Key Takeaways
- Kosmo is Hesai Group’s end‑to‑end 3D data engine that couples the Kosmo One AI spatial camera with the Kosmo World cloud service.
- The camera records 8K ultra‑HD imagery, high‑density LiDAR points, and sub‑millisecond IMU data, delivering millimeter‑level geometry and true‑to‑life material rendering.
- A single operator can digitize a 100 m² indoor area in ≈ 15 minutes and upload the scene to the cloud for robot training, virtual production, or location‑based services.
- Strategic alliances with Galbot, Dexmal, Pixar’s RenderMan, and Amap showcase cross‑industry potential in robotics, filmmaking, and navigation.
Introduction: A New Data Engine for Physical AI
Hesai Group announced Kosmo, a complete 3‑D world‑data platform designed to eliminate the data‑scarcity bottleneck that hampers embodied‑AI development. By merging the world’s first AI spatial camera—the Kosmo One—with a purpose‑built cloud repository called Kosmo World, the solution converts physical spaces into high‑fidelity digital twins that can be reused for robot perception training, world‑model simulation, and immersive content creation.
How Kosmo One Captures Reality
8K Ultra‑HD Spatial Imaging
Kosmo One integrates a one‑inch CMOS sensor capable of 8 K (7680 × 4320) video at 30 fps. The sensor works in tandem with a high‑resolution LiDAR unit (up to 1 M points per second) and a precision IMU that timestamps every frame with ≤ 10 µs latency.
3‑D Gaussian Splatting (3DGS) Reconstruction
Captured data are processed on‑board using 3DGS algorithms, which convert point clouds and textures into a continuous volumetric representation. The result is a digital asset with ≤ 1 mm geometric error and realistic lighting, shadows, and material attributes.
Operator‑Centric Workflow
- Instant start‑up: Scanning begins automatically when power is applied.
- Guided acquisition: AI‑driven prompts steer the operator to cover blind spots.
- One‑click resume: Interrupted scans can be continued without data loss.
A typical 100 m² indoor environment—such as a warehouse aisle or a conference room—can be fully digitized in ≈ 15 minutes, producing a reusable 3‑D asset of roughly 150 GB (raw) or ≈ 30 GB after compression.
Kosmo World Cloud Platform
The Kosmo World service stores, indexes, and streams the generated assets to downstream applications. Key features include:
| Feature | Description |
|---|---|
| Scalable storage | Up to 10 PB per region, with automatic tiering. |
| API‑first access | REST and gRPC endpoints for retrieval, annotation, and version control. |
| Edge inference | Built‑in GPU containers enable on‑the‑fly rendering for AR/VR headsets. |
| Security | End‑to‑end AES‑256 encryption and role‑based access control (RBAC). |
Developers can pull a scene directly into robot simulation environments (e.g., NVIDIA Isaac) or feed it to visual effects pipelines via Pixar’s RenderMan for photorealistic compositing.
Partnerships & Demonstrations
- Galbot and Dexmal (embodied‑AI startups) will use Kosmo assets to train navigation and manipulation policies, reducing their data‑collection time by > 70 %.
- A joint demo with Pixar RenderMan rendered a Kosmo‑captured factory floor at 4 K resolution, preserving surface reflectance and shadow softness.
- Integration with Amap enables real‑time location‑based services that overlay 3‑D maps onto mobile devices for indoor navigation.
Kosmo One vs. Conventional 3‑D Scanners
| Specification | Kosmo One | Typical High‑End LiDAR Scanner* |
|---|---|---|
| Resolution | 8 K video + 1 M pts /s LiDAR | 4 K video + 300 k pts /s LiDAR |
| Geometric accuracy | ≤ 1 mm (3DGS) | 2–5 mm |
| Sync precision | ≤ 10 µs (IMU‑Lidar‑Camera) | 100–200 µs |
| Capture speed | 15 min for 100 m² | 30–45 min for 100 m² |
| Data size (raw) | ~150 GB | ~80 GB |
| On‑board processing | 3DGS reconstruction, AI guidance | None (post‑process required) |
| Price (list) | $49,999 | $35,000–$45,000 |
*Values represent industry averages for 2024‑2025 models.
Impact Across Industries
- Robotics – High‑fidelity digital twins accelerate simulation‑to‑real transfer, cutting robot training cycles from months to weeks.
- Film & Gaming – Direct ingestion of photorealistic scenes eliminates manual geometry modeling, shortening VFX pipelines by 30‑40 %.
- Location Services – Accurate indoor maps improve wayfinding accuracy, reducing user positioning error from ~1.5 m to < 0.3 m.
Bottom Line
Hesai’s Kosmo platform delivers a unified hardware‑software stack that bridges the gap between the physical world and AI‑ready digital data. By marrying an 8
