ROVR Docs
Last updated
Last updated
There is no doubt that Advanced Driver Assistance Systems (ADAS) and Autonomous Driving will be the most significant transportation technology revolution in the next 20 years; at the same time, in the consumer electronics field, VR/AR technology will profoundly impact human history, much like smartphones did over the past decade. Behind all this lies the rapid advancement of AI technology. However, beneath the surface lies a massive and concerning limitation: do we have the capability to obtain massive, real-world, and precise 3-dimensional spatial data and 4-dimensional spatial-temporal data? This is not as easy as acquiring text, images, and video data on today’s internet.
For the most pressing application at the moment, Advanced Driver Assistance Systems (ADAS), it heavily relies on high-definition 3D maps, which is an obvious conclusion. However, the shortage and high cost of HD maps have severely constrained the development of related fields. People have had to resort to temporary measures, sometimes deliberately downplaying the importance of HD maps (even out of good intentions and to avoid controversy), such as using non-high-precision, non-real-time traditional electronic navigation maps, combined with multi-sensor fusion, attempting to compensate for the shortcomings in accuracy and operational effectiveness. However, this often leads to reliability and safety issues, and even heartbreaking tragedies.
On the other hand, current HD maps data collection is entirely controlled by centralized companies or organizations; it is expensive, complex, and has a very low data refresh rate. Many maps collected ten years ago are still in use. Some great and admirable DePIN projects, attempt to use decentralized methods with smartphones or dashcams to create maps. However, unfortunately, this approach cannot meet the high-precision and data quality requirements demanded by the automotive industry, leading to the project's inability to truly gain customers and achieve large-scale commercial use. This is due to an unavoidable and practical reason: centimeter-level accuracy HD maps exist as the ground truth system and are the most important infrastructure for ADAS and autonomous driving. 3-dimensional maps made solely by recording 2D video streams with smartphones and dashcams cannot prove to be accurate and stable enough.
VR/AR applications represent another exciting direction. Clearly, they involve the high precision and complete reconstruction of 3D scenes. Generally speaking, the development of social and killer applications is crucial for VR/AR. However, an awkward reality is that due to the lack of real 3D data available for AI training, VR/AR applications often fail to accurately reconstruct 3D scenes in coordination among multiple users and seamlessly integrate with other sensors. Therefore, in the foreseeable future, the shortage of high-precision 3D data will greatly limit the development of the VR/AR ecosystem.
Over the past 20 years, we have gradually come to realize that the traditional method of producing HD maps has reached its limit and cannot support the next generation of intelligent transportation revolution. ROVR will adopt a completely new approach to data collection and map creation, and pioneeringly integrate the traditional automotive industry, consumer electronics industry, Web2, and Web3, to prepare massive data for the upcoming comprehensive AI and 4-dimensional spatio-temporal computing.
1.2.1 High-precision Software and Hardware Systems Conforming to Automotive Industry Standards
ROVR will fully adopt onboard RTK (Real-time Kinematic) and PPK (Post-processed Kinematic) systems, supplemented by AGNSS and integrated navigation, to consistently improve the overall system positioning accuracy to the centimeter level
ROVR will employ automotive-grade mass-produced multi-channel solid-state 3-dimensional LiDAR and onboard survey-grade IMU systems. In the collected raw data, approximately 1 million laser reflection points with an accuracy within 2 cm are generated per second to describe the surrounding scene and for 3D reconstruction.
Traditional image sensors in dashcams and smartphones use rolling shutters, which can cause image distortion when subjects move at relatively high speeds. This distortion significantly impacts measurement accuracy. ROVR will use global shutters and fixed-focus lenses to eliminate the 'jelly effect' and image distortion commonly seen in dashcams and smartphone cameras, ensuring accurate measurements.
By employing enhanced smartphone external devices, complemented by AGNSS, RTK, and other technologies, ROVR will enhance smartphone positioning accuracy to the centimeter level. Additionally, ROVR will capture road dynamic elements in real-time and generate dynamic layers for HD Maps.
1.2.2 Open Ecosystem
ROVR is a highly transparent and autonomous project, where any organization or individual can produce, sell, or even DIY their hardware devices within the open-source framework of the ROVR project. They can also participate in data collection, storage, maintenance, and processing, and can join or leave the ROVR network at any time.
The ROVR community plays a crucial role throughout the entire ROVR network. It consists of members from various roles such as data contributors, data consumers, developers, DIY enthusiasts, hardware manufacturers, and computing platform providers. Together, they ensure the healthy operation of the ROVR network.
Despite being a decentralized open system, the ROVR network still requires an organized entity to manage and operate it. This entity is the ROVR Foundation, which is responsible for coordinating and managing software and hardware development, data collection, quality control, network maintenance, and submitting R2FCs (ROVR Request for Comments) to the community for voting.
1.2.3 Fair Token Rewards
Anyone contributing to the ROVR network, whether through data collection or other means, will receive very fair token rewards. These tokens are distributed based on actual workload and cost input. In other words, it's a return to proof-of-work (POW), but with a more meaningful and environmentally friendly approach, capable of generating significant social and economic value.
1.2.4 Permissionless and Censorship-resistant
Every human driver with a moving vehicle exceeding 15 km/h, every individual with a high-performance server, and even those with just a passion, can become contributors in the ROVR network without any permission required. Contributors remain completely anonymous and are not required to provide any personal information. They have 100% autonomy to decide when to join or leave the ROVR network.
1.2.5 Privacy Protection
ROVR places great emphasis on privacy. Data is structured and anonymized in most cases. For instance, ROVR data collection terminals typically only upload structured data without detailed features of the targets. For example, it might define a "pedestrian" and tag their speed and direction without describing any additional information.
Participating in the ROVR network is straightforward for ordinary users and offers multiple avenues:
1.3.1 Using a "Smartphone + Tarantula Mini" for Data Collection
"TarantulaX Mini" is a highly integrated and compact sensor system, including an enhanced RTK (Real-Time Kinematic) positioning unit, an Inertial Measurement Unit (IMU), and a Web3 identity and data validation module. It connects to smartphones via Bluetooth. It utilizes the smartphone's camera to capture road images, process and classify the data, and upload it, thereby earning $ROVR tokens.
Please note that, in most cases, mainstream smartphones can accomplish this task. ROVR will strive to adapt to various smartphone models as much as possible. However, not all smartphones are suitable for this task, as it depends on hardware and software configurations, particularly the camera's optical parameters, AI computing power, and encryption capabilities.
1.3.2 Using “LightCone” (Professional 3D Lidar Collection Device) for Data Collection
The “LightCone” is a professional device equipped with a multi-channel lidar (>120 channels, 200 meters measurement range), capable of directly sensing the surrounding 3-dimensional environment in real-time. It also integrates automotive-grade integrated navigation systems and large-aperture (>130mm) measurement antennas.
During usage, users still need to connect the device to their smartphones via Bluetooth or Wi-Fi. However, the smartphone is only used as a control terminal, with no special requirements and no need to stay online constantly.
It's worth noting that whether it's the “Tarantula Mini” or the “LightCone”, the user device's speed must be between 15 km/h and 140 km/h. Rewards will not be generated if the speed exceeds or falls below this range. The reward is related to the number of kilometers of quality data.
The qualification of data depends on whether it includes paved roads with clear lane markings and other traffic elements. According to road network data definitions, these are classified as Level 8 and above roads, typically encompassing highways, urban expressways, ramps, and city streets.
1.3.3 Providing R-Node nodes and staking $ROVR tokens
R-Node serves as the data storage and processing node for the ROVR network. Users can provide nodes that meet the requirements of hardware, software, and network conditions, and stake $ROVR tokens to earn $ROVR token rewards, including basic rewards and additional workload rewards.
The R-Node node plan will be launched in the second phase of the project. In the initial stage of the project, ROVR will utilize traditional cloud servers and OSS services.
ROVR Inc. is responsible for cleansing, processing, and ultimately transforming the data within the ROVR network into HD maps and other data products, which are then sold to end customers.
ROVR Inc. can provide customized maps or other data services according to the requirements of OEM/Tier 1 or other clients.
When the time is right, similar to hardware development and production, ROVR will open up third-party data processing. Any organization or individual will be able to perform secondary processing on ROVR data and sell it to end users. This will be an important component of the ROVR ecosystem.
There are two project entities: ROVR Foundation and ROVR Inc.
Responsibilities of ROVR Foundation:
Maintain the overall token economics and project ecosystem.
Output hardware reference designs and software designs.
Data collection and storage.
Maintain technological evolution and guiding principles.
Responsibilities of ROVR Inc:
Creating maps products or other forms of data products and selling them to customers.This job will be done by third parties in the near future.
ROVR deeply understands the importance of data privacy and compliance. It's important to note that in most countries and regions, the collection and use of public road map data are generally legal. However, in some countries and regions, the collection of public road map data may be restricted by law. Major considerations include privacy protection, national security, and commercial competition. In a few countries and regions (such as mainland China, etc.), unauthorized surveying is a serious offense. Therefore, in these specific countries and regions, ROVR hardware devices cannot be used, data generated will not be accepted by the ROVR network, and no rewards will be generated.
Given ROVR's decentralized nature, users are advised to understand the relevant laws and regulations of their country in advance to avoid breaking the law or causing unnecessary trouble.
Tarantula Mini serves primarily as an RTK receiver and also signs location data to prevent malicious replay attacks or forgery, ensuring its authenticity and validity.
Here's a brief schematic(Future modifications and upgrades may occur, but updates to this document may not be timely):
Tarantula Mini connects to the NTRIP service provided by RTK providers (e.g. GEODNET) through a smartphone to obtain nearby RTK base station information data. It then acquires real-time observation files from the base stations and satellite ephemeris data to perform real-time position correction at one-second intervals. Typically, Tarantula Mini can improve the positioning accuracy of a smartphone to within 1 cm.
On the other hand, Tarantula Mini uses an unreadable hardware private key to digitally sign high-precision location data to prevent data from being maliciously tampered with or forged.
For detailed technical specifications, please refer to the ROVR open-source repository.
LightCone is a much more complex system compared to Tarantula Mini. It consists of the following components:
Specs: (Future modifications and upgrades may occur, but updates to this document may not be timely)
Detect Range: 200m (905nm Lidar), 500m (1550nm Lidar), @10% NIST
Dimension: 135mm x 220mm x 120mm
Weight: 1.6kg (905nm), 3.3kg(1550nm), not include cables and support parts
Power Consumption: 35w (905nm), 50w (1550nm)
Storage: Onboard 64GB eMMC + 1T SSD
Waterproof: IP67
Accuracy: < 2cm (1sigma)
Camera Resolution: 4 K
Lidar Resolution: 0.1(horizon), 0.2(vertical)
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 °C ~ 85 °C
Storage Temperature: -65 °C ~ 105 °C
HD maps refer to map data with high spatial and temporal resolutions, typically used in applications such as autonomous driving systems and other scenarios requiring precise positioning. These maps not only contain basic geographical information like road and building locations but also include more detailed information such as lane markings, traffic signals, traffic signs, road curvature, elevation changes, and so on.
In the production process, HD map mapping typically involves the use of high-accuracy sensors and measurement devices such as LiDAR, cameras, Global Positioning System (GPS), etc., to gather map data. This data undergoes meticulous processing and map-making algorithms to generate maps with high accuracy and reliability.
In terms of data content, HD maps typically consist of three main components:
3.1.1 Road Model
This includes information at the road level, such as road geometry, road attributes (such as passability), lane information within the road, and information about objects on the road.
3.1.2 Lane Model
This contains information at the lane level, such as lane geometry, lane connectivity, lane attributes (such as direction of travel, whether it can be crossed), and the relationship with the object model.
3.1.3 Object Model
This encompasses the geometry, orientation, category, and corresponding relationships between objects and lanes.
For HD mapping, we use professional equipment, including solid-state LiDAR, a 4K camera, and high-precision positioning devices.
Parameter calibration is indeed the first step. The internal parameters of the sensors are calibrated during the manufacturing process. As for the external parameters, we have integrated an automatic calibration program, so users do not need to perform this operation.
After users complete data collection, we aggregate the local mapping results in the cloud. Typically, the requirements for HD maps are an absolute accuracy of less than 50 cm and a relative accuracy of less than 20 cm within 100 meters.
All of this data is used to automatically generate HD map data. Point cloud data provides centimeter-level ranging capabilities, while image data offers rich semantic information. Precise pose data improves the overall consistency and positional accuracy of HD maps. However, due to the high precision requirements set by OEMs, manual editing and review are still necessary after automatic generation.
Starting in Q3 2024, we will use the mini device TX for HD map mapping, and our app will support online mapping. Currently, the data transmission rate is 100kB/km, and it is expected to decrease to 50kB/km by Q2 2025.
After Q2 2025, we will release our professional device LC, which will directly support online mapping
For HD map updating:
Firstly, we will perform 3D reconstruction of consecutive images using the stereo vision method.
However, due to the latency during camera exposure, the result of visual 3D reconstruction is non-rigidly related to the existing HD map point cloud. We will simulate the imaging process to project the HD map point cloud onto the image coordinate system and match it with real-time collected images (mainly edge features) to optimize the camera's intrinsic and extrinsic parameters.
We optimize the camera's intrinsic and extrinsic parameters because the smartphone holder cannot maintain a rigid connection for a long time, causing the smartphone's extrinsic parameters to change. Additionally, the presence of the windshield can cause changes in the relative position between the smartphone lens and the windshield, leading to changes in the entire measurement system's intrinsic parameters.
After obtaining the optimized camera parameters, we will estimate the depth for each pixel, estimate the accurate depth information through monocular depth estimation, and compare it with the existing HD Map to complete the HD Map updating.
In the era of LLM, data is a precious resource, and its importance is particularly prominent in the field of vision. Unlike SORA, our 4D data generation model is based on the real world and allows for precise scene editing (such as defining accurate three-dimensional positions, dimensions, orientations, velocities, etc., for a vehicle), which is crucial for training large models for tasks like autonomous driving. Below is a demo video based on the Kitti dataset. In the video, we replaced all different types of vehicles with the same vehicle model.
Furthermore, we have obtained LOI from a Tier 2. However, for commercial reasons, the video we are showcasing is generated using the KITTI dataset, a opensource autonomies driving dataset.
“Thank you very much for your strong support for the binocular data generation project based on AIGC technology. After evaluation by the R&D team, the generated data and annotation results provided by your company meet the training requirements of the binocular perception model. I am sending this email to confirm our cooperation. Please start the formal development work from now on. The specific amount and deliverables shall be subject to the commercial contract”
— LOI email sent to us by a Tier 2 company
The data we generated has completed usability validation, and the client is already using it in a real production environment.
Cheating and spoofing is a common issue in DePIN projects. ROVR has learned from the lessons of many DePIN projects. With multiple anti-cheating / anti-spoofing technologies, ROVR will no longer require a centralized denylist to ban malicious users. In the ROVR network, all nodes are designed as trustless models, and the system only accepts verified data—nothing more. Users can attempt to cheat without worrying about being denylisted. However, cheating and spoofing will simply not succeed.
All ROVR hardware is designed or integrated with a specialized encryption chip. This chip has a feature where it stores a private key that cannot be read from the outside but can be used to digitally sign the output data of ROVR hardware devices.
During output, ROVR hardware devices, including TarantulaX and LightCone, use the encryption chip to sign their data. When sending data, the device transmits the data itself, the signed digital fingerprint, and the public key of the encryption chip. The data is also packaged according to the blockchain's data structure. The ROVR backend system receives the data and uses the public key to verify the signature, ensuring the authenticity of the data. Data with failed digital signatures will be discarded.
In DePIN projects, spoofing of device latitude and longitude information is common, with a common approach being NMEA message spoofing to carry out replay attacks. However, in addition to providing GPS latitude and longitude data, ROVR hardware devices also include raw satellite observation files and precise ephemeris files.
These files include information such as GPS time, quantity of satellites observed at that moment, their positions, velocities, heading, and DOP, etc. Additionally, this data is compared with parameter information from nearby base stations provided by RTK service providers (currently GEODNET). This makes spoofing highly complex. Additionally, for the quad-constellation GNSS system used by ROVR, the precise ephemeris files come from the GPS、GLONASS、Galileo and BeiDou. Forging such files without detection would likely be a feat only achievable by extraterrestrials.
Easily spoofed NMEA messages
Imagine forging this 250,000-line file:
ROVR hardware includes multiple sensors that create synergistic effects in the same scene. For example, when a user drives over a speed bump, the image will obviously shake. At the same time, the IMU will detect this shake, the LiDAR will experience the same shake, and the RTK system will also reflect similar vibrations in the precise positioning data.
A malicious user attempting to forge values from four different sensors and simulate the same kinematic model would find it extremely challenging.
Total fixed supply: 10 billion
Issuance Network: Solana
Allocation Structure:
51% as rewards for contributors, to participate in building the ROVR Network
20% for project founding team members, and future global core contributors for project R&D and system construction
20% for project investors
9% for ROVR ecosystem including allocation includes liquidity, market operations, promotions, and other aspects essential for sustaining ecosystem development
Category
% of Supply
Token Amount
Community Contributors
51.00%
5,100,000,000
Team
20.00%
2,000,000,000
Investors
20.00%
2,000,000,000
Ecosystem
9.00%
900,000,000
TarantulaX (TX): 1.6 $ROVR per km
LightCone (LC): 16 $ROVR per km
Starting from TGE, the base rewards for $ROVR will be halved every year
1
1.6 $ROVR
16 $ROVR
2
0.8 $ROVR
8 $ROVR
3
0.4 $ROVR
4 $ROVR
4
……
……
The quality of the collected data (e.g., clarity, brightness, RTK accuracy, etc.) will impact the amount of $ROVR received. The final assessment will be categorized into quartiles:
A - Excellent
100%
B - Good
75%
C - Average
50%
D - Below Average
25%
F - Failing
0
*”F” will be assigned to any driver using multiple TX or LC on a single vehicle
Each road will see a 50% reward reduction for every 2 times additional collections.
1st - 2nd
100%
3rd - 4th
50%
5th - 6th
25%
7th - 8th
12.5%
……
……
Road Revisits are calculated globally.
Road Revisits Records Reset Weekly: Revisit records will reset on the Monday of each week at 00:00 GMT.
$ROVR will be transferred within 1 week after data upload, data older than 2 weeks won't be accepted.
*Tip: Before driving, check the ROVR App map and follow the guidance to maximize your rewards.
The additional $ROVR reward multiplier will be set based on the requirement of future data customer, using time and spatial range (NDS 13-level grid) as criteria.
$ROVR operates on a permanent burn mechanism, with buyback funds sourced from data consumers purchasing ROVR's data products:
60% is used to buy back circulating $ROVR and permanently burn it.
20% covers RTK service fees—used to buy back and permanently burn $GEOD from GEODNET.
20% supports ROVR's operational expenses.
