Architecture
The Closed-Loop Physical AI Data Flywheel — four hierarchical layers from global data infrastructure to real-world deployment.
Closed-Loop Physical AI Data Flywheel
Modern Physical AI models — embodied foundation models and world models alike — are constrained not by model capacity, but by the structure of their data loops.
Arctivar is a vertically integrated infrastructure designed around that constraint. It enables:
- Structured large-scale embodied data generation
- Model-conditioned feedback optimization
- Continuous deployment-driven improvement
The system is organized into four hierarchical layers, spanning foundational infrastructure to real-world deployment.

Layer 1 — Global Data Infrastructure
The foundational layer provides scalable, standardized, and defensible data generation capabilities. It governs contributor access, quality control, and representation alignment.
1.1 Retail-Accessible Global Data Interface
A low-friction, human-in-the-loop interaction layer accessible via web, desktop, and mobile platforms. This interface enables globally distributed contributors to:
- Capture everyday embodied interactions
- Execute scenario-specific task instructions
- Provide corrective demonstrations on model rollouts
The system is designed for scale, geographic diversity, and task parameterization — forming a continuously expanding embodied interaction network.
In the product: Getting Started · Pre-training Tasks · Post-training Tasks. The web interface covers simulation tasks; everyday ego-centric capture arrives with the mobile app, which is in development.
1.2 Automated + Human Data Validation
All trajectories pass through a multi-stage validation pipeline:
- Automated trajectory scoring
- Failure detection and anomaly filtering
- Human validation for corrective supervision
Additionally, model-policy gap analysis identifies high corrective-value samples. Dataset growth is therefore quality-aligned rather than volume-driven.
Deep dive: Data Pipeline · How a submission is scored
1.3 Model-Ready Embodied Data Schema
All data — simulation and real-world — is normalized into a unified embodied schema. The schema standardizes:
- Scene representations
- Action trajectories
- 4D hand-pose sequences
- Subtask-level temporal alignment
This alignment layer ensures direct compatibility with model training pipelines and cross-simulator interoperability, eliminating fragmentation between synthetic and real-world sources.
Deep dive: Simulation Engine — MetaSim
Layer 2 — Closed-Loop Hybrid Data Engine
Built atop the infrastructure layer, the data engine performs structured generation and adaptive distribution shaping. It integrates three complementary modalities.
2.1 Scalable Simulation Data Collection
Simulation environments enable:
- Controllable diversity
- Task parameterization
- Distribution shaping
- Rapid policy iteration
Simulation is used to explore scaling laws and cover rare or structured task regimes.
Deep dive: Simulation Engine (RoboVerse) · Data Pipeline
2.2 Ego-centric Real-World Data Capture
To ground simulation priors, the engine incorporates high-fidelity ego-centric data, including:
- 4D hand-pose trajectories
- Subtask-level temporal annotations
- Object interaction dynamics
This establishes manipulation priors and bridges the sim-to-real representation gap.
Rollout: ego-centric capture arrives through the mobile app, which is in development, targeting 10,000+ valid hours per month — see the Roadmap.
2.3 Model-Conditioned Data Generation
Data generation is not static; it adapts to model weaknesses. Using deployment feedback and policy-gap signals, the engine performs:
- Failure mining
- Weakness-targeted task construction
- Adaptive distribution reweighting
Data generation becomes an optimization process conditioned on model behavior — the mechanism behind which post-training tasks open next on the Hub.
Layer 3 — 3D-aware Model
The foundation model layer consumes structured embodied datasets and produces evolving policies. Training consists of two tightly coupled phases.
3.1 Large-scale Pre-training
Off-policy expert demonstrations initialize:
- Task representations
- Manipulation priors
- Long-horizon planning structures
Pre-training establishes broad behavioral coverage across environments.
3.2 Post-training with Human-gated DAgger
On-policy rollouts are corrected via structured human supervision. Corrective datasets are collected directly on model-generated trajectories, enabling:
- Systematic failure recovery
- Long-horizon stability improvements
- Reduction of compounding errors
The model evolves continuously through feedback rather than isolated retraining cycles.
Deep dive: Model Training & Sim-to-Real — including the theory behind human-gated DAgger and scalable online post-training.
Layer 4 — Deployment & Policy Rollout
The top layer executes policies across real-world and simulation fleets. Deployment generates:
- On-policy rollouts
- Failure trajectories
- Distribution drift signals
These signals feed directly back into the data engine — closing the optimization loop. On the platform, this is what contributors experience as post-training tasks: every community rollout is simultaneously deployment, evaluation, and targeted data collection.
System-Level Property
The architecture forms a self-reinforcing embodied intelligence flywheel:
Global Infrastructure
→ Adaptive Data Engine
→ Model Training
→ Deployment
→ Failure Signals
→ Targeted Data GenerationEmbodied intelligence training shifts from static dataset accumulation to continuous, closed-loop optimization.
The Production Pipeline
The concrete pipeline that implements Layers 1–3 runs from task generation through data collection, refinement, and training:

| Stage | What happens | Deep dive |
|---|---|---|
| World Generation | LLM-driven task generation, asset sourcing, automated layout construction on the MetaSim scene description | Simulation Engine · Data Pipeline |
| Behavior Collection | Browser teleoperation (MuJoCo WASM) for demonstrations; policy rollout with human gating for corrections | Pre-training Tasks · Post-training Tasks |
| Data Refinement | Replay verification, cleaning, IsaacSim augmentation with domain randomization, correction aggregation | Data Pipeline |
| Model Training & Deployment | Pre-training on demonstrations, post-training on corrections, sim-to-real validation | Model Training |
Every artifact carries provenance: trajectories are signable on-chain at collection time, and datasets and models trace back to their contributors through Data IDs. See Data Provenance & Contribution Records.
Design Principles
| Principle | What it means |
|---|---|
| Data generation must be model-conditioned | The model's measured weaknesses — not a fixed curriculum — determine what data gets collected next. |
| Simulation and real-world data must share a unified representation | One embodied schema across synthetic and real sources; no fragmentation between them. |
| Post-training is as critical as pre-training | Demonstrations establish coverage; corrections establish reliability. The platform is built to produce both. |
| Deployment feedback must directly shape data distribution | Failure signals from policy rollouts feed straight back into task construction and distribution reweighting. |
Next: Simulation →