Conventional radar tells you that something is there and roughly where. 4D imaging radar tells you the shape of the scene. That distinction sounds academic until you look at what it enables: a sparse point cloud of a scene, produced directly by the radar, without a camera, in bad weather.
What "4D" actually means
The fourth dimension is Doppler velocity, and the imaging part comes from the antenna array. A radar with M transmit and N receive channels forms a virtual array of M×N elements. With virtual channel count and signal processing determining the angular resolution, you can form a plan view of the scene rather than a set of range-azimuth detections.
The XD70 in numbers
| Parameter | Specification |
|---|---|
| Frequency band | 76–81 GHz |
| Virtual channels | 8 TX × 8 RX = 64 |
| Detection range | Up to 300 m |
| Point cloud rate | Up to 20,000 points per second |
| Platform compatibility | PX4 (open-source flight stack) integration supported |
The XD70 is configured for drone and robotics platforms, where the payload budget is a few hundred grams and the compute budget is a companion computer rather than a desktop.
Why an open flight stack matters
For drone developers, the integration cost is usually software, not hardware. A radar that outputs a proprietary binary format requires a driver, a parser, and a bridge into the flight controller — and that work is not what the drone team is being paid to do. PX4 support means the point cloud can be consumed through an established path, with the developer spending their time on behaviour instead of on decoding.
Where imaging radar genuinely beats a camera
- Low light and complete darkness. Thermal and imaging radar do not need illumination.
- Adverse weather. Rain, fog, snow and dust degrade optical systems rapidly; the effect on a 77 GHz signal over 100–300 m is small.
- Direct velocity measurement. Doppler is a physical measurement, not an estimate from frame differencing, so it is far more reliable for approach rate at long range.
- Fewer false positives than pure vision in clutter. Radar geometry is unambiguous about distance, which removes a large class of depth-estimation error.
The honest limitations
Imaging radar is not a replacement for a high-resolution camera. Angular resolution is lower, so object classification from radar alone is coarser than from vision. Dense clutter — a forest canopy, a stack of containers — produces many valid returns that a camera would resolve into a single object. And the compute required to turn raw frames into a usable point cloud is not trivial; the interface rate matters.
If your application needs to identify a specific object reliably — reading a sign, recognising a model of vehicle — radar should not be your only sensor. If it needs to know that something is there, where it is, and how fast it is closing, in fog or at night, radar is the stronger instrument.
Where it is being used today
- Drone autonomous flight — obstacle avoidance and terrain following where a camera alone fails in haze or at dusk.
- Port and yard logistics — container and vehicle tracking in low-light yards where lighting is expensive to install.
- Robotics platforms — AGVs and inspection robots needing reliable navigation in dust and darkness.
- Perimeter and traffic — vehicle classification and counting where an optical system would need continuous maintenance.
Frequently asked questions
How is imaging radar different from a camera plus optical flow?
Optical flow measures apparent motion of texture at the image plane, which requires illumination and textured surfaces, and it degrades sharply in fog, rain or darkness. Imaging radar measures range and Doppler directly from the radio signal, so it keeps working when the optical solution cannot.
What does the 64 virtual channel count buy in practice?
It determines the angular resolution and therefore how finely the radar can separate objects in bearing, and how well it can form a plan view rather than a set of range-azimuth points. More virtual channels generally means a denser, more usable point cloud and better separation of two targets at similar range but different angles.
Is 20,000 points per second realistic, or a headline figure?
It is achievable with the full array at work, but it depends on the measurement configuration. In practice a programme runs a lower point rate with a shorter integration window for responsiveness, or a higher rate for mapping tasks. We size the configuration to your application rather than quoting the maximum.
Does PX4 support mean I can just plug it in?
It means the integration path is established and documented, so you are not reverse-engineering a driver. You still configure the radar for your detection requirement, and the companion computer still needs the capacity to process the cloud. We provide the configuration guidance as part of the package.
Can I use it alongside a camera?
That is the recommended architecture in most cases. Use the radar for detection, ranging and velocity, and the camera for classification and recognition. Fusing the two gives a system that is more robust than either alone.