Electronics Era: Physical AI is increasingly being discussed as the next evolution in autonomous systems. How do you define Physical AI, and why has it become such a critical focus for the automotive industry?
Matthew Carrey: Physical AI is where the advances in software-enabled intelligence meet atoms – of which automotive is one of the messiest intersections. AI contributions started in software, while automotive has traditionally been a market rooted in the physical world. Now, these two worlds are intersecting with the goal of safety/efficiency – advanced sensing is the critical bridge that connects them.
Electronics Era: How does Physical AI differ from traditional AI approaches currently deployed in ADAS and autonomous driving platforms?
Matthew Carrey: Traditional ADAS AI is pattern-recognition: detect, classify, react. It works well when the world fits the training set. Physical AI is context-aware; it incorporates signal behavior, material properties, and environmental effects into the perception stack. Instead of just identifying a pedestrian and braking according to strict rules, for example, it understands how a host of inputs can affect an outcome and responds accordingly based on human training data (smoothly changing lanes instead of breaking due to water on the road).
This shift is what enables consistent performance in the conditions where today’s systems degrade most.
Electronics Era: How can advanced sensing technologies bridge the gap between perception and real-time vehicle intelligence?
Matthew Carrey: Advanced sensing technologies enable safer, more advanced vehicles by providing reliable, high-resolution data that can be trusted in every environment. This is core to the mission of Teradar – we’re inventing a new, untapped modality that can overcome the limitations (weather, resolution, range) that current vehicle systems face. Unlike other sensors on the market, our sensor offers both high resolution (you can see a child in the street over 200 meters away) and reliable performance in all weather conditions. While sensors that operate near visible light (like cameras, lidar, or traditional eyeballs) struggle in weather conditions, our technology can operate in any kind of weather – rain, snow, fog, dust, and glare. This level of information is pivotal to unlocking the vehicle of tomorrow in both functionality and safety.
Electronics Era: As the industry moves toward higher levels of automation, how do you see the sensing stack evolving over the next five years?
Matthew Carrey: As the industry moves toward greater automation, the sensing stack is being forced to evolve beyond what today’s hardware supports. Automakers are already rethinking their architectures – including pulling lidar from next-generation stacks – because the mixture of reliability in real-world conditions and cost has become the primary factor for Level 2++/Level 3 autonomy.
We’re also seeing increased demand for ‘orthogonal modalities’, meaning sensor data is decoupled from each other as much as possible to improve safety. While Lidar and cameras use wavelengths very close to each other (yielding similar failure modes), Terahertz waves don’t suffer from those same downsides and are increasingly attractive as a second source of truth.
If we want to really unlock true autonomous driving, the way that most people think about it (which is L3), you need to be able to have a sensor stack that sees everything all the time. It’s clear we don’t have that today – conversations with even the advanced autonomy players are full of ‘we switch to a different, degraded performance in snow, rain, or fog due to sensor limitations’ – we’re building a sensor that won’t force that tradeoff.
Electronics Era: What are the biggest limitations of today’s camera, radar, and LiDAR-based perception systems?
Matthew Carrey: Cost, resolution, and all-weather reliability are the three biggest gaps in today’s sensing stacks, and each of the legacy modalities falls short in at least one of them. Cameras struggle similarly to our eyes, performing poorly in direct sunlight, darkness, or a mix of the two at sunset. Radar is cost-effective, but it doesn’t provide the resolution needed for next-level autonomy. Lidar is costly, affects design, and degrades when exposed to particulates such as rain, fog, snow, or dust. All of these constraints hold back the safe development of safe or autonomous driving.
Electronics Era: How can emerging sensing technologies improve environmental perception and decision-making accuracy?
Matthew Carrey: Terahertz (THz) waves, which lie between radar and lidar on the electromagnetic spectrum, have always held great promise for sensing applications due to their unique wavelength, which enables both high resolution and all-weather penetration. Previously constrained by available technologies, this new category of sensing is now viable following Teradar’s breakthrough in terahertz chip design. Because THz waves are far less affected by environmental interference, they maintain performance through rain, fog, snow, dust, smoke, glare, and low-light conditions – precisely the scenarios where today’s sensors degrade or fail. This consistency across real-world weather conditions gives autonomous systems a clearer, more reliable view of their surroundings, dramatically improving the accuracy of AI-driven perception and decision-making.
Electronics Era: While significant progress has been made in assisted driving, widespread L3 adoption remains limited. What are the biggest factors slowing deployment today?
Matthew Carrey: Consistency. L3 requires OEMs to take liability for the vehicle’s actions, something they have (wisely) hesitated to do.
To keep people safe (and therefore reduce accident litigation to a minimum), the standard must be incredibly high – even in operational conditions that vehicles struggle with today, such as bad weather, poor lighting, and inconsistent driving environments. OEMs have to do all of this while keeping the car within budget – a herculean task.
Teradar can help make this possible.
Electronics Era: How important is sensor reliability versus software intelligence when addressing current autonomy limitations?
Matthew Carrey: Sensor reliability is just as important as software intelligence when it comes to overcoming today’s autonomy limitations, because even the most advanced AI can only make decisions based on the data it receives. Over the past few decades, we’ve made immense strides in software development, but the limiting factor is still the hardware: today’s sensors don’t deliver the consistent, high-resolution, all-weather reliability needed for safe L3 and beyond. No matter how good the software is, if the underlying signal is degraded by rain, fog, glare, dust, or low light, the system will hesitate, hand control back to the driver, or fail outright.
Electronics Era: Safety remains the defining challenge for autonomous mobility. What metrics are regulators and automakers quietly prioritizing behind the scenes?
Matthew Carrey: Regulators and automakers are increasingly focused on the operational metrics that keep surfacing in safety investigations – things like the number of unplanned stops per 1,000 miles, the frequency of minimal-risk maneuvers, weather-related perception dropouts, and how often a vehicle exceeds its operational design domain.
Regulators are also tracking low-speed collision rates, near-miss events with pedestrians and cyclists, emergency-vehicle interactions, and the consistency of sensor performance across lighting and weather conditions – all areas where today’s robotaxis continue to struggle. By addressing these concerns and delivering visibility and reliability through fog, rain, snow, and sun glare, Teradar’s terahertz sensors could help prevent up to 150,000 road deaths worldwide each year.
Electronics Era: How are OEMs measuring real-world safety performance beyond traditional disengagement statistics?
Matthew Carrey: Each OEM uses a wide range of statistics going beyond normal accident detection or disengagement requests. While they are specific to each company, they generally capture information about the driver’s attentiveness, road conditions, and (where Teradar comes in) what the sensor stack was seeing prior to any accident or incident.
Once a failure at the sensor fusion level is found, OEMs will peel back the layers to discern which sensors failed to correctly understand the environment around the vehicle. If one sensor (e.g., a lidar in a foggy situation) keeps underperforming, the data is used to build a model that leverages lidar data less in foggy situations. Teradar’s goal is to provide a sensor that never needs to be discounted, instead serving as a source of truth even in difficult environments.
Electronics Era: What role will scenario-based validation and simulation play in future regulatory frameworks?
Matthew Carrey: The vehicles of the future will be incredibly complex machines, likely driven by a form of end-to-end neural networks. This means that the vehicles of tomorrow will likely be certified and regulated not by code inspection and rules-based, deterministic algorithms, but by statistics (similar to aircraft safety standards today). Prior to any new model introduction on the road, scenario-based simulation, evaluation, and validation will be mandatory to prevent incidents from occurring in the real world.
This means that the sensor stack (and its limitations) will also have to be modeled and characterized – something Teradar is actively working on.
Electronics Era: Looking ahead, what developments do you expect to have the greatest impact on autonomous vehicle performance over the next three years?
Matthew Carrey: While I might be a bit biased, I truthfully believe one of the most meaningful shifts will come from the next generation of sensing technologies entering broad vehicle programs- including the introduction of our terahertz vision sensor, Summit, which is targeting the start of production in 2028. The combination of simplicity, reliability, and high resolution provides a step function in sensing, as expressed by our OEM partners.
But more broadly, I’m also excited about the increased compute arriving on vehicles, enabling more/better data processing, along with robust, complete training data for our AI systems to learn on. It’s going to be exciting (and safer!).
