Perciv AI
Perciv AI is radar perception software for automated vehicles and robots. Four AI modules cover localisation, free-space estimation, object classification and radar-camera fusion, targeting LiDAR-grade performance from low-cost off-the-shelf radars in rain, fog, dust or darkness.
What is Perciv AI?
Perciv AI is radar perception software developed by Perciv B.V., a Delft-based deep-tech company founded by TU Delft alumni and incubated at YES!Delft. Its stated ambition is to be the world's first dedicated AI for radar perception. Most automated vehicles and robots still rely on LiDAR, which the company describes as expensive, fragile and unreliable in rain, night, fog, dust or smoke. Perciv replaces it with AI-driven radar and radar-camera fusion running on affordable, off-the-shelf radars.
The product line answers four perception problems. Perciv EgoTrack derives ego-motion and odometry from one or several radars alone, needing neither GPS nor IMU, with translational drift claimed as low as 2% (1% with an IMU), which matters in parking garages, tunnels and urban canyons. Perciv FreeRoad produces radar-only occupancy and height maps that flag static and dynamic obstacles regardless of class, up to 150 m and with recall of up to 99.99% for vulnerable road users. Perciv Objects outputs an object list carrying 3D pose, velocity, class and track; the company claims LiDAR-level classification, first place on the View-of-Delft and ManScenes benchmarks, and four times better classification than the leading Tier-1 solution. Perciv Fusion adds RGB or thermal cameras in early, late or multi-level fusion, and every module ships in a radar-only and a fused version.
Three properties define the offer. It is sensor-agnostic across radar grades, compute-agnostic thanks to networks under 2 TOPs, and modular, each module being configurable and fine-tunable from the company's SDK. Foundational models are trained on heterogeneous data and then fine-tuned to a customer's application, mounting and use case.
Two hardware kits support this. The AI Radar Kit is a turnkey, pre-calibrated, ROS-ready IP67 unit sold in an Off-Highway variant with a Bosch radar, while a Long-Range variant built on the ZF FRGEN21 is announced. The AI Development Kit records training data from sixteen or more PTP-synchronised sensors. Deployments span cars, trucks, forklifts, off-roaders and drones.
What it does
- Detect, classify and track cars, pedestrians and two-wheelers from radar alone, with additional object classes available on request
- Estimate free space and produce an occupancy or height map flagging every collision-risk obstacle up to 150 m
- Compute ego-motion and odometry from radar only, with no GPS and no IMU in the loop
- Fuse radar with RGB or thermal cameras using early, late or multi-level fusion
- Keep perceiving through rain, night, fog, dust and smoke
- Run inference on lightweight embedded compute, with networks under 2 TOPs
- Record calibrated, time-synchronised training data with the AI Development Kit and fine-tune the models to a specific mounting and use case
When to use Perciv AI / When not to
A quick filter to help you decide if Perciv AI is the right fit.
When to use Perciv AI
- ADAS and autonomous-driving teams that need perception to hold up when cameras and LiDAR degrade in rain, fog, dust or darkness
- Robotics and off-highway manufacturers fitting forklifts, tractors, construction machines or drones with obstacle detection and collision avoidance
- Hardware integrators trying to take an expensive LiDAR out of the bill of materials without losing detection and classification quality
- Engineering teams working in GPS-denied environments such as tunnels, parking garages, warehouses and urban canyons, where odometry has to come from the sensors themselves
- Road authorities and research consortia mapping infrastructure or benchmarking weather-robust perception, as with Rijkswaterstaat and the EU EVENTS project
When not to use Perciv AI
- Anyone looking for a general-purpose AI assistant: Perciv AI only turns radar and camera signals into 3D perception output, nothing else
- Buyers who need published pricing and a self-service checkout, since the online store lists no product and every call to action leads to a contact form
- Developers expecting a hosted API with public documentation; an SDK is mentioned repeatedly but is not documented anywhere on the site
- Teams that want to evaluate a tool on their own first: there is no free plan, no free trial and no online demo
- Projects with no radar hardware and no embedded compute budget, because the software always runs on a sensor plus an onboard compute unit
How to use Perciv AI
A typical end-to-end flow, from setup to results.
- Identify your use case, on-road, off-road, indoor or in-air, and the radar hardware you already run
- Get in touch through the contact form: there is no self-service signup and no online purchase path
- Choose between integrating the software on your existing radar and ordering a turnkey AI Radar Kit
- For a kit, pick the Off-Highway variant with its Bosch radar, available now, or wait for the announced Long-Range variant built on the ZF FRGEN21
- Mount the hardware: sensor mount 180x200x85, compute mount 312x195x98, 12 V DC input, typical IP67 rating
- Power the unit up, since kits are delivered synchronised and calibrated and no field calibration is expected
- Consume the perception output through the ROS-ready interface
- Select and configure the modules you need, EgoTrack, FreeRoad, Objects or Fusion, from the SDK
- If the models need adapting to your mounting and scenario, record data with the AI Development Kit, which can be driven from a phone
- Ask for a custom sensor setup or a dedicated perception model when the standard modules do not cover the case
Pros & Cons
Pros
- Takes the LiDAR line out of the bill of materials while claiming comparable perception quality
- Built for the conditions that degrade optical sensors: rain, night, fog, dust and smoke
- Works without GPS or IMU, which covers tunnels, parking garages, warehouses and urban canyons
- Modular by design, with four independent modules that are configurable and fine-tunable from the SDK
- Small compute footprint, under 2 TOPs, widening the choice of embedded targets
- Verifiable research grounding through arXiv and TU Delft publications, public benchmarks and the EU EVENTS consortium
- Named institutional references such as Rijkswaterstaat, RDW, CES 2025 and NVIDIA Inception, backed by a EUR 2.5 million seed round
Cons
- No public pricing whatsoever: the online store lists no product and every purchase path runs through a sales contact
- No terms and conditions and no legal notice, so contractual terms and data ownership stay invisible until you talk to sales
- No free plan, no free trial and no online demo, which makes independent evaluation impossible
- No public API or SDK documentation, even though the SDK is central to the modularity argument
- The Long-Range Radar Kit was still marked as coming soon when this record was reviewed
- Headline performance figures are published without a measurement protocol or a linked source
- No security or functional-safety certification is mentioned, despite safety-critical target applications
Pricing & Plans
Perciv AI publishes no price. The site does carry a Squarespace store, but it lists no product, and every product call to action, including the Order now button on the AI Radar Kit, leads to the contact form. No permanent free plan and no free trial are advertised, and neither is explicitly ruled out. Pricing is therefore quotation-based, and the only cost statement made is a relative one: the company positions its kits at a fraction of the cost of comparable perception systems.
Data, GDPR & hosting
A consolidated view of how Perciv AI handles your data.
GDPR overview
Implementation is concrete and unusually detailed for a company of this size. The privacy statement, last updated 30 July 2026, names Perciv B.V. (Dutch Chamber of Commerce number 87714019, Delft) as controller and cites its legal bases explicitly: Article 6(1)(f) legitimate interest for road recordings, Article 6(1)(b) for business relationships and job applications. Article 11 GDPR is invoked for recordings in which no one can be identified. Perciv says it has documented the balancing assessment, available on request, and carried out a data protection impact assessment as a precaution. Subprocessors are named, Squarespace and Google, with the EU-US Data Privacy Framework and Standard Contractual Clauses covering transfers. Recording campaigns are announced at least two weeks ahead and vehicles carry a QR-coded notice. The Dutch supervisory authority is named. No DPO and no Article 27 representative are designated.
Who owns the data?
The site publishes no terms and conditions, so nothing on it defines who owns the data a customer would process with the software; that gap has to be closed in a contract before any deployment. What the privacy statement does establish is narrower. Perciv B.V. acts as controller for the road recordings made by its own marked test vehicles and stationary sensor units, and states that this data comes solely from sensors on its own equipment and from no other source. Blurred recordings are kept on Perciv's own access-controlled systems in the Netherlands, are never sold, and are not shared with subprocessors or any other third party.
Reuse rights
Because Perciv AI publishes no terms and conditions, the site grants no licence and states no reuse rights for a customer: what may be done with the perception outputs is left entirely to a negotiated contract. The only reuse rules published concern personal data. Road users may exercise the full set of GDPR rights, including access, rectification, erasure, restriction, portability and objection, by writing to info@perciv.ai or using the form on the privacy page. A request to access or remove a specific recording must give the location, date and approximate time; Perciv answers within one month at the latest, extendable by two further months for complex cases. Article 11(2) GDPR is invoked because Perciv cannot identify individuals in the data it keeps, and it commits to acting on removal requests and deleting the material where feasible even when the GDPR does not apply to a given recording.
Data retention & training
Hosting summary
Perciv B.V., established in Delft, the Netherlands, is the controller and the applicable jurisdiction is Dutch, with the Autoriteit Persoonsgegevens named as supervisory authority. Sensor recordings, once faces and licence plates have been blurred, are kept on Perciv's own access-controlled systems in the Netherlands and are not shared with subprocessors or any other third party. The website and its contact form run on Squarespace, while business email and document storage sit with Google. Data collected through the website and the contact form may therefore be processed in the United States by those two providers. Both are described as certified under the EU-US Data Privacy Framework, with the European Commission's Standard Contractual Clauses applying as a fallback. No hosting region is stated beyond these two countries, and no data residency option is offered to customers.
Things to keep in mind
Risks and trade-offs to weigh before adopting Perciv AI.
- Headline performance figures are published without a protocol: 99.99% recall on vulnerable road users, four times better classification than the leading Tier-1 solution, and drift as low as 2% cannot be reproduced from what the site shows
- With no terms and conditions online, liability, warranty and data ownership only become visible once a commercial discussion has started, which is late in the decision process
- This is a safety building block: a perception failure on a vehicle or a machine has physical consequences, so it should be validated in your own conditions rather than trusted on marketing claims
- The all-weather framing invites over-confidence, since neither range nor detection quality is quantified condition by condition
- Test vehicles record public roads, so images of people and licence plates are captured before blurring, on a legitimate interest basis; anyone who may have been filmed should know the objection and removal procedure exists
- Depending on a single supplier for a critical link in a perception chain creates a real continuity risk for a young company
- No functional-safety or information-security certification, such as ISO 26262, ISO 27001 or SOC 2, is mentioned anywhere on the site
Setup & Integrations
Technical difficulty
High, and unavoidably so: this is embedded perception, not a web tool. The turnkey route is the easiest, since the AI Radar Kits ship synchronised and calibrated with a ROS-ready output, need only 12 V DC power and are rated IP67, but they still have to be mounted on a vehicle or machine. The software-only route means integrating the modules onto your own radar and compute through an SDK that is not publicly documented. Fine-tuning to a specific mounting and use case involves recording data with the AI Development Kit and working with the vendor.
Integrations
Behind Perciv AI
Fundraising
Social
Resources
All the official URLs gathered for verification and reference.
Frequently asked questions
What exactly is Perciv AI?
How does it replace a LiDAR?
Which modules make up the product line?
Does it work without GPS?
What hardware do I need?
How much does it cost?
Is there a free trial or a demo?
Is there a public API?
How does Perciv AI handle personal data and the GDPR?
Who are the customers and partners it names?
Should you pick Perciv AI?
Perciv AI is a narrow, serious product rather than a general-purpose AI tool. It addresses one question, how a vehicle or a robot can perceive its surroundings reliably when optical sensors fail, and it answers with radar plus AI instead of LiDAR. The technical credibility is easy to check: publications on arXiv and in the TU Delft repository, results on the public View-of-Delft and ManScenes benchmarks, and membership of the EU-funded EVENTS consortium. Commercial traction is documented too, with a EUR 2.5 million seed round in December 2024, a tunnel-mapping mission for the Dutch road authority, high-speed testing at the RDW proving ground, a CES 2025 appearance and the claim of being already in production with selected partners.
What a buyer will not find is anything resembling a self-service product. There is no price, no terms and conditions, no free plan, no trial and no demo. The only route in is a contact form, which is normal for embedded perception sold to manufacturers but makes an independent first assessment impossible. The performance claims are striking, up to 99.99% recall on vulnerable road users, classification four times better than the leading Tier-1 solution, drift as low as 2%, yet none of them comes with a measurement protocol, and no functional-safety or security certification is mentioned for applications that are plainly safety-critical.
The company itself is young, with a second anniversary celebrated in October 2024 and a team of roughly twenty people. Its privacy statement is unusually thorough for that stage, with a documented balancing assessment, a data protection impact assessment, automatic blurring and a 90-day deletion rule. Worth a conversation for anyone integrating perception into a vehicle or a robot; worth asking hard questions about benchmarks and contract terms before committing.
- Choosing a selection results in a full page refresh.
- Opens in a new window.