Signal-chain integrity for autonomous systems
Know when
your sensors
stop telling
the truth.
Sensors degrade in silence, and every autonomous system keeps trusting them. obsurver watches the signal chain from the outside and catches the degradation before a wrong picture becomes a wrong decision.
45 minutes with the founding team, on making your autonomous systems safer
- Read-only
- Software-only
- Sensor-agnostic
In the field, sensors soil, drift and misalign over years. That slow slide is sensor degradation: not a hard failure, but a signal that quietly stops being right. The built-in self-test only knows two states, healthy and dead. Everything in between is invisible to it. This is how the industry watches a sensor age today:
Years pass between healthy and dead, and nothing watches them. Perception keeps acting on the degraded signal, confidently wrong. The built-in test asks "is it alive". obsurver asks "is it still right".
The problem
The blind spot every autonomous system runs with.
An autonomous stack is judged on its outputs, and nothing in the chain verifies input integrity beyond alive-or-dead health checks. When the signal chain degrades, perception keeps producing confident output from corrupted input. No fault code ever fires. You end up paying for it everywhere except the place it started.
Training data
Degraded frames enter datasets unflagged. KPIs turn unstable and nobody can say why.
Triage
Phantom bugs that never reproduce. Engineering weeks burned chasing a model that isn't broken.
Validation
SOTIF evidence gaps and slower sign-off, because input integrity was never measured.
In the field
A perception fault with no attribution. The blame lands on the stack, not the sensor.
You don't have to take our word for it.
Browse the full research library →ADAS are not always reliable in long-term operation. An inspection body states that assistance systems cannot be assumed safe over the vehicle's life.
AAA · 2021Measured sensor performance loss from real-world wear. The American Automobile Association documented degradation in ADAS sensors under everyday conditions.
NHTSA / US DOTThe US regulator studied it formally. A federal report on the safety implications of potential ADAS sensor degradation.
US Navy · 2019Defense is funding detection R&D. The US Navy commissioned dedicated sensor-degradation detection for uncrewed vessels, because nothing existed.
IET · 2025Robot IMUs degrade without warning. Peer-reviewed work shows industrial-robot sensors wearing gradually, with failures arriving unannounced.
Journal of Sensors · 2025Sensor drift is unresolved in agriculture. Heat, humidity and dust degrade field sensors; the only mitigation today is manual recalibration.
The product
One engine. Three ways to run it.
The obsurver engine measures signal-chain integrity independently of your perception stack. Every deployment below runs the same detection core.
The obsurver engine
The detection core
A fixed library of physical metrics, measured at both interfaces against each sensor's own baseline. Read-only, sensor- and vendor-agnostic.
01 · Embedded
The watchdog
In the vehicle
The engine runs on the vehicle compute as a real-time watchdog. Read-only, no cloud connection required. It monitors the full signal chain while the system operates, and raises a flagged degradation event with severity the moment a sensor departs from its baseline.
02 · Off-device
On recorded logs
In development and service
The same engine runs offline on the drive logs you already record. It surfaces degradation patterns across recordings over time, produces the input-integrity evidence SOTIF and type approval ask for, and turns blanket service intervals into condition-based maintenance.
03 · Fleet
The cloud layer, fed by both
Aggregates embedded and off-device runs
Everything the watchdog and the log runs produce flows together here, on your infrastructure. One view of which vehicles are degrading, where and why, ranked by severity, for fleet operators and program managers.
What an alert looks like
Every metric is a defined physical quantity, computed in real time against the sensor's own baseline. When one departs, the flag names the sensor, the metric, the deviation and the severity, time-stamped.
Nothing new on the operator's screen, and nothing new to do: alerts ride your existing interfaces, at asset level for maintenance and at system level for the decision.
sharpness-18% · FLAGGED · severity 2
noise
saturation
timing
flag out: front camera · sharpness · -18% vs baseline · severity 2 · time-stamped (illustrative)
How it works
Read-only on two interfaces. Never inside your models.
obsurver clamps onto the sensor frame before perception and onto the perception output after it: a black-box approach that needs no access to your models, weights or targets. A fixed library of physical metrics, measured against each sensor's own baseline. No model to train, none to drift.
Sensor hardware
camera · radar · lidar · IMU / GNSS
Signal chain
drivers · ISP · time sync · transport
Perception
fusion · detection · tracking
Localization
pose · HD map
Prediction · Planning · Control
trajectory · actuation
obsurver
read-only sensor-integrity module
At the pre-perception interface
Every frame is measured against the sensor's own baseline: sharpness, noise, saturation, soiling, geometric drift and timing.
At the post-perception interface
The output is correlated against the pre-perception signal. Environment is separated from real degradation, and each fault is attributed to its source sensor.
No hardware. No write access. No visibility into your models, weights or targets, and never in the control path.
Without measurement
A degraded frame passes as valid. The output is confident and wrong, the dataset quietly picks it up, and the failure surfaces in the field.
With obsurver
The frame is flagged at ingest, the fault is attributed to its sensor, and the record becomes homologation evidence. Your models never notice a thing.
Use case
The same degraded sensor comes back three times.
Once in development, once at homologation, once in the field. Each time it costs more, and each time your stack takes the blame. The sensor can be a camera, a radar, a lidar or an inertial unit: the sequence is the same.
01
Silent drift
A sensor in the fleet degrades. Every health check still reports healthy.
Development
02
Phantom regression
Degraded frames enter the dataset. A perception KPI drops, and weeks go into triage on the model side.
Development
03
Evidence gap
Type approval asks for proof of the sensing performance limits. Nothing in the toolchain ever measured them.
Homologation
04
Field incident
The OEM sees a perception fault in service. The cause sits upstream, in the signal chain.
Field
05
Attribution dispute
You cannot show the inputs were out of spec, so the fault stays with your stack.
Field
Caught at the source
The sensor is named in development
Degradation is flagged pre-perception, confirmed and attributed post-perception, and written to an integrity record.
Illustrative sequence. The order does not vary, only the interval between the moments.
A drift no health check flags. The same braking command ends three meters later, inside the crossing.
ROI
An unmeasured signal chain is a tax on every budget below it.
Your stack rests on one silent assumption: that the inputs are what you think they are. obsurver turns that assumption into a measurement, and every budget below it gets cheaper.
Data budget
Contaminated frames stop polluting the dataset. Every kilometer driven yields more usable data.
Recovered in: data collection & curation
Engineering capacity
Every KPI regression starts with an answer instead of a hypothesis: was the input healthy?
Recovered in: R&D triage, every release
Validation & homologation
SOTIF and type-approval evidence accrues as a by-product of operation. Sign-off arrives sooner.
Recovered in: the program timeline
Field operations
Environment is separated from real degradation, and condition-based maintenance becomes a feature you resell.
Recovered in: operations & warranty
Why now
The ROI is immediate. Regulation makes it mandatory.
The budget case above stands on its own. What turns it from smart into unavoidable is that European engineering standards, type approval and liability law now converge on one requirement: prove that the sensing your stack depends on performs within spec, at approval and across the whole service life.
The SOTIF standard
ISO 21448 · ISO/TS 5083
Safety of the intended functionality
Sensor performance limitations, including degradation, must be identified as triggering conditions and tied to measurable safety KPIs. ISO/TS 5083 carries the duty into automated-driving safety, and ISO/PAS 8800 extends it to AI-based systems.
Demands: evidence of performance limits
The functional-safety standard
ISO 26262 · Ed. 3
Functional safety, revised
The upcoming third edition of the automotive functional-safety standard tightens the duty to monitor safety-relevant electronics in operation. Detecting degrading sensing moves from good practice toward hard requirement.
Demands: degradation monitoring in operation
The type approval
UN R157 · EU 2022/1426 · UN GTR
Approval for automated driving
Demonstrated perception performance, an audited safety concept and continued in-service reporting, with a UN Global Technical Regulation on automated driving in development to carry the same duties worldwide.
Demands: proof at approval, then in service
The liability regime
EU PLD 2024/2853
Product liability for software and AI
Software is explicitly a product under strict liability, component suppliers are jointly liable, and defectiveness is presumed where evidence is not disclosed. Applies from 9 December 2026.
Demands: a record you can produce in court
Industries
One observation principle, every autonomous platform.
obsurver enters through the automotive stack, where regulatory pressure is highest, and the same capability carries to every platform that depends on its sensors to act in the world.
Autonomous stack providers
When perception fails, the stack takes the blame. The stack that can prove its own inputs wins the bid, passes homologation sooner and holds the evidence when it matters. obsurver embeds as that proof, without touching your models.
Defense and uncrewed systems
A sensor that wears out and a sensor that is attacked look the same to the platform. obsurver reports which sensor stopped being trustworthy, runs fully offline where no cloud is allowed, and uses no AI in the watchdog: nothing to train, nothing to poison.
Humanoid and industrial robotics
Collisions and line stoppages start as drift in safety margins. One read-only module covers mixed fleets across vendors, and what one platform teaches carries into every other.
Autonomous agriculture
Dust, vibration and long seasons make field machinery one of the hardest places a sensor will ever work, and harvest the most expensive time to find out one has drifted. obsurver replaces blanket manual checks with condition-based maintenance from the logs machines already record.
The company
Built by people who helped write the standard.
obsurver was founded in Sindelfingen, in the middle of the German automotive industry, by three founders who have spent their careers in autonomous sensing. Two of them co-founded the IEEE P2020 working group on automotive image quality, and the team has published peer-reviewed research on sensor degradation since 2021.
obsurver is developed together with stack providers, suppliers and research partners, so that it fits real architectures rather than idealized ones.
Fabian Schmidt
CEO & Co-Founder
Serial founder with 7+ years of tech leadership and an M.Sc. in Entrepreneurship from Babson College. Built companies across consulting, supply chain and mobility, and is one of Germany's youngest guest lecturers for AI and entrepreneurship.
LinkedIn →
Benjamin May
CTO & Co-Founder
M.Sc. in Physics, University of Greifswald. Spent 15+ years leading ADAS and autonomous-driving system development for global OEMs, and co-founded the IEEE P2020 working group on automotive image quality.
LinkedIn →
Dr. Sven Fleck
CSO & Co-Founder
M.S. and Ph.D. in Computer Science, University of Tübingen. Advises premium OEMs on imaging, is co-founder and Vice-Chair of IEEE P2020, and has 35+ scientific publications alongside expert-reviewer work for the European Commission.
LinkedIn →
Backed by
Next Mobility Labs
obsurver was built inside Next Mobility Labs, the global mobility venture studio. NML acted as institutional co-founder: strategic guidance and access to a mobility and industrial network across Europe.
nextmobilitylabs.com →
Göran Göhring
MD, Next Mobility Labs · Angel Investor
Peter Mertens
Former Audi Board & Volvo Cars CTO
Joachim Langenwalter
Board dSPACE · Former NVIDIA Director
Sabina Jeschke
Advisory Board Rheinmetall & Aumovio
Dirk Wollschläger
Former General Manager, IBM Global Automotive
Bram Schot
Former Audi CEO & Board of Volkswagen
Next step
See what your signal chain is really delivering.
A 45-minute technical session with the founding team, on your architecture, your sensor set and your own recorded data.