction and maintains the shared world model. Vehicles receive selected geometry, occupancy, motion, confidence and uncertainty instead of running the entire RF inference pipeline themselves.
This could reduce duplicated perception computation. The amount of onboard power or hardware saved must be measured; it does not eliminate the vehicle’s own safety-critical compute.
A compatible petrol, diesel, hybrid or electric vehicle could access the same infrastructure. The ambition is to extend the useful life of suitable existing cars, without making a change of powertrain a condition of access to shared perception.
Access could eventually be subscription-based, as familiar as a digital service. That familiar access model must be backed by vehicle-specific integration, validation and approval; a subscription alone does not enable automated driving.
01 / EquipIntegrate vehicle-specific OEM or professionally engineered aftermarket steering, braking and throttle actuation, with feedback and fault detection.
02 / ValidateIntegrate localization, secure communications, onboard perception, planning and control. Validate the vehicle’s operating conditions and safe fallback.
03 / ConnectAccess the infrastructure’s spatial service in covered areas. Automation remains subject to vehicle compatibility, safety evidence and applicable approval.
Ego-motion / Distributed viewpoints / NLOS
The journey starts at home. The context starts ahead.
Imagine a covered city corridor: once localized and connected, a car leaving home queries spatial context along a route extending kilometres ahead. Nearby nodes deliver relevant updates as it moves. That is our deployment vision.
01 / Before departure
The vehicle registers its position in the shared spatial frame and requests route-relevant state. A view of the covered corridor can arrive before the vehicle has traversed it.
Ego-motion is localization, not visibility
A moving frame needs registration.
Ego-motion estimates the vehicle’s own movement. Pose errors and time misalignment can distort the placement of perceived objects; visual methods can also struggle with poor texture or lighting. An occluded object is a related but different sensing problem.
Cosmic’s patent registers node observations into a common spatial frame and maintains entities across node regions. Vehicles still need reliable localization and time alignment to use that model. Collaborative-perception research identifies pose error, latency and communication loss as continuing challenges.
NLOS is an opportunity, not certainty
Different positions. Additional evidence.
A roadside node may observe a road user hidden from the car. Multiple transmitter–receiver paths can supply complementary RF evidence; physics-informed inference tests candidate spatial states against those observations.
The goal is to reduce viewpoint blind spots, not promise a completely occlusion-free world or anticipation of every hazard. RF propagation, materials, geometry, calibration and freshness constrain what can be resolved. Confidence and uncertainty travel with the estimate.
Public safety / Vision Zero alignment
A wider view. A safer-streets ambition.
Austin’s Vision Zero programme in Texas addresses deaths and serious injuries through street design, data and coordinated safety interventions.
Cosmic’s proposed AI World Recreation Engine could support research into earlier awareness at obscured crossings, vulnerable-road-user conflicts and infrastructure safety analysis. These are potential applications—not a claimed municipal partnership or a demonstrated crash-reduction result.
The safety case must stay with the vehicle
Shared context. Local responsibility.
Connectivity cannot be the only path to a safe response. Vehicle integration must account for stale data, outages, degraded perception and leaving the supported operating area, with a validated fallback to a minimal-risk condition.
Useful context. Not just a detection.
A spatial estimate matters only when the consuming system can interpret it.
Where things may be
Geometry and position connect an observation to the physical road environment.
How they are changing
Motion and temporal continuity help describe a developing situation.
How much is known
Confidence and provenance preserve the limits of the estimate and the evidence behind it.
Built to complement
A broader view. The same responsibility.
Infrastructure supplies spatial context. The vehicle evaluates that context alongside onboard perception. Planning, control and safety decisions remain with the consuming system.
Robotic perception in automation.
See how RF perception infrastructure supports autonomous mobile robots and industrial co-bots.
