URMAP®

Most radar perception models are tied to the hardware they were trained on. Change the radar, and the model breaks. URMAP — the Universal Radar Model for Active Perception — is Provizio’s AI perception layer for radar, trained on diverse data from 30+ OEM engagements so it works on hardware it hasn’t seen before. Plug in a radar point cloud — any radar — and start detecting, classifying and tracking.

DATA FLYWHEELURMAPPERCEPTION AS A SERVICEMODEL01CAPTUREYour environment, your data01CAPTUREYour environment, your data02LABELEvery frame, every object02LABELEvery frame, every object03TRAINTuned to your application03TRAINTuned to your application04EVALUATEQuality-assured before shipping04EVALUATEQuality-assured before shipping05DEPLOYOver-the-air to your radar05DEPLOYOver-the-air to your radar

Applications

One perception model for every application — on any radar you choose.

Automotive

ADAS and autonomous driving on public roads. URMAP detects, classifies and tracks vehicles, pedestrians and cyclists — out of the box, on whatever radar architecture you're integrating.

Industrial

The Industrial variant is specifically optimised for precision in environments with large, high-RCS machinery and challenging environmental conditions that swamp typical perception systems.

Robotaxi

Autonomous fleet operators need a single perception model that travels across vehicle variants and radar generations. URMAP's universal training base means switching radars doesn't reset the perception roadmap.

Evaluation

New to radar perception? Skip the months of data collection and ML buildout. URMAP runs on your hardware day one, then refines through Perception as a Service as your requirements mature.

How it works

A universal radar perception model with a data flywheel underneath.

STEP 01

One model, many radars

URMAP is trained on diverse, crowdsourced radar data from 30+ OEMs and use cases — automotive, industrial, robotaxi. That breadth means the model arrives already adaptable to new radar architectures, instead of needing months of hardware-specific training before producing useful output.

ONE MODEL, MANY SOURCESAutomotiveIndustrialRobotaxiSmart cityURMAPTrained on datafrom 30+ OEMs

STEP 02

Works with any radar point cloud

URMAP runs on any imaging radar that outputs a point cloud. Because it consumes that universal radar output rather than a vendor-specific format, you're free to choose — or change — the radar hardware without rebuilding your perception stack.

ANY POINT CLOUD INURMAPDifferent radars,different cloudsDetected, classified, tracked

STEP 03

Custom fine-tuning via PaaS

Need the model optimised for your specific application? As part of Perception as a Service, the data flywheel handles it for you — field data uploads automatically, Provizio labels and fine-tunes URMAP on top of the universal base, and the improved model deploys back to your radar over the air.

THE DATA FLYWHEELField dataLabelFine-tuneOTA updateURMAPYour field data improves your model, automatically

STEP 04

Better with every release

Every customer dataset feeds back into the universal base, so detection accuracy compounds over time. PaaS subscribers receive each refinement as an over-the-air update, delivering better mAP (mean average precision) detection accuracy with every version.

ACCURACY COMPOUNDSmAPv1v2v3v4v5Every release lifts accuracy, delivered over the air

Featured deployment

95% object detection from less than 0.1% of training data.

Provizio benchmarked URMAP on a customer A-sample radar the model had never seen before. With less than 0.1% of the customer's data added to URMAP's universal training base, the model reached 95% object detection on that hardware — proof that the universal training approach decouples high-fidelity radar perception from the lengthy, hardware-specific data collection cycles every customer used to face.

  • 95% detection rate on a customer A-sample radar the model had not seen
  • <0.1% of customer-specific training data needed to reach that result
  • Decouples radar perception from hardware-specific data-collection cycles
  • Available continuously through Perception as a Service (PaaS) — OTA

Variants

Three URMAP variants — one for each deployment shape.

Same universal training base, but tuned and packaged for different deployment scenarios. Pick the variant that matches your compute and your environment.

AUTOMOTIVE

On-road perception

Tuned for road driving scenarios — vehicles, pedestrians, cyclists, road users. Targets ADAS and autonomous driving applications on public roads. Runs on NVIDIA DRIVE Orin and similar GPU-class hardware.

Optimised for: ADAS · L2+ autonomy

INDUSTRIAL

Off-road perception

Trained for unstructured environments — construction, mining, agriculture. Optimised for environmental robustness (dust, weather, vibration), small-target discrimination next to high-RCS machinery, and multipath handling off metallic surfaces.

Optimised for: Mining · agri · construction

LITE

Edge-SoC perception

Knowledge-distilled from the full URMAP — compact, edge-optimised, runs directly on the radar's SoC. It delivers the same output interface as the full model at a fraction of the compute.

Optimised for: On-chip · no external GPU

Integration

Runs on standard compute, ships OTA

URMAP perception runs on GPU platforms — including the NVIDIA DRIVE Orin — and the Lite variant runs directly on embedded radar SoCs. Output is delivered using the same interface across deployment targets, so your downstream integration doesn't change when you move between compute tiers.

  • Automotive and embedded GPU platforms
  • Embedded radar SoCs (via URMAP Lite)
  • C++ and Python APIs
  • Standard messaging and time-sync interfaces
  • MISRA-compliant code
  • OTA model updates

Explore the stack

URMAP is the AI core of Provizio's 5D Perception stack.

Radar Odometry

Calculates vehicle velocity and yaw rate from the radar point cloud — feeds URMAP with the ego-motion needed for point cloud accumulation.

Radar Freespace

Identifies the drivable area from radar data — runs alongside URMAP to give planners both what's there and where it's safe to drive.

Tracking

Persistent object identities and velocities on top of URMAP’s detections — the layer planners consume directly.

Classification

Applies object-class labels to every detection — a perception module built on URMAP’s CNN core.

Questions, answered

URMAP FAQ

URMAP — the Universal Radar Model for Active Perception — is Provizio's AI perception model for radar. It interprets a radar point cloud into detected, classified and tracked objects. Unlike hardware-specific models, URMAP is trained on diverse data from 30+ OEM engagements, which means it adapts to new radar architectures fast — typically reaching production performance with a fraction of the customer-specific training data conventional models require.

Any imaging radar that outputs a point cloud. URMAP consumes that universal radar output rather than a vendor-specific format, so it works across radar architectures and vendors — that's the universal part of the name. Perception as a Service then lets you fine-tune it on your specific hardware without rebuilding the model from scratch.

PaaS is a subscription model where you provide raw radar data captured in your real-world environment, Provizio labels and annotates it, fine-tunes URMAP on your data, and deploys the resulting model to your radar units via OTA updates. You keep your annotated dataset, you receive continuous model improvements, and Provizio's universal model gets better through the data flywheel.

Three variants share one universal base. Automotive — tuned for on-road ADAS and L2+ autonomous driving, running on GPU platforms such as the NVIDIA DRIVE Orin. Industrial — for off-road environments where large, high-RCS machinery and small pedestrian workers coexist. Lite — knowledge-distilled to run directly on a radar's embedded SoC with no external GPU. The output interface is the same across all three, so you can prototype on one variant and deploy on another without changing the integration.

URMAP models are deployed and updated over the air. PaaS subscribers receive ongoing model improvements as the universal model evolves.

Radar perception, ready on day one.

If you're integrating radar and need detection, classification and tracking before you've collected months of training data, URMAP is the universal model that lets you start producing perception output from your existing radar sensor stack — then refine it for your application through PaaS.

Contact us