The Compounding Data Engine
The flywheel that makes data compound, the three cornerstones and three products that expose it, and the four-layer engine beneath.
Most robot data pipelines are one-directional: collect data, train a model, ship it, start over. Arctivar is built as a compounding data engine — a closed loop where every contribution improves the next policy, and each improved policy determines what data to collect next. Data stops being a one-time cost and becomes an asset that appreciates.
What makes the coverage possible is unifying the two most scalable data sources there are: high-fidelity simulation and ego-centric human data. Neither alone reaches the diversity general-purpose manipulation needs; together they compound it at foundation-model scale.
This page is the map of the whole system: the flywheel that makes data compound, the three cornerstones and three products that expose it to the world, and the four-layer architecture that runs underneath. Every other page zooms into one piece of this map.
Why It Compounds

The engine runs as a loop, not a line:
- Generate the task and world — diverse scenes, objects, and skills, produced on demand rather than hand-built.
- Collect behavior — contributors teleoperate in simulation and capture real-world interactions, producing demonstration and correction data.
- Process the data — verify, clean, and expand each trajectory into thousands of photorealistic, domain-randomized training samples.
- Train & deploy — policies learn from the data, run on real and simulated fleets, and expose exactly where they still fail.
- Feed back — those failures decide which tasks open next, and the cycle repeats on harder, higher-value data.
Because step 5 targets step 1, the loop is self-reinforcing: the more it runs, the more valuable each new contribution becomes. A trained policy's weaknesses are not a dead end — they are the next data request.
The Three Cornerstones
Underneath the products sit three reusable cornerstones. They are what every part of the data business runs on, and what keeps quality consistent no matter which source the data came from.
| Cornerstone | What it does |
|---|---|
| Universal Task Generation Engine | Continuously streams diverse scenarios and comprehensive coverage of atomic skills to drive collection |
| Distributed Global Workforce | A decentralized network that removes the physical and geographical limits of data gathering |
| Unified Processing Pipeline | One standardized system that filters, augments, and transforms raw inputs into high-quality, golden datasets |
The Three Products
The cornerstones reach the world as three products — the surfaces contributors and customers actually touch. Together they cover the full loop.
| Product | Flywheel stage | Scale | |
|---|---|---|---|
| Web | Simulation Data Platform | Generate → Collect (simulation) | 10,000 valid trajectories per hour |
| Mobile | Ego-Centric Data Collection | Collect (real-world) | In development — target 10,000+ valid hours per month |
| Infrastructure | Data-to-Model Pipeline | Process → Train | — |
How the Task Generation Engine Feeds Both
Every task a contributor sees, on either platform, comes out of one generation engine — which is why diversity is engineered rather than left to chance. From a natural-language instruction, the engine multiplies five dimensions to generate exponentially growing data at low marginal cost:
- Embodiment — one native multi-embodiment infrastructure spanning 13+ major robots, from single-arm to humanoid, integrating new hardware at near-zero marginal cost.
- Object — the full spectrum of manipulation, from simple rigid items to complex articulated mechanisms for home and industrial use.
- Spatial — layout randomization that samples object placements across the robot's entire reachable workspace.
- Semantic — asset randomization that translates LLM-level knowledge into physical variety, generating sub-types within any object class.
- Visual — procedural randomization of textures, lighting, and camera poses for zero-shot robustness.
Web — Simulation Data Platform
Data collection shifted from legacy offline, sequential methods to an online, fleet-wide, massively parallel human-in-the-loop engine in the browser — no hardware, no install. It powers both halves of the loop: expert demonstrations for pre-training, and human-gated corrections (HG-DAgger) on live policy rollouts for post-training. Retail-friendly for contributors (zero learning curve) and hardware-agnostic for partners (plug-and-play with any robot or simulator). This is the product contributors experience as Pre-training and Post-training Tasks.
Mobile — Ego-Centric Data Collection
Real-world capture shifts from expensive, hardware-heavy rigs to a zero-barrier mobile ego-centric app. Pairing SOTA real-time hand-pose tracking with a global workforce, it translates human vision into robotic motion at scale — across everyday routines and domain-specific workflows, with modular setups from a single iPhone to multi-camera headbands, wrist cameras, and UMI grippers. Output is 4D hand-pose reconstruction paired with dense language annotation.
Infrastructure — Data-to-Model Pipeline
The central processing hub, and model-centric data engineering rather than raw accumulation. A rich toolset turns diverse inputs into training-ready formats: cleaning (filter invalid data, smooth jitter, resample to a common frequency), domain-randomized augmentation across visuals (camera, lighting, textures) and physics (object poses, clutter, mass, friction), and annotation (segmentation, context-aware descriptions, and frame-by-frame atomic labels syncing text, video, and 4D trajectories). Customers reach the same layer through automated evaluation toolkits and open-source datasets, taking raw data through to foundation models. The deep mechanism lives in Data Pipeline.
Products ↔ Architecture
The three products are not the same thing as the four architecture layers. Products are the outward capability surfaces — the flywheel stages, productized. The architecture is the engine underneath: four hierarchical layers from global data infrastructure up to deployment. This table is the single place that maps one to the other.
| Product (capability surface) | Flywheel stage | Runs on architecture layer(s) |
|---|---|---|
| Web — Simulation Data Platform | Generate → Collect (sim) | L1 assets, schema & global data interface → L2 structured generation and simulation modality → L3 post-training loop |
| Mobile — Ego-Centric Data Collection | Collect (real) | L1 mobile interface → L2 ego-centric modality |
| Infrastructure — Data-to-Model Pipeline | Process → Train | L2 model-conditioned generation → L3 model-ready data and model training → L4 deployment & rollout |
For the mechanism behind each layer, follow into Architecture. The three products above describe what each does and who it's for; Architecture describes how it's built.
Next: Roadmap →