The Physical AI Perception Stack

Executive Summary

Physical AI begins with a deceptively simple requirement: a machine must know what is around it before it can decide what to do. That makes perception a prerequisite for autonomous vehicles, industrial robots, drones, defense systems and humanoids. It does not, however, make every sensing supplier an attractive investment.

Our central conclusion is that the opportunity is in the perception stack, not in a single winning sensor. Cameras provide semantic detail at low cost. Radar adds velocity and all-weather operation. LiDAR supplies direct three-dimensional geometry. Thermal and short-wave infrared extend performance into darkness and obscurants. Force-torque and tactile sensors close the loop when a machine makes contact with the world. Each modality has a failure envelope, so higher-reliability systems tend to combine sensors and process their outputs through increasingly capable edge compute.

That architecture creates two distinct types of public-market exposure:

1. Qualified component suppliers sell hard-to-replace sensors, analog components and safety-certified silicon. They can benefit from rising sensor content, but hardware price erosion and customer concentration often limit upside.
2. Integrated perception platforms combine hardware, software, calibration and workflow data. Their economics can improve as software attach, installed-base learning and qualification lock-in deepen.

Within the companies reviewed, Cognex has the cleanest platform characteristics in industrial machine vision. NXP is the strongest diversified component franchise, particularly in automotive radar, processing and networking. Leonardo DRS offers differentiated defense exposure where qualification and program duration matter more than consumer-style unit economics. Mobileye has the strongest software and data architecture but also the most direct competition from integrated automotive compute platforms. Ouster is the clearest public LiDAR exposure, yet it remains a bet on both company execution and continued demand for the modality itself.

The report does not assume that LiDAR wins every vehicle socket, that humanoid production ramps on a fixed schedule, or that market-share estimates are audited facts. Those are the variables to monitor, not foundations to take on faith.

The Investment Question

The broad thesis—machines need sensors—is obvious. The investable question is narrower: which companies can preserve value as sensors become cheaper, compute becomes more integrated and customers demand complete perception rather than raw data?

Three forces shape the answer.

First, no modality is universally sufficient. Cameras struggle with glare, darkness and ambiguous depth. LiDAR performance can deteriorate in rain, fog, dust and highly reflective scenes. Radar works through poor weather and measures velocity directly, but historically offered less angular detail. Thermal imaging detects heat rather than visible texture. Reliable systems manage these limitations through redundancy, sensor fusion and carefully bounded operating conditions.

Second, hardware tends to commoditize. Semiconductor scaling, integration and competition reduce the cost per sensing channel. That is good for adoption but not automatically good for supplier returns. A falling sensor price must be offset by more sensors per system, share gains, software revenue or a larger served market.

Third, qualification creates friction. Automotive and defense components require long validation cycles. Industrial vision systems become embedded in production recipes, lighting configurations and quality-control workflows. A nominally replaceable sensor can therefore be commercially sticky even when competing hardware exists.

This produces a useful distinction. A bottleneck is difficult to replace. A platform captures more value as usage grows. The best investments can be both; most are only one.

How the Perception Stack Works

The stack can be read from left to right:

The farther right a supplier operates, the greater the potential for recurring software and workflow lock-in. The trade-off is higher competition from customers and large compute vendors that want to own the complete stack. Component suppliers face less platform competition but more price pressure.

Latency is the common technical constraint. A mobile robot or vehicle cannot wait for a cloud round trip before reacting. Camera streams must be corrected for exposure and motion, neural networks must run locally, and outputs from sensors with different fields of view and update rates must be synchronized. This is why perception growth pulls through not only sensors but also image-signal processors, accelerators, memory, networking and safety control.

Modality by Modality

Cameras: the default semantic sensor

Cameras are inexpensive, information-rich and supported by the largest software ecosystem. They recognize color, text, lane markings, faces, defects and object classes that range sensors cannot infer as naturally. Automotive-grade devices add high dynamic range, LED-flicker mitigation, temperature tolerance and functional-safety requirements.

The investment case rests on content growth and qualification. More capable vehicles and robots use more cameras, at higher resolution, with more local processing. The risk is that image sensors remain components in a value chain increasingly controlled by the processor and software provider. onsemi has established automotive and industrial sensing positions, but its Intelligent Sensing Group represented only a minority of total 2025 company revenue. The sensing thesis is therefore diluted by the larger power and analog businesses.

LiDAR: differentiated geometry, contested economics

LiDAR measures distance by timing or analyzing emitted light. It provides explicit three-dimensional geometry and can be valuable in mapping, industrial automation, robotaxis and other applications where accurate free-space measurement matters.

The debate is not whether LiDAR works. It is whether its incremental reliability justifies cost and complexity in each application. Mobileye’s 2024 decision to stop internal FMCW LiDAR development illustrates the pressure: the company cited progress in computer vision and imaging radar, along with falling prices for third-party time-of-flight LiDAR. Mobileye still uses front LiDAR in its eyes-off architecture, so the evidence supports reduced dependence, not a universal zero-LiDAR conclusion.

Ouster is the most direct listed exposure in this report. Its 2025 revenue increased 52% to $169.4 million, but roughly $22.8 million came from royalties, making $146.6 million of product revenue the cleaner measure of operating demand. The company shipped more than 25,000 sensors and improved gross profit, evidence that industrial and robotics demand is real. The remaining questions are whether product growth can outrun price erosion and whether operating scale arrives without repeated dilution.

Radar: the durable complement

Radar measures range and velocity using radio waves and performs well in darkness and poor weather. Imaging radar improves angular resolution with larger antenna arrays and more processing, narrowing part of the gap with LiDAR while retaining direct Doppler velocity.

Radar is the least binary modality in the automotive debate because camera-first and multi-sensor systems both use it. NXP has long described itself as a market leader in automotive radar and offers transceivers, processors and surrounding vehicle-compute products. Historical company materials show that its technology shipped in more than half of new radar sensors in 2016; that supports leadership, but it does not prove a current 50% market share. We therefore treat the exact share as an external estimate rather than a reported fact.

The risk to radar suppliers is integration. Single-chip solutions can expand adoption while reducing content per module. The stronger franchise belongs to the vendor that wins across transceiver, processing, safety and vehicle networking rather than relying on one chip socket.

Edge vision and perception compute

Raw sensor growth creates a compute problem. Multiple high-resolution streams must be processed within strict power, thermal and latency budgets. Image-signal processors correct the incoming image; accelerators run neural networks; safety processors supervise decisions; high-speed links move data across the machine.

This is the strategic battleground between merchant specialists and integrated platforms. Ambarella offers efficient computer-vision silicon. Mobileye couples EyeQ processors with perception software and mapping. NXP spans control, radar, networking and edge processing. NVIDIA and Qualcomm can bundle perception compute into broader platforms. Integration can enlarge the overall market while compressing the number of independent sockets.

Touch, motion and non-visible sensing

Visual perception is only part of physical intelligence. Force-torque and tactile sensing tell a robot whether it has made contact, how hard it is gripping and whether an assembly operation is aligned. VPG supplies precision sensing components based on strain-gauge expertise. Novanta owns a broader motion and vision portfolio, including encoders, motors, drives and force-torque products through ATI Industrial Automation.

These exposures are appealing because manipulation requires feedback, but the humanoid revenue base is still small relative to the narratives attached to it. VPG reported $1.0 million of humanoid-related orders in the first quarter of 2026 and discussions with a fourth developer. That is evidence of engagement, not proof of mass production. Supplier names, per-robot content and launch timing should remain scenarios until disclosed by the companies or customers.

Thermal, short-wave infrared, event-based vision and inertial sensors serve narrower but important failure cases. Defense and industrial deployments may support better economics because performance, ruggedization and export controls matter more than minimum unit cost. Leonardo DRS is the clearest integrated defense exposure in this group; Teledyne provides broader thermal and imaging exposure.

Where the Economics Accrue

We evaluate each company through two lenses.

Toll-road quality asks whether the product is difficult to bypass, concentrated within a qualified supply base and insulated from direct hyperscaler competition. This is a durability test.

Platform potential asks whether value migrates from a component sale toward software attach, data, workflow ownership or a larger installed ecosystem. This is an upside test.

This framework changes how we read the same headline. Falling LiDAR prices are negative for a component ASP, positive for adoption and potentially positive for a software platform that can deploy more widely. A single-chip radar is negative for socket count but positive for total units. More robot cameras help image-sensor suppliers, yet the largest value may accrue to the inspection software trained on those images.

Company Scorecard

The table is a research ranking, not a current buy list. “Bottleneck” measures durability; “platform” measures the opportunity to compound beyond a unit sale. Current signals and valuation work remain on the individual company pages.

Photonics suppliers such as Coherent and Lumentum can participate through lasers and optical components, but their investment theses are driven increasingly by data-center communications. We treat LiDAR as an optional contributor rather than the primary reason to own either company. Ambarella is a focused edge-vision candidate, while Qualcomm and NVIDIA are powerful competitors whose perception exposure is diluted inside much larger compute franchises.

The Investment Cases

Cognex: the closest thing to a perception platform

Cognex sells into a bounded domain where accuracy has measurable economic value: a missed defect, failed barcode or stopped production line costs the customer money. Hardware, lighting, software and application knowledge are deployed together. Once an inspection workflow is validated, switching requires more than swapping a camera.

That gives Cognex a better path to platform economics than most sensor vendors. The upside case is increasing software content and easier deployment of AI-based inspection across its installed base. The downside case is familiar: factory automation is cyclical, customers can defer projects, and general-purpose AI tools may reduce differentiation at the application layer.

NXP: the diversified radar and control toll road

NXP’s attraction is breadth. It participates in radar, vehicle networking, safety control and edge processing rather than depending on one modality. In the second quarter of 2026, company revenue reached $3.50 billion, up 19% year over year, while automotive revenue was $1.94 billion. Those figures demonstrate scale and automotive exposure; they do not disclose radar revenue separately.

The result is a high-quality but less pure perception investment. NXP can benefit regardless of whether a vehicle emphasizes cameras, radar or LiDAR because all architectures require control, connectivity and safety. Its limitation is upside dilution: investors own the whole semiconductor portfolio, and the most interesting perception products are not reported as stand-alone segments.

Mobileye: software leverage with strategic competition

Mobileye generated $1.894 billion of 2025 revenue, up 15%, and shipped approximately 35.7 million EyeQ SoCs and SuperVision systems. Its advantage is the combination of purpose-built compute, perception software, OEM relationships and mapping data. That moves it farther right in the value chain than a sensor supplier.

The debate is whether that installed architecture compounds or is gradually displaced by automakers and integrated compute vendors. Mobileye’s decision to stop internal FMCW LiDAR development is capital discipline if third-party sensors and imaging radar meet requirements; it is not evidence that perception hardware no longer matters. The key indicators are SuperVision and Chauffeur program conversions, software value per system, operating leverage and the durability of the OEM pipeline.

Ouster: the purest modality bet

Ouster offers the most direct upside if digital LiDAR adoption broadens across industrial automation, robotics, mapping and smart infrastructure. The 2025 product-revenue growth and shipment record support the commercial case, while royalties show an additional route to monetize intellectual property.

Purity creates risk as well as upside. Ouster must offset falling industry prices, maintain differentiation and fund the path to consistent profitability. Its merger and acquisition history also makes organic comparisons important. Investors should track product revenue separately from royalties, sensor shipments, product gross margin, adjusted EBITDA and cash usage.

Precision sensing and defense: VPG, Novanta and Leonardo DRS

VPG and Novanta are complementary rather than interchangeable. VPG is closer to the underlying measurement element; Novanta combines sensing with motion and subsystem capabilities. Their humanoid opportunity is credible but early, so bookings and disclosed production revenue matter more than prototype announcements.

Leonardo DRS sits in a different economic regime. Defense sensing is bought through programs, integrated into mission systems and supported over long periods. The company reported first-quarter 2026 revenue of $846 million and funded backlog of $4.7 billion. That backlog provides visibility, but not all of it is perception-related. The thesis depends on continued demand for EO/IR, network computing and sensor-fusion capabilities, plus disciplined conversion of funded programs into revenue and cash.

Risks, Gaps and Disconfirming Evidence

The most important uncertainty is modality substitution. Better camera models and imaging radar could reduce LiDAR units in passenger vehicles. Conversely, lower LiDAR prices and tougher safety requirements could expand adoption. Evidence should be judged application by application; robotaxi, industrial and mass-market passenger-vehicle economics are not the same.

A second gap is segment disclosure. NXP does not report radar revenue, onsemi does not isolate automotive image-sensor revenue within all relevant disclosures, and diversified suppliers rarely disclose robotics content cleanly. Market-share figures from industry researchers are useful directional inputs, but they should not be presented with the certainty of filed financial statements.

A third risk is narrative outrunning revenue. Humanoid design discussions, prototype content and vendor nominations can precede production by years—or never reach volume. Customer attribution and per-unit content assumptions remain outside the core case until they are independently disclosed. The proof points are orders, recognized revenue, repeat programs and manufacturing investment.

Other disconfirming signals include:

• Sensor count rises but total supplier revenue does not, implying faster ASP erosion.
• Integrated processors eliminate more stand-alone content than adoption creates.
• Software attach and recurring revenue remain immaterial for purported platforms.
• Automotive qualification fails to produce pricing power or share stability.
• Ouster’s product growth slows after separating one-time royalty revenue.
• Mobileye’s advanced-program wins fail to convert into higher value per system.
• Defense backlog grows without timely revenue and cash conversion.

Conclusion and Monitoring Framework

Perception is a structural growth market, but the value pool will move as the architecture matures. We expect cameras and radar to remain the volume foundation, LiDAR to persist where direct geometry earns its cost, and specialized sensors to expand as machines enter less controlled environments. Edge compute and fusion software should grow faster than raw sensing units because every additional channel creates more data and calibration work.

The portfolio implication is to avoid making the entire thesis depend on one modality. A balanced watchlist combines a qualified automotive supplier such as NXP, an industrial platform such as Cognex, a software-rich autonomy company such as Mobileye and selective higher-risk exposure such as Ouster. Defense and precision sensing add different demand cycles through Leonardo DRS, Novanta and VPG.

We would increase conviction when software revenue grows faster than hardware, disclosed sensing revenue outgrows the parent company, production programs replace prototypes and product gross margins hold despite falling sensor prices. We would reduce conviction when integration destroys more content than adoption creates or when a company’s claimed exposure remains too small to affect reported results.

The durable thesis is not that every machine will use every sensor. It is that useful autonomy requires reliable perception—and the companies that own qualification, fusion and workflow will retain more value than those selling undifferentiated sensing elements.

Data Quality and Sources

This report separates three evidence levels:

1. Filed or issuer-reported facts include revenue, segment data, shipments, backlog, product announcements and stated roadmaps. Direct links appear in the source library below.
2. External estimates include market share, market size, sensor content and industry pricing. They are directional and may use different definitions or measurement dates.
3. PXS Research analysis includes the bottleneck and platform assessments, risk ranking and synthesis across modalities. These are judgments, not company guidance.

Financial figures are point-in-time observations through August 14, 2026. The report intentionally excludes live prices and price targets so that a dated industry thesis is not confused with a current trading signal. Readers should use the linked company pages for current PXS Research signals and review dates.