Journey into ROVR - Unveiling the Veil
Hello everyone, Welcome to the first article in our Journey into ROVR series—Unveiling the Veil. This article will provide a detailed introduction to our project's basic information, principles, and product plans.
First, who are we and where do we come from?
We are a four-person startup team from mainland China. Our primary work backgrounds span the fields of automotive, autonomous driving, sensor manufacturing, communications, big data, and financial investment. We have previously worked at world-renowned companies such as Tencent, Alibaba, Baidu, the European Space Agency, and China Ping An Insurance.
Our team members have led the development and production deployment of the world's first pure vision crowdsourced map updating algorithm, as well as the R&D of Velodyne Lidar in the mainland China market.
Prior to this, we also have successful entrepreneurial experience. This May, we secured angel funding from Outlier DePIN Accelerator, in collaboration with Borderless and Peaq.
Second, what are we doing?
Our project is ROVR, which is a decentralized, highly precise multidimensional digital twin platform. The vast amounts of high-definition 3D data it collects will fuel the engine of next-generation intelligent transportation and advanced spatial artificial intelligence.
What exactly are we doing?
The era of AI has arrived, and AI models require vast amounts of training data. Natural language models, such as ChatGPT, benefit from thousands of years of human written civilization, providing abundant text data for models to learn from. However, in the field of visual models, the lack of 3D data is a significant challenge.
Our aim is to address this by leveraging the DePIN (Decentralized Physical Infrastructure Network) approach. We plan to use specialized hardware and software to collect real-world 3D data. This data will then be provided to companies or individuals in need, with the profits shared among the contributors.
Our principle
Data Empowerment, Rejection of Monopolies!
In the era of AI, we aim to break the monopoly of large corporations over core resources like 3D data. We want the vast contributors to not only be consumers but also beneficiaries of the AI era—collecting data and sharing in the profits. We invite everyone to join us in building the greatest project of the AI era. Join the ROVR Official Community and become part of our team!
Why us?
In the market, there are many competitors using dashcams to map road elements, traffic flow, and more. They claim they can generate corresponding elements using data from onboard dashcams. However, why do we believe we can succeed?
There's a simple and clear principle: the success of a venture isn't just about how far we've come but whether we started on the right path from the beginning. While many competitors are on similar paths, does that necessarily mean those paths are correct? Below, we'll explain why ROVR is on the right track from both a methodological and outcomes perspective. If you agree with our perspective, we hope you'll join us.
Despite our competitors using similar methods for similar tasks, it appears they haven't clarified whether their claimed data generation from a single device revisit is possible or requires multiple revisits.
For generating high-precision data like HD maps, our answer is clear: we require multiple revisits.
Actually, this question can be reframed as whether expensive professional surveying equipment can be replaced by consumer-grade crowdsourced devices. This prompts us to ask, why use professional-grade equipment at all?
From the perspective of error analysis theory, any measurement system exhibits two types of errors:
Systematic Error: This arises due to inherent characteristics of the system during measurement. It is regular and predictable.
Random Error: This occurs due to unpredictable factors during measurement. It is irregular and unforeseeable.
Expensive professional surveying equipment, owing to its excellent component performance, minimizes both random and systematic errors. In contrast, consumer-grade crowdsourced devices, due to their lower-quality components, exhibit larger random and systematic errors.
Fortunately, random errors in nature generally follow some form of distribution (typically normal distribution). This allows us to mitigate systematic errors by increasing the number of observations, estimating the distribution of random errors, and subsequently reducing them.
After addressing random error issues, the next challenge is to reduce systematic errors, particularly the positioning error critical to map collection systems. Let's analyze the specific applications of systematic and random errors in map collection systems:
Random Error: The primary random error in map collection systems typically stems from the ranging system, especially in 3D detection or mapping using pure visual methods. For instance, in pure visual methods, a 3% error rate is common at a distance of 30 meters (based on engineering experience). By revisiting multiple times and conducting statistical analysis, we can estimate the distribution of random errors and determine their expected value.
Systematic Error: The most crucial aspect of systematic error is positioning accuracy, which needs to be minimized as much as possible. However, the challenge lies in determining how low we need to reduce positioning errors. Generally, a measurement system needs to exceed the accuracy requirements of the final product by an order of magnitude. For HD maps, this typically translates to an absolute accuracy requirement of 50 cm and relative accuracy of 20 cm. Therefore, the positioning error requirement should be less than 5 cm, the positioning accuracy of 15-50 cm is sufficient, is a lie.
ROVR's hardware devices
Based on the methodology discussed above and practical experience, we are preparing to launch the following two hardware devices: Tarantula X & LightCone.
TarantulaX
Dual-band GNSS (L1/L5)
NMEA and RTCM3.2 Messages (RTK/PPK)
Positioning Accuracy: ±2 cm
Hardware Data Encryption for Blockchain Applications
Dual Patch Antenna: 35mm x 35mm + 25mm x 25mm
6-axis IMU
Waterproofing Grade: IP67
LightCone(Coming soon)
Lidar Detect Range: 200m (180m @10% NIST)
Lidar Resolution: 0.2° * 0.2°
Accuracy: ±5cm
Camera: 4K HD
IMU bias instability (Allan):
Gyro: XY<2.6 degree / h, Z<2.1 degree /h;
Accelerator: <45 ug
IMU bias (XYZ): 0.13 degree / √h, 0.050m/ √h
AI Computing Power: 8 TOPS
Working Temperature: -45℃ ~ 85℃
Storage Temperature: -65℃ ~ 105℃
Waterproof: IP67
Tarantula X (TX) Device:
We utilize a smartphone camera paired with an external GNSS antenna (with built-in RTK service) to achieve high precision positioning (2 cm). It is worth mentioning that we have partnered with GEODNET, one of the most competitive RTK service providers. All our devices will integrate GEODNET's RTK services. Compared to competitor dashcams, we opt for smartphone cameras for the following reasons:
Cost Efficiency: By reusing existing smartphone cameras, we significantly reduce user costs, enhance ROI, and rapidly scale our community with low-cost hardware.
Reuse of Computational Hardware: The computational power within smartphones far exceeds that of typical dashcams available on the market.
Scalability: Leveraging a smartphone app allows easier software upgrades. Unlike integrated dashcam solutions that require OTA updates for each device upgrade—increasing system complexity and failure rates (more than tenfold for software updates).
Of course, using smartphone internal cameras presents new challenges, namely the differences between cameras of various models. To address this issue, we will employ software-based fixed focus settings while capturing intrinsic and distortion parameters.
Focal Length: We will standardize the focal length to EFL=3.9mm, corresponding to a wide-angle camera. This focal length is based on our experience from selecting visual mapping equipment over thousands of kilometers.
Camera Distortion Parameters: We will utilize the retrieved distortion parameters to first perform distortion correction. The image below shows our calibration test results: the left image depicts before distortion correction, while the right image shows after distortion correction.
Through these measures, we can minimize differences among various smartphone camera models. Based on actual testing, mapping accuracy can achieve an absolute precision of 50 cm and relative precision of 20 cm (tested prototype, performance consistent with production devices).
LightCone (LC) Device:
For our second professional device, LightCone (LC), we have integrated a LiDAR system and designed it as an all-in-one solution. Pre-sales are expected in Q4 2024, with official release in Q1 2025. The differences between LC and TX are as follows:
LiDAR Integration: LC incorporates LiDAR technology, offering significant improvements in ranging accuracy and range compared to pure visual solutions. For instance, a competitor's stereo camera with a 15 cm baseline achieves roughly 3% accuracy at 30 meters (empirical value), whereas lidar can achieve ±3 cm accuracy at 150 meters with 10% NIST. This represents more than a five-fold increase in ranging capability.
Automotive-grade LiDAR: LC uses automotive-grade LiDAR, with substantial potential for cost reduction in the future (expected 90% price decrease within three years) and high reliability.
User Benefits: LC's powerful ranging capabilities and support for 3D data generation mean users can earn more token rewards compared to TX for the same effective mileage.
These features position LC as a robust choice for advanced data collection needs, offering enhanced precision and future scalability in comparison to TX.
ROVR's Data Products
Our decentralized HD Map and 3D Generation Data technologies have already been adopted and endorsed by several international vendors. We have secured multiple Letters of Intent (LOIs), and commercial orders are currently underway.
Decentralized HD Map
Here is our test data from in Porte de Saint-Cloud, Paris. We consider this to be an extremely complex scene, covering urban roads, highways, roundabouts, and overpasses. Additionally, we support output for traffic signs, road markings, streetlight poles, and other elements (not just traffic signs).
In fact, we believe that mature, mass-producible HD map data products are more convincing than any amount of promotion, even a series of tweets. So, let's go straight to the comparison.
Other Project
ROVR
Features
Traffic Sign
Traffic Sign, Crosswalk, Lane Marking, Road Marking, pole, etc.
Attribution
Lat, Lon, Height, Heading
+ Length, Width, Color and Type
Online/Offline
Offline
Online
Recall Rate
60%+ in pink hex (small-scale testing, in Palo Alto)
90%+
Accuracy
Absolute Accuracy. 1.5m
Absolute Accuracy. 0.5m Relative Accuracy. 0.2m
Geographic Units
Uber's H3 which is used for ride-hailing, logistics, etc.
NDS level 13 Grid which is used for navigation and autonomous driving
3D Generation Data
ROVR's 3D Generation Data is oriented towards training AI models, significantly differentiating it from marketing-focused data/video generation models like SORA. We provide training data to Visual Language Models (VLMs). Our primary clients are autonomous driving technology firms and AI model companies.
ROVR's 3D Generation Data is used for training models in production environments, not just for demo.
Our sample data have been widely applied in our partners' model development efforts. The video below showcases a road-side perception model trained using ROVR's 3D Data Generation technology (Original footage cannot be displayed due to commercial reasons).
ROVR’ Product Plan
Hardware Product Plans
TX: We plan to start pre-sales of TX devices in Q3 2024 and gradually scale production. We anticipate delivering 10,000 units by Q4 2025. (Join the waitlist for a chance to receive free trial devices!)
LC: Pre-sales for LC devices will commence in Q4 2024, with an estimated delivery of 1,000 units by Q4 2025.
Data Product Plans
Decentralized HD Map: We completed technical validation in Q2 2024 and aim to release a complete city-scale HD Map by Q2 2025. By Q4 2025, we plan to offer HD Map data covering a larger area (expected to exceed 100,000 km, covering an entire country).
3D Generated Data: Our first batch of production-oriented 3D generated data will be available by Q3 2025, supporting customizations for clients.
In conclusion
Thank you very much for taking the time to read through our article. We eagerly look forward to your participation in accelerating ROVR's growth and achieving greatness together.
For more information, please visit our official media channels:
🔸 ROVR Network (***https://rovr.network/#/home***) 💬 X: ROVR (***https://x.com/ROVR71776***) 🗯 Discord (***https://discord.com/invite/RjV3E3u4F2***)
Thank you!
ROVR Team
Last updated
