You might have noticed that over the past few years, presence detection has become quite popular across applications ranging from smart homes and buildings to automotive, industrial, and commercial spaces. It’s a perfect case for edge AI; however, the gap between basic detection and true radar intelligence still remains significant. The question is: Why?
Presence detection for objects, animals, and humans has historically relied on Passive Infrared (PIR) sensors, which are motion-activated. While PIR-based solutions are decently competent, we can’t really call them “smart”. You’ve probably been in a situation where you were working at your office desk, having achieved a caffeine-fuelled flow state, when suddenly the automatic lights switch off. Reluctantly, you get up from your chair and start waving your hands as if signaling a friend at a crowded concert.
Thankfully, radar implementation has been one of the most exciting developments in this area; a key advantage being that it can detect presence even when objects or people are stationary. And this can indeed work extremely well in controlled environments. But what happens when the system needs to distinguish between humans and pets or objects? Of course, you can add a powerful application processor in the mix at the cost of battery life, or try to make sense of the data via complex processing in the cloud, adding extra latency and data privacy issues.
It’s a good time for neuromorphic compute.
Radar data is inherently temporal, making it a natural fit for neuromorphic compute. By processing signal changes over time rather than static snapshots, neuromorphic architectures enable accurate, ultra-low-power presence detection directly on-device.
Real intelligence: let’s start at the beginning
Before we dive into why neuromorphic and radar are a match made in heaven, we have to take a few steps back and look at the smart systems governing the real-world sensor edge. Most AI systems today require very high-power compute, which is typically not possible at the sensor edge.
To bypass this, the real focus at the edge should be on figuring out whether anything meaningful is happening at all.
To accurately solve this problem, we must shift the intelligence forward in the chain. Systems that can detect meaningful patterns first and only then activate higher-power components will offer an unbeatable advantage for radar-based products that need to operate continuously, locally, and efficiently. For future edge devices, context-dependent input processing will always trump raw input processing.
A processing architecture built for intelligent sensing
Neuromorphic is the future. This isn’t a statement that we make lightly.
In very practical terms, this new way of computing enables sensor-based devices to process information only when something relevant occurs. This is crucial for two reasons. Firstly, sensor data is inherently temporal; for radar, this means presence detection is a result of detecting minuscule changes over time (like a person’s chest moving when they breathe). Secondly, the best way to process this type of signal is to look at embedded patterns over time rather than isolated snapshots.
While the radar sensor alone can produce rich 3D data about an environment, most systems still use that data for simple binary detection: present or absent. That means much of the signal’s value is left unused. Radar can capture subtle patterns such as gait, breathing, gestures, and movement signatures, but interpreting those patterns efficiently requires an architecture built for temporal data. When combined with Spiking Neural Processing (the fundamental architecture that powers our neuromorphic chips), radar systems can move beyond basic presence detection to classify what is present, distinguish humans from pets or objects, and separate relevant activity from background noise.
This enables systems to recognize a person even when they are stationary, while ignoring irrelevant disturbances such as moving branches or passing objects.
A chip like Pulsar can power many such applications, ranging from smart smoke and occupancy detection to smart doorbells with human presence detection, gesture recognition, activity recognition, and vital sign monitoring.
You might think that, given the scalability and multi-application compatibility, neuromorphic chips would be quite power-hungry. Short answer? They’re not. In fact, this balance of performance, power, and latency is exactly where neuromorphic compute stands apart. Traditional CPUs, or CPU-plus-CNN architectures, either become too slow, too power-hungry, or simply unable to run this level of sophisticated intelligence so close to the sensor.
While traditional AI architectures follow a brute-force approach, processing all data continuously, neuromorphic compute challenges this high base-level energy consumption by significantly reducing unnecessary activity while still being always-on at an extremely low power envelope. Because Spiking Neural Networks only fire when needed, the power draw for most Pulsar-based applications remains at sub-milliwatt levels.
Future-proofing radar
A growing trend in edge AI is the development of highly specialized, single-function chips optimized for specific tasks like object detection, keyword spotting, or always-on vision detection. While it’s true that neuromorphic compute deployed in these solutions can help accelerate initial deployment, it also introduces constraints that become more apparent over time.
Real-world products rarely perform a single function. In the future, a typical radar-based consumer device may need to detect presence, recognise gestures, classify activity, manage power, and integrate multiple sensor inputs. Achieving this with single-purpose solutions increases system complexity, cost, and design overhead. It also limits adaptability. As requirements evolve, systems built around fixed-function hardware often need to be redesigned rather than updated.
Building a reusable intelligence layer for radar
A more sustainable approach is to treat radar intelligence as a reusable layer rather than a fixed function. Neuromorphic architectures enable this by supporting multiple sensing tasks on a single platform, with behavior defined in programmable software rather than hardware.
This allows the same system to be adapted across different radar-based applications, from smart home devices to industrial monitoring or automotive sensing. It also enables sensor fusion, where radar data can be combined with other sensor inputs to improve robustness and context awareness.
From a product perspective, this reduces bill-of-materials complexity and accelerates development cycles. From a market perspective, it creates a foundation for scaling across use cases rather than optimizing for one.
The next phase of presence detection will be defined by how radar systems can move beyond binary detection. This requires systems that understand context, operate efficiently, and scale across real-world applications. It requires architectures that prioritize when and how to act.
The companies that succeed in this transition will be the ones who look beyond isolated detection features and focus on building flexible, efficient foundations for intelligence.
