Spexi Taps 10,000 Drone Pilots to Map Earth for Physical AI Models

Closing the Sim2Real Gap with Living Drone Imagery With Spexi — Ai4

Ai4

Bridging the sim-to-real gap requires continuous, decentralized drone captures because static satellite imagery and synthetic simulations cannot provide the geometric ground truth physical AI demands.

AI-assisted summary of Closing the Sim2Real Gap with Living Drone Imagery With Spexi — original source. Check the original for full context.

Source published: 2026-08-18

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 18, 2026, this presentation features Bill Lakeland, CEO of Spexi, discussing how a living, geometrically accurate aerial data layer can close the sim-to-real gap in Physical AI. The session highlights how a decentralized network of over 10,000 drone pilots captures ultra-high-resolution, continuously refreshed data to power spatial world models and assist partners like Niantic Spatial.

## TL;DR Bill Lakeland presents Spexi's decentralized aerial capture platform, which deploys off-the-shelf micro-drones across standardized H3 hexagonal 'specagons' to generate a living, geometrically accurate ground-truth layer for Physical AI. Moving beyond stale satellite feeds and synthetic-first simulations, the network executes automated, privacy-preserving flights verified via blockchain-based proof of capture. Spexi uses a predictive freshness and change-detection engine to trigger targeted reflights, optimizing capture unit economics across dynamic urban environments. Through city-scale Gaussian splatting in partnership with Niantic Spatial, Spexi provides the critical geometric anchors needed to align multimodal aerial and street-level datasets.

## Atmosphere & style The presentation maintains a pragmatic, authoritative, and forward-looking tone, driven by a founder with decades of domain experience in aerial surveying. Lakeland communicates with crisp technical clarity and commercial realism, comfortably shifting from geospatial indexing and unit economics to engaging banter about visually striking 3D Gaussian splats. The delivery makes an ambitious decentralized data infrastructure feel thoroughly operational, grounded, and accessible rather than speculative.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) The core bottleneck in Physical AI is shifting from model architecture to having access to a living, geometrically accurate, and continuously refreshed representation of the physical world. - [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) While synthetic simulation provides cheap and infinitely repeatable training data, real-world-first training is essential to ground world models in observable physical truth. - [03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) Aerial imagery must possess strict geometric fidelity and true-to-scale dimensions to serve as a digital fabric that contextualizes other spatial data streams. - [04:49](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=289s) Ground autonomous fleets miss rooftops and structural envelopes, satellites lack resolution, and crewed planes are economically prohibitive for continuous refreshes. - [06:00](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=360s) Spexi organizes distributed capture across Uber H3 resolution 9 spatial cells called 'specagons' (~25 acres), standardizing flight paths and embedding complete spatial telemetry. - [08:12](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=492s) Autonomous micro-drone missions take 7 to 10 minutes and maintain an altitude ceiling calibrated to ensure personal identifiable information (faces, license plates) is never captured. - [10:32](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=632s) Spexi ensures data provenance and eliminates synthetic spoofing by minting soulbound NFTs that immutably link raw imagery to drone sensor telemetry and GPS timestamps. - [10:32](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=632s) A network of over 10,000 decentralized pilots has mapped 7 million acres across nearly 250,000 missions, generating 40 million standardized images. - [11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) Distributed drones transform mapping unit economics by enabling targeted, high-cadence reflights of changing areas instead of repeatedly flying entire static cities. - [12:54](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=774s) A self-directed capture loop uses coarse change detection (satellites, permits, ground sensors) and freshness decay clocks to predict changes and trigger drone reflights. - [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) In partnership with Niantic Spatial, Spexi transforms standardized drone imagery into city-scale 3D Gaussian splats with sub-centimeter fidelity. - [15:31](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=931s) Dense aerial geometry serves as an anchor to align and unlock massive amounts of unstructured ground-level imagery, provided data streams are closely synchronized in time. - [16:44](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=1004s) The living drone imagery platform is actively deployed in production for physical AI world models, enterprise mapping, and property insurance analytics.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) Ground Truth and the Sim-to-Real Challenge

Closing the sim-to-real gap requires more than algorithmic tweaks; it demands a continuous, living baseline of the physical world. Bill Lakeland introduces himself as the CEO of Spexi, a Vancouver-based company focused on mapping the Earth from the air. Lakeland has spent his entire career in aerial surveying, having spent years flying airplanes to capture data that many consumers encounter daily on platforms like Google Maps or Apple Maps via municipal and governmental data-sharing agreements. However, traditional aerial and satellite mapping pipelines have left consumer maps stale. Lakeland points out that looking at one's own house on existing mapping platforms typically reveals outdated imagery—a baseline that cannot support the emerging requirements of Physical AI and spatial world models.

To fix this, Spexi transitioned away from satellites and airplanes toward drones, assembling what Lakeland describes as the largest standardized drone network in the world. While futurists have long envisioned swarms of drones buzzing continuously over urban centers, Lakeland emphasizes that this reality is actively taking shape right now through decentralized, programmatic capture. The goal is to build an automated, repeatable data layer that provides a verified ground-truth input for world models.

[01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) Synthetic First vs. Real-World Training

When teaching artificial intelligence to understand and navigate the physical world, two primary training methodologies have emerged across the industry:

```mermaid flowchart TD accTitle: Two Approaches to Training World Models subgraph SyntheticApproach["Synthetic-First Simulation"] A["Imagined / Procedural Data"] --> B["Fast, Low-Cost Generation"] B --> C["Infinitely Repeatable Sim-to-Real"] end subgraph RealWorldApproach["Real-World-First Foundation"] D["Observable Physical Truth"] --> E["Complex Empirical Capture"] E --> F["Geometrically Grounded Reference Layer"] end ```

The first paradigm is synthetic-first simulation, which represents the majority of existing projects. Companies utilizing synthetic generation benefit from rapid iteration, lower upfront costs, and infinite repeatability. However, Lakeland highlights a critical drawback: the foundational training source remains "imagined data" rather than verifiable, observed truth.

The second paradigm is real-world-first training, which relies directly on physical observation. While far more complex and time-consuming to construct, this approach produces data grounded in observable reality. Spexi aligns itself with this second camp, arguing that real-world AI training requires strict standardization and protocols to establish a consistent, dependable reference layer.

[03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) Geometric Precision and the Digital Fabric

Recent visual imagery is valuable for capturing the immediate state of the physical landscape, but it becomes vastly more powerful when it is geometrically accurate. Lakeland defines this usability through strict spatial criteria: positional correctness, true-to-scale dimensions, and measurable geometry. Drawing on Spexi’s heritage in precision mapping and photogrammetry, the platform applies these engineering principles to low-altitude drone imagery.

The result is an ultra-high-resolution "digital fabric of the built space" that acts as a base framework. This continuously refreshed aerial layer provides the coordinate anchors necessary to contextualize and calibrate other incoming data streams, giving disparate real-world inputs the geometric fidelity required for foundational world models.

[04:49](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=289s) The Sensory Gaps of Ground Fleets, Satellites, and Planes

A massive volume of real-world data is collected and refreshed every day, yet it remains fragmented and blind from the air:

- **Consumer & AR Apps:** Pervasive games like *Pokémon GO* and various augmented reality applications capture street-level slices of the world. - **Autonomous & Commercial Fleets:** Automotive camera networks, dashcams, and mobile LiDAR arrays on Uber and Waymo vehicles continuously scan transit corridors. - **Logistics Networks:** Delivery and transport vehicles record continuous telemetry and operational imagery.

Lakeland notes that virtually all of this ground data is proprietary, non-standardized for external model training, and hoarded for private competitive advantage. More critically, ground-level collection cannot capture rooftops, backyards, structural envelopes, or upper urban profiles, leaving massive holes throughout cityscapes.

Satellites and conventional surveying aircraft fail to bridge this divide. Lakeland deliberately excludes them from real-time dynamic pipelines because satellite imagery is too coarse in spatial resolution to guide physical AI interactions, while crewed aircraft surveys are far too infrequent and economically prohibitive to maintain a truly living model.

[06:00](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=360s) The Architectural Framework: Specagons and Autonomous Execution

Standardized drone imagery addresses these economic and physical limitations. Drones operate beneath the clouds, capture millimeter-to-centimeter detail, and exist in massive numbers globally. Spexi built an end-to-end software platform to coordinate this distributed hardware capacity, automating everything from flight planning to final imagery ingestion.

```mermaid flowchart LR accTitle: Spexi Data Capture and Ingestion Pipeline subgraph FlightPrep["Mission Setup"] A["H3 Res 9 Specagon (~25 Acres)"] --> B["Pilot Claims Cell & Hits Fly"] end subgraph AutonomousCapture["Autonomous Execution"] B --> C["Automated Flight Plan (7-10 Min Mission)"] C --> D["Terrain Tracking & Privacy Altitude"] end subgraph Processing["Verification & Delivery"] D --> E["Proof of Capture (Soulbound NFTs)"] E --> F["Mapping Products & API Pipelines"] end ```

To structure capture grids, Spexi partitions geography using Uber’s open-source H3 spatial index at resolution level 9. Each cell covers roughly 25 acres and is termed a "specagon." Within these spatial hexagons, the platform deploys standardized flight trajectories and capture rosettes tailored for specific machine learning and photogrammetric outputs.

Every captured frame embeds critical spatial metadata—precise GPS coordinates, 3D positioning (XYZ), timestamps, camera sensor telemetry, and orientation angles—ensuring straightforward integration into client training pipelines and external APIs.

[08:12](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=492s) Autonomous Micro-Drone Operations and Five Capture Modes

The operational workflow for field pilots is deliberately stripped of friction. Rather than manually steering the aircraft, the pilot simply travels to a claimed specagon and presses a single "fly" button within the Spexi mobile app. The operator's role is strictly supervisory: monitoring the airspace, tracking local conditions, and ensuring regulatory compliance.

The network relies entirely on decentralized consumer hobbyists and enthusiasts operating off-the-shelf micro-drones. Lakeland notes that micro-drones dramatically lower regulatory and safety barriers while still yielding aerial data superior to manned aircraft flown at higher elevations. Flights are swift, lasting only 7 to 10 minutes per cell.

Spexi embeds five distinct mission profiles into the specagons:

1. **Panoramas:** Elevated 360-degree street-view-style captures spanning entire municipal skylines. 2. **Photogrammetric Mapping:** Traditional nadir and overlapping grids optimized for orthomosaic generation. 3. **Hybrid Flights:** The network's most frequent mission profile, blending oblique panoramics with nadir mapping patterns. 4. **3D Reconstruction & Gaussian Splatting:** Hyper-dense viewing angles engineered specifically for novel view synthesis, neural radiance fields (NeRFs), and Gaussian splatting models. 5. **Specialized Structural Missions:** Targeted rosettes tuned for detailed geometry and infrastructure training data.

During every run, the automated flight control software maintains constant terrain-following altitude. This altitude ceiling is specifically calibrated to protect civilian privacy: drone cameras remain high enough that human faces, license plates, and house numbers cannot be resolved, eliminating personal identifiable information (PII) at collection time.

[10:32](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=632s) Blockchain-Based Proof of Capture and Network Scale

To guarantee that incoming imagery cannot be spoofed, manipulated, or synthetically fabricated, Spexi implemented a backend verification protocol called "proof of capture." As the micro-drone records data in the air, the platform mints soulbound NFTs in real time to immutably bind sensor telemetry, GPS timestamps, and device hardware signatures to the raw imagery. Lakeland stresses that for a foundation model, data provenance and authenticity are non-negotiable: "If we're creating ground truth layer, that's an absolute must." The system has maintained a zero-spoofing track record across its entire deployment history.

Once an upload passes automated validation, the pilot is paid immediately. Lakeland remarks that some dedicated operators earn between $500 and $1,000 in a single day using an off-the-shelf drone that costs roughly $600.

In just over a year of operation, the decentralized network has achieved significant operational milestones across North America:

- **Total Automated Missions:** Nearly 250,000 unique, pre-planned flights successfully completed. - **Pilot Community:** Over 10,000 active, decentralized drone pilots enrolled. - **Geographic Footprint:** 7 million acres of high-resolution, geometrically consistent coverage mapped. - **Curated Repository:** Nearly 40 million standardized images processed and optimized for physical AI and real-world foundation model training.

[11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) Beating Satellites and Aircraft: Unit Economics of the Living Layer

Satellite and fixed-wing aircraft imagery both face fundamental constraints when attempting to generate a living digital twin of the earth. While satellites suffer from resolution that is far too coarse for physical AI and detailed spatial reconstruction, aircraft can achieve decent resolution at a workable price point only if they fly over massive, contiguous geographic areas. That requirement creates an operational roadblock: flying entire regions by airplane makes continuous refreshes cost-prohibitive. Spexi’s distributed drone capture changes those underlying unit economics. Because the marginal cost of flying localized targets is realistic, the system can sustain a truly "living" layer across entire regions.

Rather than wasting resources repeatedly flying over an entire metropolis that has remained static, Spexi shifts its operational focus to targeted reflights. Once a city baseline is established, drones are dispatched at a much higher cadence specifically to the areas undergoing physical change. This selective, high-frequency approach forms the backbone of Spexi's strategy to keep spatial digital twins as fresh and continuously updated as possible.

[12:54](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=774s) Triggering Reflights and the Self-Directed Capture Loop

Because cities change unevenly across space and time, continuously reflying whole urban footprints makes little sense. Spexi tracks a "freshness clock" inside every individual Spexagon cell, comparing the time elapsed since the last flight against estimated physical change. To bootstrap this detection mechanism, Spexi pulls from multiple coarse signals: satellite imagery, ground-based sensor outputs, and municipal construction site permits. While these upstream data sources are effective at signaling that something has changed and flagging the general coordinates, they lack the spatial resolution to reveal *what* actually changed.

```mermaid flowchart TD accTitle: Spexi Self-Directed Capture Loop subgraph Detection["1. Coarse Change Detection"] S["Satellite Imagery"] P["Construction Permits"] G["Ground Sensors"] end

subgraph Dispatch["2. Targeted Aerial Capture"] FC["Spexagon Freshness Clock"] DR["Drone Reflight Dispatch"] end

subgraph Intelligence["3. Processing & Prediction"] IL["High-Res Identification Layer"] PM["Predictive AI Model"] end

S --> FC P --> FC G --> FC FC --> DR DR --> IL IL --> PM PM -.->|"Feedback Loop Refines Targets"| FC ```

Once a coarse trigger fires, Spexi dispatches localized drone flights to capture ultra-high-resolution imagery, revealing the exact nature of the change and generating a refreshed identification layer. Taking this architecture further, Spexi is developing its own predictive model to establish a fully "self-directed capture system." This model forecasts where physical alterations are bound to happen next, creating a closed-loop engine that continually learns, directs drone flights, absorbs new high-resolution feedback, and refines future predictions across repeatedly surveyed areas.

- **Trigger data feeds**: Satellite alerts, construction permits, and ground-level Internet of Things (IoT) sensors identify geographic anomalies without semantic detail. - **Spexagon freshness tracking**: Each hexagonal spatial unit monitors its own decay clock to balance cadence against cost. - **Autonomous learning loop**: Predictive models continuously direct pilot captures, ingesting newly gathered drone data to refine future prediction accuracy.

[14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) City-Scale Gaussian Splatting and the Niantic Spatial Partnership

Putting this pipeline into production, Spexi has worked for several years alongside Niantic Spatial to construct an end-to-end aerial pipeline transforming raw drone data into city-scale 3D Gaussian splats. Through this collaboration, municipal governments, enterprise infrastructure teams, and construction managers can directly task local drone networks to ingest imagery and output Gaussian splats straight into their daily engineering and visualization workflows.

Showcasing both older baseline scans and recent high-fidelity captures across San Francisco, the results demonstrate broad metropolitan coverage combined with sub-centimeter surface detail. Bill Lakeland jokes with the audience while showcasing the 3D renders on screen: "I'm not sure you guys heard anything I said on this slide. You're probably looking at this awesome splat made by Niantic."

[15:31](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=931s) Multimodal Ground-Air Alignment and Real-World Models

A central insight gained from these city-scale captures is that effective world models cannot rely on a single vantage point—they must be multimodal. Aerial perspectives and street-level ground captures cannot live in silos; "they need to connect, they need to drive together." However, bridging the perspective gap between top-down drone shots and ground-level street photography remains an inherently difficult computer vision challenge.

Spexi solves this alignment problem by serving as the underlying geometric anchor. Because Spexi data is processed into a dense, geometrically consistent framework covering the entire urban fabric, it provides the structural features required to register and bind disparate visual modalities together. This enables teams to unlock immense value from vast troves of unstructured, incidental ground data. Crucially, alignment depends on temporal synchronization: the closer datasets are in capture time, the tighter the spatial bond. Recency is what prevents dynamic urban elements from breaking multi-sensor feature matching.

- **Geometric binding layer**: Spexi’s high-density aerial photogrammetry provides ground-truth anchors that ground-level sensor arrays can lock onto. - **Unlocking incidental data**: Vast catalogs of unaligned ground imagery and drive-by captures become usable once tethered to consistent aerial geometry. - **The recency requirement**: Sensor alignment degrades if aerial baselines and ground feeds drift apart in time, reinforcing the need for continuous refresh cycles.

[16:44](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=1004s) Production Reality, Industry Verticals, and Closing the Sim-to-Real Gap

Far from an abstract concept, the platform is fully operational and actively providing foundational training data to world model architectures today. Spexi is aggressively expanding its footprint across several key commercial sectors: - **Physical AI & World Models**: Ground-truth aerial data pipelines feeding simulation environments and spatial AI. - **Insurance & Enterprise Mapping**: High-frequency, high-resolution validation of real-world property risks and spatial assets. - **Distributed Pilot Network**: Continued expansion of the 10,000+ pilot community into new cities to execute daily localized reflights.

Bill Lakeland concludes by reiterating that bridging the sim-to-real gap fundamentally relies on accessible, authentic, and living aerial foundations. Because the physical world is constantly shifting, autonomous systems and physical AI architectures require continuous awareness of where and how that change unfolds. Lakeland invites attendees to explore their work, visit their exhibition booth, or access sample datasets directly via spexi.com.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) One of the next major bottlenecks in Physical AI isn’t model architecture, it’s access to a living and geometrically accurate representation of the real world. — Bill Lakeland

> [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) imagined data — Bill Lakeland

> [03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) digital fabric of the built space — Bill Lakeland

> [06:00](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=360s) specagon — Bill Lakeland

> [08:12](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=492s) fly — Bill Lakeland

> [10:32](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=632s) proof of capture — Bill Lakeland

> [10:32](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=632s) If we're creating ground truth layer, that's an absolute must. — Bill Lakeland

> [12:54](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=774s) freshness clock — Bill Lakeland

> [12:54](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=774s) self-directed capture system — Bill Lakeland

> [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) I'm not sure you guys heard anything I said on this slide. You're probably looking at this awesome splat made by Niantic. — Bill Lakeland

> [15:31](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=931s) they need to connect, they need to drive together. — Bill Lakeland

## Mentioned Bill Lakeland, Spexi, Ai4, Niantic Spatial, Google Maps, Apple Maps, Pokémon GO, Uber, Waymo, Uber H3, San Francisco

## Video description Bill Lakeland - CEO, Spexi

One of the next major bottlenecks in Physical AI isn’t model architecture, it’s access to a living and geometrically accurate representation of the real world.

This session explores why closing the sim-to-real gap requires foundational aerial data with three essential characteristics: ultra-high resolution, continuous refresh, and consistent quality at scale.

Learn why these attributes are becoming critical for training world models and how a distributed network of more than 10,000 drone pilots can deliver them where satellites and traditional aerial imagery cannot, and see how this approach is already helping train Niantic Spatial’s foundation models.

[Source](https://www.youtube.com/watch?v=EO0P8a0GZwE)

中文简报

Spexi动员万名无人机飞手测绘地球，服务物理AI模型弥合从模拟到现实的差距需要持续、去中心化的无人机采集，因为静态卫星图像与合成模拟无法提供物理AI所需的几何地面真值。 Bill Lakeland介绍了Spexi的去中心化航空采集平台。该平台在标准化的H3六边形“specagon”区域中部署现成的微型无人机，旨在为物理AI生成动态且具有精确几何形态的地面真值层。该网络摆脱了陈旧的卫星数据源和以合成模拟为主的模式，执行具有隐私保护的自动化飞行，并通过基于区块链的采集证明进行验证。Spexi利用预测性数据新鲜度与变化检测引擎来触发针对性的重新飞行，从而优化动态城市环境中采集的单位经济效益。通过与Niantic Spatial合作开展城市级高斯泼溅（Gaussian splatting），Spexi为对齐多模态航拍与街景数据集提供了关键的几何锚点。物理AI的核心瓶颈正从模型架构转向能否获取动态、几何精确且持续更新的物理世界表征。 (0:04) 尽管合成模拟能够提供廉价且可无限重复的训练数据，但以现实世界为先的训练对于将世界模型建立在可观测的物理真实之上至关重要。 (1:51) 航拍图像必须具备严格的几何保真度和真实比例尺寸，才能作为连接并赋予其他空间数据流上下文背景的数字化基底。 (3:16) 地面自动驾驶车队会遗漏屋顶和建筑外围护结构，卫星分辨率不足，而有人驾驶飞机若用于持续更新则在经济成本上难以承受。 (4:49) Spexi在被称为“specagon”（约25英亩）的Uber H3第9级分辨率空间网格中组织分布式采集，实现飞行路径标准化并嵌入完整的空间遥测数据。 (6:00)

Reading summary

Spexi CEO Bill Lakeland argues that physical AI and spatial world models require authentic, geometrically precise real-world data rather than purely synthetic simulations, which existing satellite, airplane, and ground fleets fail to adequately provide. Spexi addresses this gap by coordinating a decentralized network of consumer micro-drones to construct a continuously refreshed, standardized aerial data layer. Limitations of Existing Terrestrial and Orbital Mapping (4:49) While synthetic-first simulations offer rapid iteration and low upfront costs, Lakeland asserts that world models trained purely on procedural generation lack empirical physical truth. Conversely, existing physical observation networks possess severe structural gaps. Ground fleets—such as autonomous vehicles, delivery vans, and smartphone AR apps—capture continuous transit corridors, but the data is largely proprietary and incapable of viewing rooftops, structural envelopes, or upper urban profiles. Conventional aerial methods also fail to provide a living foundation. Lakeland notes that satellite imagery is too coarse in spatial resolution to guide physical AI interactions, while crewed aircraft surveys are far too costly and infrequent to maintain currency. In response, Spexi captures low-altitude imagery with millimeter-to-centimeter detail to build an ultra-high-resolution digital fabric of built spaces, providing the positional accuracy and scale needed to contextualize foundational spatial models. Decentralized Specagon Capture and Provenance Verification (6:00) Spexi partitions geography using Uber’s open-source H3 spatial index at resolution 9, dividing land into roughly 25-acre cells called "specagons." Independent hobbyists operating off-the-shelf micro-drones claim these cells via a mobile app and initiate automated 7-to-10-minute flights. The system supports five distinct capture profiles—panoramas, photogrammetry, hybrid flights, structural rosettes, and trajectories optimized for novel view synthesis techniques like NeRFs and 3D Gaussian splatting. Flight control software enforces a constant terrain-following altitude to prevent the resolution of faces or license plates, filtering out personal identifiable information at capture time. To ensure data integrity for foundation models, Spexi implements a "proof of capture" protocol that mints soulbound NFTs in real time, immutably linking sensor telemetry, GPS timestamps, and drone hardware signatures to raw imagery. Lakeland reports a zero-spoofing record across the platform's history. Operating for just over a year, the decentralized network has enrolled over 10,000 active pilots, completed nearly 250,000 automated missions, and assembled nearly 40 million images covering 7 million acres across North America. Selective Reflight Economics and Self-Directed Loops (12:54) Traditional aerial surveys must cover large contiguous areas to remain economically viable, making continuous urban refreshes cost-prohibitive. Spexi shifts these unit economics by dispatching micro-drones only to locations experiencing physical alterations. Each specagon cell tracks an individual "freshness clock" that measures time elapsed since its last flight against expected physical change. To trigger localized reflights, Spexi ingests coarse external indicators, including municipal construction permits, coarse satellite imagery, and ground IoT sensors. Because these sources signal the coordinates of change without providing semantic resolution, Spexi sends drones to capture high-resolution identification layers. Furthermore, the company is developing a predictive model to forecast where physical shifts will occur next, creating a closed-loop system where newly collected drone data continually refines future targeting algorithms. Multimodal Anchoring and City-Scale Gaussian Splatting (15:31) Through a multi-year partnership with Niantic Spatial, Spexi translates raw aerial imagery into city-scale 3D Gaussian splats, producing models across metropolitan areas like San Francisco that merge regional coverage with sub-centimeter surface detail for municipal and enterprise infrastructure workflows. Beyond standalone 3D rendering, Lakeland emphasizes that robust world models require multimodal integration between aerial imagery and street-level views. Disparate ground and aerial feeds are historically difficult to match, but Spexi's geometrically consistent aerial photogrammetry provides a structural base framework that grounds and aligns unstructured ground-level imagery. Lakeland highlights that this cross-perspective alignment is time-sensitive: temporal drift between aerial baselines and ground feeds degrades multi-sensor calibration, underscoring the necessity of high-cadence refresh cycles. Lakeland emphasizes that closing the sim-to-real gap depends on treating the physical world as a living, continuously changing baseline rather than a static map. While Spexi is deploying this aerial data across physical AI, insurance, and municipal mapping, its long-term impact relies on maintaining active pilot density across diverse geographies and proving that automated predictive reflight loops can accurately anticipate real-world urban development.

精炼阅读摘要

Spexi 首席执行官 Bill Lakeland 指出，缩小物理 AI 与空间世界模型的“仿真到真实”（Sim2Real）差距，关键在于建立持续更新且具备严格几何精度的真实物理世界基准层。该公司通过构建去中心化的消费级微型无人机网络，替代传统卫星与载人航测，提供低成本、高频次、可验证的动态空域数据底图。真实世界数据对齐与传统航测感知断层 (1:51) 在训练物理世界 AI 模型时，行业主要存在以合成模拟为主和以真实世界为基础两种路径。Lakeland 认为，合成数据虽然迭代快、成本低且可无限复现，但其本质是“想象数据”；而以客观物理观测为核心的路线虽构建复杂，却能提供具有可度量几何精度与真实空间尺度的可靠基准。当前的物理空间数据捕获存在明显断层：自动驾驶车队、物流网络及移动 AR 应用采集了庞大的地面数据，但此类数据大多专有封闭，且受地面视角限制无法捕捉建筑屋顶、后院与立面结构；与此同时，卫星图像的空间分辨率过于粗糙，传统载人飞机测绘则成本高昂、更新周期过长，导致消费级与工业级地图长期处于陈旧状态。 Specagon 网格自主采集与区块链确权机制 (6:00) 为实现标准化低空采集，Spexi 基于 Uber 开源的 H3 空间索引（第 9 级分辨率），将地理区域切分为覆盖约 25 英亩的六边形单元（称为 Specagon）。网络依赖持有约 600 美元常见微型无人机的去中心化业余飞手。飞手抵达网格后仅需在 App 中一键触发，无人机即自动执行 7 至 10 分钟的地形跟随预设航线，飞手仅需负责空域安全监管。航线预设全景、正射摄影测量、混合航拍、高斯溅射（Gaussian Splatting）三维重建及特殊结构扫描五种任务，巡航高度严格限制在无法解析人脸与车牌的范围以消除隐私隐患。针对基础模型对数据真实性的严苛要求，系统在飞行中实时铸造灵魂绑定 NFT（Soulbound NFTs），将硬件签名、传感器遥测及 GPS 时间戳不可篡改地绑定至原始图像，实现了零欺诈记录。该网络上线一年多已吸纳超 10,000 名飞手，完成近 25 万次自动化任务，测绘面积达 700 万英亩，沉淀了近 4,000 万张标准化图像。动态刷新经济学与自导向闭环调度 (11:55) 与载人飞机必须大面积连续飞行以均摊成本不同，微型无人机网络的局部复飞边际成本极低，使高频维持“活体数据层”成为可能。由于城市物理环境的变化在空间上高度不均衡，连续全量重飞缺乏经济合理性。Spexi 在每个 Specagon 单元内设置“新鲜度时钟”，结合卫星影像异常、市政施工许可证和地面物联网传感器等粗粒度信号，识别可能发生变化的区域，进而靶向调度无人机复飞以获取超高分辨率图像。进一步地，Spexi 正在开发预测模型构建“自导向捕获系统”，用于预判物理环境即将发生变动的热点区域，使模型在下发任务、吸收新影像与优化未来预测之间形成自主迭代闭环。城市级高斯溅射与空地多模态协同 (14:27) 在落地应用层面，Spexi 与 Niantic Spatial 合作建立了将无人机原始影像直接转化为城市级 3D 高斯溅射（Gaussian Splats）的端到端管线，已在旧金山等城市展示了兼具大范围覆盖与亚厘米级表面精度的渲染成果，供市政、基建和工程团队调取。更深层的价值在于空地多模态对齐：地面车辆拍摄与高空俯拍视角差异巨大，在计算机视觉中极难直接匹配；Spexi 生成的连续密集几何骨架可充当底层几何锚点，将杂乱碎片化的地面随车影像和街景数据精准配准。Lakeland 强调，多模态对齐具有严苛的时间敏感性，基准底图的采集时间若与地面数据脱节，城市动态要素的变化将导致特征匹配失效，因此高频的活体更新是多源传感器融合的关键前提。 Lakeland 指出，Spexi 目前正将这一活体空域数据层推向物理 AI、世界模型构建以及保险与资产风控等商业领域。然而，该模式的长期有效性仍取决于其在不同空域监管政策下飞手网络的扩张效率、动态预测模型的调度准确度，以及面对光照、季节等环境扰动时保持高精度几何一致性的技术鲁棒性。

Detailed notes — zh

## Background 本内容由 Ai4 于 2026 年 8 月 18 日发布在 YouTube 平台。页面简介显示，Spexi 首席执行官 Bill Lakeland 指出当前物理 AI 的核心瓶颈在于缺乏真实世界的活态几何表征，并介绍了其如何通过超 10,000 名分布式无人机飞手网络提供高分辨率、高频刷新的底层航空数据，以助力训练 Niantic Spatial 基础模型并弥合 Sim2Real 差距。 ## TL;DR Spexi 首席执行官 Bill Lakeland 深入阐述了如何利用去中心化无人机网络打破物理 AI 与世界模型在数据层面的瓶颈。针对传统卫星分辨率粗糙、有人机航测成本过高且难以高频复飞的问题，Spexi 通过 Uber H3 六边形网格与自动化飞行控制，让微型无人机高效采集具备严密几何精度的高清航测数据。同时，平台结合区块链 Proof of Capture 防伪鉴权与主动变化预测模型实现动态按需复飞，并携手 Niantic Spatial 打造城市级 3D Gaussian Splatting 处理管线，为空地多模态融合与前沿世界模型训练提供了高保真、活态的物理世界底座。 ## Atmosphere & style 演讲整体呈现出务实、严谨且极具行业洞察力的技术布道风格。发言者凭借深厚的航空测绘实战背景，从容而笃定地剖析行业底层痛点，语言精炼且逻辑层次分明。现场展示城市级 3D 高斯泼溅渲染效果时，讲者适时穿插轻松幽默的互动，既展现出对技术突破的强烈自信，又为前沿空间计算与物理 AI 的落地勾勒出一幅充满工业级实操质感的生动画卷。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) 弥合物理 AI 与具身智能的 Sim2Real 差距，核心瓶颈在于缺乏具备超高分辨率、实时新鲜度与几何精度的“活态（living）”航空基准数据。 - [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) 物理世界 AI 训练存在两大流派：合成优先（Synthetic-first）方法成本低但缺乏真实物理观测；真实世界优先（Real-world first）方案虽采集难度高，但具备完全可验证的客观真值。 - [03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) 单纯依赖卫星遥感（分辨率粗糙）或载人航测（成本高、周期长）无法构建活态模型，且地面数据存在生态封闭与低空俯瞰盲区，亟需去中心化微型无人机网络提供厘米级超高清数字织物。 - [06:38](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=398s) Spexi 基于 Uber H3 算法将空域拆分为约 25 英亩的六边形单元（Spexagons），飞手自备低门槛微型无人机即可一键启动全自动化作业，并覆盖测绘、全景与高斯泼溅等五大飞行模式。 - [10:13](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=613s) 飞行全程采用地形跟随技术维持恒定离地高度，既保证了高质量成像，又在物理层面规避了识别面部或车牌等隐私信息。 - [10:13](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=613s) 引入基于区块链的“捕获证明（Proof of Capture）”和灵魂绑定 NFT（Soulbound NFT），实现对空中影像防篡改鉴权并支持即时结算报酬。 - [10:13](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=613s) 网络规模在一年多内爆发式增长，汇聚超 10,000 名飞手，在北美执行近 25 万次自动化任务，覆盖逾 700 万英亩区域并沉淀近 4,000 万张高精度影像底料。 - [11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) 设立网格“鲜活度时钟”，结合卫星影像、施工许可及地面传感器等粗粒度信号动态评估变化，仅对发生物理变动的区域进行定向无人机高频复飞，极大优化作业经济性。 - [13:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=831s) 研发专有变化预测模型构建“自主导向数据采集系统”，使平台从被动应对告警转向主动预测城市空间物理演变，形成自我进化的采集与训练闭环。 - [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) 携手 Niantic Spatial 推出城市级 3D Gaussian Splatting 处理管线，并以高几何精度航空数据作为粘合剂，精准对齐零散的地面多模态传感器数据。 - [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) 跨视角多模态数据对齐对“时效性（Recency）”要求极高，时间维度越接近，数据缝合精度越高，物理世界模型结构也越稳固。 - [16:44](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=1004s) Spexi 数据底座已投入工业化落地，正直接支持前沿世界模型团队训练，并在保险科技、企业级空间测绘及新城市网络扩展中加速推进。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) 弥合 Sim2Real 差距：活态无人机影像与世界模型基座在物理人工智能（Physical AI）与具身智能蓬勃发展的当下，如何让 AI 真正理解并融入现实物理世界成了一个核心瓶颈。Spexi 首席执行官比尔·莱克兰（Bill Lakeland）开门见山地指出，解决“仿真到真实世界（Sim2Real）”差距的关键，在于为世界模型（World Models）提供具备超高分辨率、实时新鲜度与几何精度的“活态（living）”航空基座数据。总部位于温哥华的 Spexi 长期专注于从空中测绘地球，莱克兰本人职业生涯的大部分时间都在驾驶飞机进行航空测绘，许多人在 Google Maps 或 Apple Maps 上看到的市政公开地理数据，很多便出自他早年的测绘成果。伴随技术演进，他毅然将重心从人造卫星与传统航测飞机转向无人机，并搭建起全球最大的标准化去中心化无人机网络。当前主流地图平台上的影像往往严重滞后，普通人打开地图查看自己的住所时，看到的通常是数月甚至数年前的旧景。莱克兰坚信，随着物理 AI 与世界模型的兴起，未来的地图与空间模型必须成为物理世界高保真、高频更新的真实镜像（Ground Truth）。行业过去常畅想无人机漫天穿梭于城市上空的未来场景，而如今 Spexi 正在通过一种标准化的网络架构，让这种大规模自主空域数据捕获在现实中真正落地运转。 ### [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) 物理世界 AI 训练的双重路径：合成仿真 vs. 真实观测在教授 AI 理解物理世界的技术路线上，业内目前主要分化为两大流派。第一种是当前被广泛采用的“合成优先（Synthetic-first）”模型。这类方法通过计算机模拟来构建虚拟世界，其核心优势在于生成速度极快、边际成本极低，且具备近乎无限的可重复性。然而莱克兰尖锐地指出，这类方案致命的短板在于其底层数据源本质上是“想象中的数据（imagined data）”，缺乏物理世界中复杂而不可预测的真实观测真值。另一种方法则是以真实环境为基底的“真实世界优先（Real-world first）”训练。这一路径直接取材于物理实体世界，尽管采集与处理复杂度极高、构建周期更长，但它具备不可替代的优势：完全可验证，且牢牢扎根于可观测的现实真理之上。Spexi 选择全面押注这一方向，并与业内顶尖的先锋机构展开紧密合作。莱克兰强调，想要在真实世界基底上稳定训练模型，必须建立起一套严密的标准化协议与统一参照系，而 Spexi 的核心目标正是打造这套不可或缺的底层数据基座。 ```mermaid flowchart TD accTitle: 物理世界 AI 训练的两大路径对比 subgraph Synthetic["合成优先路径 (Synthetic-First)"] S1["计算机仿真生成"] --> S2["优势：极速、低成本、无限重复"] S2 --> S3["痛点：属于'想象数据'，缺少物理真实性"] end subgraph RealWorld["真实世界优先路径 (Real-World First)"] R1["物理世界直接观测采集"] --> R2["优势：绝对真实、可验证、客观真值"] R2 --> R3["挑战：复杂度高、耗时长、需统一标准化"] R3 --> R4["Spexi 解决方案：标准化无人机采集网"] end ``` ### [03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) 几何精度与空中盲区：构建人造环境的超高清数字织物真实世界影像的价值并不仅体现在视觉观感的新鲜度上，更取决于其几何精度（Geometric Accuracy）。这意味着数据必须精确匹配地理坐标，且具备严密的可量化与可测量属性。Spexi 将多年积累的高精度测绘工程经验沉淀为底层原则，构建出一套覆盖人造建筑环境的高精度基准数据集。该数据集以持续更新的图像图层为核心，宛如一张覆盖物理空间的超高分辨率“数字织物（digital fabric）”。它不仅能作为世界模型的底层底图，还能将来自地面车辆、传感器等其他模态的数据进行对齐与几何校正，极大增强多模态数据在物理模型训练中的实用价值。现实中并不缺少地面维度的鲜活数据源——例如以《Pokémon Go》为代表的 AR 移动应用、Uber 和 Waymo 自动驾驶车辆上配备的行车记录仪及激光雷达（LiDAR），或是各类物流车队在移动中不断积累的视觉片段。然而，这些地面数据几乎全部被各家公司封锁在封闭生态内作为竞争壁垒，缺乏对外输出的统一格式；更关键的是，它们完全缺失俯瞰视角的低空维度，导致整座城市的数字表达存在大量死角与缝隙。而传统卫星与有人航测飞机同样无法胜任： - 卫星遥感分辨率过于粗糙，无法满足自动驾驶和微观世界模型的几何需求； - 载人飞机测绘成本极度高昂、调度周期冗长，根本无法维持“活态模型”所需的高频刷新节奏； - 唯有去中心化的标准化微型无人机网络，兼具厘米级超高清分辨率、坚固的几何刚性与极高的重访刷新率，在经济性与全球覆盖能力上展现出绝对优势。 ### [06:38](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=398s) 自动化端到端采集体系：Uber H3 六边形网格与五大飞行模式为了将散落全球的无人机硬件转化为标准化的测绘力量，Spexi 自主研发了一套端到端的全自动采集与产品输出平台。系统将复杂的空域任务解构为统一的离散地理单元——基于 Uber H3 算法划分的 Level 9 六边形空间单元，面积约 25 英亩，被称为“Spexagons”。飞手在 App 端认领一个 Spexagon 后，其工作被简化到了极致：只需在起飞点点击“开始飞行（Fly）”按钮，随后在现场承担空域安全监控职责，确保符合航空监管规范。 ```mermaid flowchart LR accTitle: Spexi 端到端自动化数据采集与处理流程 subgraph Pilot["飞手操作端"] A["认领 Spexagon 六边形单元"] --> B["一键启动自动化飞行"] B --> C["现场空域监控与安全规避"] end subgraph Autonomous["全自主飞行控制"] C --> D["软件接管：微型无人机全自动作业"] D --> E["五大预设捕获模式: 测绘/全景/高斯泼溅等"] end subgraph DataPipeline["数据处理与交付"] E --> F["嵌入全量元数据: GPS/时间戳/姿态角"] F --> G["API 对接外部工作流 / 基础底图产品"] end ``` 网络中的硬件设备完全由全球各地的无人机爱好者和兼职飞手自备，且系统限定仅支持微型无人机（Micro Drones）。这类设备不仅大幅降低了法规门槛与作业风险，其在低空捕获的画质表现反而显著超越了高空飞行的航测飞机。飞手点击起飞后，软件即刻全面接管，在 7 至 10 分钟内全自动执行精密设计的飞行航线。所有影像自动嵌入精准的 GPS 坐标、时间戳、空间 XYZ 位置、航向角、姿态及传感器参数，可通过 API 无缝接入客户的物理 AI 模型训练流程。针对不同任务场景，平台预设了五种专属航线模式： - **全景空巡模式（Panoramas）**：类似于在城市上空铺开的“空中版 Google 街景”； - **航测正射模式（Mapping Flights）**：为传统摄影测量工作流提供严谨的正射基准； - **混合倾斜模式（Hybrid Flights）**：结合倾斜全景与测绘路径，也是目前应用最广泛的综合模式； - **高斯泼溅专用模式（Gaussian Splatting Missions）**：针对超高保真 3D 重建与辐射场算法进行视角与重叠度专项优化的飞行模式。 ### [10:13](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=613s) 地形跟随、Proof of Capture 区块链认证与百万级规模突破为了在全域范围内保持成像的一致性并兼顾社会伦理，无人机在飞行全程采用实时“地形跟随（Terrain Tracking）”技术，始终维持相对地面的恒定高度。莱克兰特别说明，这种飞行高度既保证了测绘模型所需的超高画质，又从物理层面天然规避了隐私侵犯——在当前分辨率下，系统无法识别行人面部、车牌号或门牌号码，不采集任何可识别的个人身份信息。在数据确权与防伪层面，Spexi 构建了一套基于区块链的“捕获证明（Proof of Capture）”后端机制。当无人机在空中快门触发时，系统会实时生成灵魂绑定 NFT（Soulbound NFT）对影像进行防篡改鉴权与来源追溯，彻底杜绝了飞手伪造 GPS 或串改数据的可能。一旦影像通过验证，飞手将立即获得即时结算报酬。莱克兰透露，许多飞手仅凭一台价值 600 美元的入门级微型无人机，通过全职或密集执行任务，单日收益甚至能达到 500 至 1,000 美元。在短短一年多的运营时间内，Spexi 去中心化网络的规模实现了爆发式扩张，展现出传统航测无法企及的落地效能： - **自动化任务量**：在北美地区完成了近 25 万次独立的端到端自动化飞行任务； - **飞手网络生态**：吸引并激活了超过 10,000 名持有合规硬件的注册无人机飞手； - **地理覆盖面积**：累计完成超过 700 万英亩高精度、标准化人造空间环境的数据覆盖； - **基座数据体量**：数据中心已沉淀近 4,000 万张专为真实世界模型训练清洗、调优的超高清影像底料。 ### [11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) 相比卫星与传统航空摄影的经济性优势与动态变化检测机制与卫星图像和传统载人飞机摄影相比，Spexi 的核心差异化优势集中在两点：分辨率与成本结构。卫星遥感的分辨率较为粗糙，很难支持精细几何建模与物理 AI 的高精度需求；传统飞机虽然在大面积航拍时具备一定的成本效益，但其作业模式必须一次性覆盖极其庞大的区域，这种沉重的机制导致其无法频繁复飞，根本无法维持数据的“鲜活性”（Living data）。而 Spexi 凭借去中心化无人机网络的单位经济效益，让高频复飞在商业成本上切实可行。 ```mermaid flowchart TD accTitle: 动态变化检测与高精度无人机复飞工作流 subgraph TriggerSources ["多源粗粒度变化线索"] A["低分辨率卫星图像"] B["建筑施工许可数据"] C["地面传感器异常信号"] end subgraph DetectionEngine ["变化检测与决策引擎"] D["监测 Spectagon 鲜活度时钟"] E["评估区域实际变化概率"] D --> E end subgraph SpexiAction ["精确复飞与数据更新"] F["触发定向无人机复飞"] G["生成超高精度识别图层与 3D 重建"] end TriggerSources --> DetectionEngine E --> F F --> G ``` 在实际运营中，Spexi 不再需要毫无目的地对整座城市进行重复全量扫描。从高空俯瞰，城市并非在同一时间发生全盘改变，因此盲目全覆盖飞行不仅浪费资源，也毫无必要。Spexi 为每个被称为“Spectagon”的六边形空域网格设立了“鲜活度时钟”（Freshness clock），通过将时间流逝与区域内的实际变化幅度进行比对，动态决定复飞优先级。系统会结合卫星图像、建筑施工许可证申报以及地面传感器读数等外部信号作为前置触发机制。这些外部数据源虽然能够灵敏地指出“某地发生了变化”，但受限于精度无法解析“具体改变了什么”。Spexi 借此触发针对性的无人机复飞，通过超高分辨率影像精确捕捉地表地貌的真实演变，进而生成全新的高精识别图层。 - **卫星遥感痛点**：空间分辨率粗糙，无法满足物理 AI 与精细三维建模对几何精度的严苛标准。 - **载人飞机航拍痛点**：单次作业面积过大、调动成本高昂，无法支持高频次局部刷新，数据极易过时。 - **Spectagon 鲜活度机制**：以六边形空间网格为单元，结合外部变化线索动态跟踪时效衰减，精准调度高频复飞。 --- ### [13:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=831s) 自主导向的数据采集与变化预测模型为了进一步提升空间数据更新的自动化水平，Spexi 正在自主研发一套专有的预测模型，旨在构建一个“自主导向的数据采集系统”（Self-directed capture system）。该系统不再完全依赖事后产生的施工许可或低分辨率卫星告警，而是能够主动预测城市空间中最可能在何时、何地发生物理改变。这种架构构成了自我进化的闭环：系统持续学习历史数据、向采集端反馈策略、精细化调整预测精度，并自主向无人机飞手网络下达重定向采集任务。随着无人机在各个 Spectagon 网格内完成一次又一次的复飞，这种“采集-反馈-训练”的闭环不断提高预测算法的准确性。在 Spexi 已经建立起高密度、持续性飞行覆盖的核心区域，这种预测驱动的自主采集模式已经初具雏形并展现出显著的运行效率。 --- ### [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) 携手 Niantic Spatial：城市级 3D Gaussian Splatting 与多模态数据对齐在技术落地层面，Spexi 展示了与 Niantic Spatial 团队深度合作两年的成果。双方联合打造了一条从无人机原始航拍数据直接生成城市级 3D 高斯溅射（3D Gaussian Splatting）的完整处理管线（Pipeline）。目前该产品线已正式推向市场，市政管理部门、工程建设施工方等终端客户能够直接下发无人机采集任务，快速生成逼真的 3D Gaussian Splat 模型，并无缝注入其现有的日常数字化业务工作流中。现场展示了早期概念验证案例与近期在旧金山采集的高精度成果对比。针对物理世界模型的构建，多模态融合存在巨大技术难点：空域航拍数据与地面端传感器数据之间往往存在严重的割裂，难以精准对齐与缝合。而 Spexi 提供的航拍影像具备几何精度的严谨性与空间覆盖的连续性，能够天然充当将地面数据与空中数据紧密咬合的“粘合剂”（Bond）。这种高密度的视觉锚点使得海量零散、偶发的地面传感器数据（Incidental data）得以精准挂载，成倍释放其潜在价值。在此过程中，“时效性/近时性”（Recency）至关重要——只有当多模态数据集在时间维度上高度接近时，跨视角的数据缝合才足够牢固，物理世界模型的结构才能更加严密精确。在展示由 Niantic 生成的极其惊艳的 3D Gaussian Splat 渲染画面时，演讲者打趣道：“我甚至怀疑大家根本没听我这页 PPT 讲了什么，你们的注意力肯定全被 Niantic 制作的这个震撼的 Splat 画面吸走了。” - **产品化落地**：支持城市治理与建筑工地按需调用无人机生成 3D Gaussian Splats，实现空间资产的高保真三维重建。 - **空地多模态对齐**：凭借厘米级几何精度与高密度特征点，将散落的地面传感器数据与高空全局图层紧密绑定。 - **时效性协同**：多模态数据采集在时间跨度上越接近，跨模态数据对齐的精度就越高，模型误差越小。 --- ### [16:44](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=1004s) 现实世界模型落地、垂直领域拓展与缩小 Sim2Real 差距这一整套空间数据基础设施并非停留在纸面理论阶段，Spexi 的平台已全面投入工业化运行，各大前沿世界模型（World Models）研发团队目前正在直接使用该数据开展基础模型的实际训练。展望未来发展路径，Spexi 正在加速拓展合作伙伴生态，并在多个核心垂直业务领域深耕拓展，包括： 1. **保险科技与企业级测绘**：提供经得起法律与风险审计的高精地面真实情况（Ground truth）； 2. **Physical AI 与世界模型团队**：为其提供持续刷新、具备严格几何一致性的三维物理世界基底； 3. **飞手网络与城市覆盖扩张**：持续吸纳新飞手进入超过 10,000 名认证飞手的分布式网络，将业务推向更多新城市，同时保持既有覆盖区域的每日高频刷新。演讲最后总结强调：真正弥合“从模拟到真实”（Sim-to-Real）鸿沟的核心起点，在于具备**超高分辨率、高频持续更新、几何精准且高度真实**的底层基础航空数据。物理世界每时每刻都在经历动态重塑，物理 AI 和世界模型必须深刻理解这些变化发生于何处、演变成了何物。Spexi 邀请参会者访问 spexi.com 或前往展台体验实际数据集进行测试接入。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) 物理人工智能与具身智能蓬勃发展的当下，如何让 AI 真正理解并融入现实物理世界成了一个核心瓶颈。 — 比尔·莱克兰 > [00:04](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=4s) 未来的地图与空间模型必须成为物理世界高保真、高频更新的真实镜像。 — 比尔·莱克兰 > [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) 这类方法通过计算机模拟来构建虚拟世界，其核心优势在于生成速度极快、边际成本极低，且具备近乎无限的可重复性。然而……这类方案致命的短板在于其底层数据源本质上是‘想象中的数据’，缺乏物理世界中复杂而不可预测的真实观测真值。 — 比尔·莱克兰 > [01:51](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=111s) 直接取材于物理实体世界，尽管采集与处理复杂度极高、构建周期更长，但它具备不可替代的优势：完全可验证，且牢牢扎根于可观测的现实真理之上。 — 比尔·莱克兰 > [03:16](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=196s) 真实世界影像的价值并不仅体现在视觉观感的新鲜度上，更取决于其几何精度。这意味着数据必须精确匹配地理坐标，且具备严密的可量化与可测量属性。 — 比尔·莱克兰 > [06:38](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=398s) 飞手在 App 端认领一个 Spexagon 后，其工作被简化到了极致：只需在起飞点点击‘开始飞行’按钮，随后在现场承担空域安全监控职责，确保符合航空监管规范。 — 比尔·莱克兰 > [10:13](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=613s) 许多飞手仅凭一台价值 600 美元的入门级微型无人机，通过全职或密集执行任务，单日收益甚至能达到 500 至 1,000 美元。 — 比尔·莱克兰 > [11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) 从高空俯瞰，城市并非在同一时间发生全盘改变，因此盲目全覆盖飞行不仅浪费资源，也毫无必要。 — 比尔·莱克兰 > [11:55](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=715s) 这些外部数据源虽然能够灵敏地指出‘某地发生了变化’，但受限于精度无法解析‘具体改变了什么’。Spexi 借此触发针对性的无人机复飞，通过超高分辨率影像精确捕捉地表地貌的真实演变。 — 比尔·莱克兰 > [14:27](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=867s) 我甚至怀疑大家根本没听我这页 PPT 讲了什么，你们的注意力肯定全被 Niantic 制作的这个震撼的 Splat 画面吸走了。 — 比尔·莱克兰 > [16:44](https://www.youtube.com/watch?v=EO0P8a0GZwE&t=1004s) 真正弥合‘从模拟到真实’鸿沟的核心起点，在于具备超高分辨率、高频持续更新、几何精准且高度真实的底层基础航空数据。 — 比尔·莱克兰 ## Mentioned Spexi, Bill Lakeland, Ai4, Niantic Spatial, Google Maps, Apple Maps, Uber H3, Waymo, Pokémon Go, 3D Gaussian Splatting, Proof of Capture, Soulbound NFT, Spexagons ## Video description Bill Lakeland - CEO, Spexi One of the next major bottlenecks in Physical AI isn’t model architecture, it’s access to a living and geometrically accurate representation of the real world. This session explores why closing the sim-to-real gap requires foundational aerial data with three essential characteristics: ultra-high resolution, continuous refresh, and consistent quality at scale. Learn why these attributes are becoming critical for training world models and how a distributed network of more than 10,000 drone pilots can deliver them where satellites and traditional aerial imagery cannot, and see how this approach is already helping train Niantic Spatial’s foundation models. [Source](https://www.youtube.com/watch?v=EO0P8a0GZwE)

Source quotes

[0:04] 比尔·莱克兰: 物理人工智能与具身智能蓬勃发展的当下，如何让 AI 真正理解并融入现实物理世界成了一个核心瓶颈。 [0:04] 比尔·莱克兰: 未来的地图与空间模型必须成为物理世界高保真、高频更新的真实镜像。 [1:51] 比尔·莱克兰: 这类方法通过计算机模拟来构建虚拟世界，其核心优势在于生成速度极快、边际成本极低，且具备近乎无限的可重复性。然而……这类方案致命的短板在于其底层数据源本质上是‘想象中的数据’，缺乏物理世界中复杂而不可预测的真实观测真值。 [1:51] 比尔·莱克兰: 直接取材于物理实体世界，尽管采集与处理复杂度极高、构建周期更长，但它具备不可替代的优势：完全可验证，且牢牢扎根于可观测的现实真理之上。 [3:16] 比尔·莱克兰: 真实世界影像的价值并不仅体现在视觉观感的新鲜度上，更取决于其几何精度。这意味着数据必须精确匹配地理坐标，且具备严密的可量化与可测量属性。 [6:38] 比尔·莱克兰: 飞手在 App 端认领一个 Spexagon 后，其工作被简化到了极致：只需在起飞点点击‘开始飞行’按钮，随后在现场承担空域安全监控职责，确保符合航空监管规范。 [10:13] 比尔·莱克兰: 许多飞手仅凭一台价值 600 美元的入门级微型无人机，通过全职或密集执行任务，单日收益甚至能达到 500 至 1,000 美元。 [11:55] 比尔·莱克兰: 从高空俯瞰，城市并非在同一时间发生全盘改变，因此盲目全覆盖飞行不仅浪费资源，也毫无必要。 [11:55] 比尔·莱克兰: 这些外部数据源虽然能够灵敏地指出‘某地发生了变化’，但受限于精度无法解析‘具体改变了什么’。Spexi 借此触发针对性的无人机复飞，通过超高分辨率影像精确捕捉地表地貌的真实演变。 [14:27] 比尔·莱克兰: 我甚至怀疑大家根本没听我这页 PPT 讲了什么，你们的注意力肯定全被 Niantic 制作的这个震撼的 Splat 画面吸走了。 [16:44] 比尔·莱克兰: 真正弥合‘从模拟到真实’鸿沟的核心起点，在于具备超高分辨率、高频持续更新、几何精准且高度真实的底层基础航空数据。

Source quotes

[0:04] Bill Lakeland: One of the next major bottlenecks in Physical AI isn’t model architecture, it’s access to a living and geometrically accurate representation of the real world. [1:51] Bill Lakeland: imagined data [3:16] Bill Lakeland: digital fabric of the built space [6:00] Bill Lakeland: specagon [8:12] Bill Lakeland: fly [10:32] Bill Lakeland: proof of capture [10:32] Bill Lakeland: If we're creating ground truth layer, that's an absolute must. [12:54] Bill Lakeland: freshness clock [12:54] Bill Lakeland: self-directed capture system [14:27] Bill Lakeland: I'm not sure you guys heard anything I said on this slide. You're probably looking at this awesome splat made by Niantic. [15:31] Bill Lakeland: they need to connect, they need to drive together.

Original transcript

[0:04] Yeah, closing the the the sim tore gap with living drone imagery. So resolution, freshness, [0:10] accuracy, what foundational data needs to actually ground a world model. So my [0:17] name is Bill Aland. I'm the CEO of Specsy. We're we're a Vancouver company um that's obsessed with [0:23] with mapping the Earth from the air. Uh I've been doing that my entire career. uh [0:28] flying airplanes for many years. Uh mapping a lot of you folks might have seen some of my data on [0:34] an a mapping application, Google Maps, Apple Maps, somewhere like that where cities actually [0:38] give the cities and governments give the data back to over the years. Uh I've been steeped [0:43] in this space for a long time and obsessed with trying to make it better. Um and so uh you know [0:50] we're doing this with drones now. I switched from satellites to doing this with drones and we've [0:56] built now the largest standardized drone network uh in the world. And when I say standardized, [1:00] this drone network is automatically capturing capturing data like satellites and planes do [1:05] over large over large geographies. And why are we doing that? Well, I don't know if you guys have [1:11] looked at mapping platforms recently. You know, you look at your house, it may be a little bit [1:16] out of date. I'm envisioning this becoming very different in the next few years and we really [1:20] want to be a part of that and be a part of this next movement in physical AI and world models [1:25] and being a really ground truth input into what that is and you know in the drone space people [1:30] have visioned at some point in this world that drones are flying all over cities all the time [1:37] but has anybody ever really seen that in reality it's actually happening uh now and we're a part [1:43] of what that is and I'm going to dive into how and why why that's actually happening. [1:51] So when we talk about teaching AI about the physical world, they're really two main [1:56] ways to train. Uh this can be done by simulating the world um like synthetic first type models. [2:03] This is this is the most that are out there. And there are a number of cool models and use cases [2:07] around and great companies that are working on some of these things. These methods are [2:12] are typically, you know, relatively fast. um less costly and like infinitely repeatable. The [2:21] the only issue from my perspective is that it's imagined data as the source training data [2:26] as opposed to the real observed truth of what of what it could be. Um the other method can be done [2:34] by training on what is actually physically present. So the real real world first [2:41] and there are a few incredible companies doing this as well. Um, [2:45] and this method typically is more complex. It takes longer to build, but it's verifiable [2:52] and it's grounded in observable truth. This is the group we're working more closely with [2:58] today. But we do believe that there needs to be some standardization, some protocols to be able [3:03] to create a reference to be able to train on top of that's reliable and consistent. And and we're [3:10] we're bent on creating a a major part of what that is. [3:16] So when it comes to real world imagery, um you know, recent visual data is fantastic [3:23] and it it really shows the current state of the physical world. Um [3:28] but even better yet, if it is geometrically accurate, it is so much more usable. Meaning, [3:35] is it positionally in the right location? Is it is there is are the dimensions measurable? [3:45] Our company was formed uh for many years of of this type of precision mapping where accuracy, [3:50] high quality measurements and geometry have always been kind of core to to the business [3:54] and the functions. Um and so we've simply taken that those principles and applied [4:01] them to this this new model, this new this new drone model in the platform. [4:06] And what we're doing now is building this highly accurate reference data set [4:12] across the built environment that can easily be ingested into training workflows [4:16] and real world real world first models. So the imagery layer is continually updated. [4:23] Uh and so if you want to imagine the super high resolution digital fabric of the built space [4:31] that can be used as a framework, a base map to layer other information, [4:37] a modalities against and also to make those data sets geometrically accurate as well [4:43] and therefore more usable in real world first models. [4:49] So, as we all know, there's incredible amount of of real world uh data that already exists of which [4:56] can easily be refreshed. There's AR applications out there. There's game applications. Pokémon Go [5:00] is a great example of one of those. There's fleets of sensors like dash cams and LAR systems and cars [5:08] like Ubers and Whimos. There's transportation and logistics companies that are constantly collecting [5:13] data as they're moving around. Most of which are non-standardized for external consumption [5:20] uh and quite often used for their own competitive advantage. Uh and none of these that I mentioned [5:26] are from the air, which means there's a gap. There's holes all across these cityscapes. [5:33] I've deliberately not included satellites and airplanes in this particular [5:39] incidental bucket. Um either they are too coarse in resolution to be meaningful for [5:45] the type of purposes we're talking about or too infrequent or expensive [5:50] to acquire at the resolution and the pace needed for a truly living model. [6:00] So consistent drone drone imagery is becoming the next essential data set for real world models. [6:05] That's what we're seeing because it's ultra high resolution. It's accurate and it's solid geometry. [6:12] It has incredible capacity globally, meaning drones are everywhere now. [6:18] And it makes the economics feasible. You just need a software platform that connect them and [6:22] put them to work in a very standardized way. and it's extremely refreshable [6:29] which is a key requirement for connecting all new modalities into the training fabric. [6:38] So how does this actually work? We created a platform to automate collection [6:42] through to image and product output. So this end end to end sort of collection output model. [6:48] Uh by automating drone flights, we can completely standardize the capture and task capture patterns [6:56] that are optimized for model training. Meaning we're extremely smart about how we fly, [7:01] where we fly, when we fly, and exactly what images are taken. We then take those captured [7:08] images and either process them into mapping products or feed them in through like an API [7:13] uh system to customers pipelines and and their internal workflows. Then all images have embedded [7:21] metadata such as GPS timestamp, XYZ, orientation information, all sensor details. So you know all [7:29] the things that make them easily ingestable and and uh and useful inside of an external workflow. [7:37] And if we're going to dig a little bit deeper into this as some people might kind of be interested to [7:41] understand how it actually works, we use a we use a unit of flight which is an Uber H3 level [7:46] 9 cells roughly 25 acres in size. We call them specagons which is a hexagon spatial hexagon. [7:53] Um, and in these cells we've embedded uh we can embed flight missions [7:58] and any type of capture m pattern or rosette that we really want inside of these, but we [8:03] specifically focus on five right now uh that are unique to creating outputs that are most valuable. [8:12] So once a pilot claims the specs are gone, their job is [8:16] incredibly simple. They just hit a fly button on the app. They're there. They monitor the [8:22] airspace. They maintain safety standards and protocols associated with all regulation. [8:29] These are drone enthusiasts. These are like hobbyists. This is a decentralized network. They [8:34] own their hardware. We only use micro micro drones in this system. That's incredible global capacity [8:42] and and lower risk barriers. There's a bunch of great reasons why we actually use these and [8:47] I could talk about that at length. the data is still better than what we get out of airplanes [8:51] at a low at a low elevation. So, it's quite incredible what this capacity can actually do. [8:56] We're flying at an elevation that outputs a resolution that's better than the status quo [9:00] uh and and high enough in quality to train models from and to reference groundbased data sets too. [9:09] So, it's a perfect ground truth ground truth data layer standardized and automated. [9:14] So, once the pilots hit the fly button, it's fully automatic. Our software takes over [9:19] executing perfect pre-planned flight plans and it guarantees every single data capture [9:25] is consistent. Each flight is fast about 7 to 10 minutes. [sighs] And we remove the complexity of [9:32] manual flight just to make sure that our data set is reliable, standardized, consistent and usable [9:37] every single time. Uh and when I mentioned the embedded flights, we've got five different ones. [9:42] Some of them are like panoramas. So if you can envision Google Street View but in the sky across [9:47] in the sky across an entire city. There's mapping flights that what you would get from [9:50] an airplane for photoggramometric workflows. We have hybrid flights which we do the most which are [9:56] oblique panoramics with mapping flights. And we have missions that are optimized for [10:01] dodge and splatting. So super high fidelity 3D reconstruction and uh and working with partners [10:07] one in the room which I'll talk about here in a little bit um with those which is really exciting. [10:13] So depending on the mission type, they can be different lengths. And drones also track [10:17] the Earth. They stay the exact same distance above ground at all time. That also ensures [10:21] that we maintain privacy in terms of you can't recognize people. You can't read license plates, [10:26] house numbers. We're not collecting any personal information at the altitudes that we're flying at. [10:32] So it's fast and easy for pilots to then go and upload. We actually created something called pre [10:36] capture. It's a blockchainbased backend system. So when the drones are flying, [10:39] we're actually creating soulbound NFTs in real time that authenticate to the [10:43] images as they're being collected. So pilots can't spoof the data whatsoever. [10:48] This being truth data, the authenticity, the provenence is absolutely paramount to this. [10:53] If we're creating ground truth layer, that's an absolute must. [10:57] And and we've created a system to be able to do that, which is quite incredible. It does [11:01] use the blockchain. It's pretty excit I'm really excited about um because we've not had one image [11:06] spoofed ever. Pilots get paid immediately once the imagery passes proof proof of capture. [11:12] Some pilots have been making $500,000 a day doing this off a $600 drone. So, it's pretty wild. [11:20] So, in a little over a year, we've uh [sighs] we've flown nearly 250,000 unique uh flights [11:27] across North America. That's like, you know, a quarter of a million separate automated missions. [11:32] Uh those flights were performed by over 10,000 pilots now in the network. And [11:37] we're looking at 7 million acres uh of coverage at this sort of standardized high resolution [11:42] geometrically accurate imagery. And we now have nearly 40 million images uh in that database that [11:47] are optimized and primed to feed into world uh real world uh model training workflows. [11:55] So I brought up uh or brought up aircraft and satellites earlier and I'd like to to touch [12:03] a little bit at how we're different than than them. And it really comes down to a couple [12:06] things. Resolution obviously is one of them from satellites. Uh it's just harder to do what we're [12:11] trying to do at a coarser resolution. And cost is the other big one. Aircraft can do at a pretty [12:16] good cost, but they must fly in a big area. Um and so it makes it really hard to make things living. [12:22] It makes it really hard to keep it fresh. And so, you know, we can actually keep regions living [12:28] because the unit economics can allow it without an unrealistic cost associated to flying. [12:36] So, now we're really able to focus on reflights on what is actually just changed. We cover a city and [12:41] now we just fly at a much higher cadence what has changed and keeping this layer as as living as as [12:48] as possibly we can. And that is the big next step that we're focused on. So because of that [12:54] really understanding change is becoming more and more important to us and when you look at a city [13:00] from above you know it doesn't all change at once. So why would we continually fly the whole thing? [13:07] What we do is measure the freshness clock within each spectagon and compare that to how much [13:14] has actually how much change has actually occurred. And to to to start this off, we you [13:21] know, we do use some satellite imagery. It's great at saying something's changed. Doesn't know what [13:25] has changed. It's it's identified that something has changed in this region. Same with using like [13:29] construction site permits and groundbased sensor outputs. Again, they're all really [13:34] good at saying something's changed and where it's changed, but they don't know exactly what it is. [13:41] So, this is how we can trigger at first our reflights and then our imagery actually shows [13:45] what the train change actually is. and we can create a new identification layer. [13:51] And so we're pushing this notion even further now and building a model ourselves. Uh [13:56] this one predicts where change will occur in our effort to create a self-directed capture system. [14:04] This learns, feeds back, refineses, and redirects [14:08] continually. And as we fly each spec time and time again, the feedback to model training loop [14:14] continues to refine and improve. And we're really starting to see the beginnings of of this actually [14:19] taking shape in regions where we have flown a lot continuously. [14:27] So I'd like to uh share a great partnership example here. uh our team has been uh working [14:32] closely for you know a couple years now uh with with the team at Niantic Spatial who have been I [14:39] think a great presence also at this conference and uh and who we've recently partnered with to create [14:45] a drone to 3D god and splat pipeline at city scale and so those products are actually just starting [14:50] to go to market now where a city or construction site uh end user can actually you know task drones [14:57] to collect imagery for god and splats that are injected into their individual workflows. [15:02] So, working through some really cool go to markets with those types of products. And I'll show you [15:08] a little bit more of what we do. This is more of the the city a city size. This is actually [15:13] one of the first ones I think we ever did together. It's a little bit older now, [15:18] but you can catch broad broad areas and super high fidelity. This one is uh a little bit [15:24] higher in resolution. Um more a more recent frame flight in San Francisco. And one thing we've like [15:31] deepened our understanding around is that a real world model needs to be multimodal. Air, [15:37] ground, they need to connect. They need to they need to drive together. And that's actually like [15:42] a really difficult problem to to solve that we've been seeing. Um, and so because PEXD is processed [15:50] into this geometrically accurate format and covers everything, it becomes that bond. [15:57] And because it's dense enough and has enough detail for the other modalities to [16:03] connect to it, can see the details and connect themselves to it. This increases the value [16:08] usability of all that incidental data I mentioned earlier. It's an absolutely ginormous resource. [16:15] But a key factor here is also recency. It becomes incredibly important. Ideally, the connecting data [16:21] sets need to be close in time, making the bond much stronger and the models tighter. [16:29] So, recency again is core to what we do at Spexi. [16:35] And uh I'm not sure you guys heard anything I said on this [16:37] slide. You're probably looking at this awesome splat made by Antic. [16:44] So, this is not just a a great theory. The platform's up and running. Um, and it's real. [16:49] And I'm excited to say that world models are actively being trained on this data today now. [16:56] And so what's next really for for us, you know, we're we're growing [17:00] we're growing our partner base, some of whom are in the room today. [17:03] Uh we're deepening our our vertical depth with insurance enterprise mapping, [17:08] physical AI, and world model teams with our ground truth capabilities and continually increasing [17:13] coverage and advancing our pilot network into new cities, refreshing existing coverage every day. [17:21] So, I'd like to wrap our our our view of the world here up a little bit and and and [17:27] that closing the sim tore gap really starts with foundational aerial data. Ultra high resolution, [17:33] frequently updated, accurate, authentic, and real. [17:38] The world never stops changing. Physical AI and world models need to know where and what that [17:43] change is. And that's the world we live in. So, please go to speciot.com to learn more about who [17:50] we are or come by our booth. We'd be happy to set set folks up with the data set to test. Thank you.