Roadmap
2026 milestones and the road ahead β from engine launch to the world's first large-scale DAgger dataset.
Our roadmap is driven by two core objectives: closing the product loop from data generation to model training, and building a thriving, globally scalable open ecosystem.
| Phase | Timeline | Focus | Headline |
|---|---|---|---|
| 0 β Foundation | Q1 2026 β | Core loop validation | Browser teleop β cleaning β training, end-to-end |
| 1 β Engine & Dataset | Q2 2026 β | Engine launch + flagship dataset | Franka Sim Dataset β 207 tasks, 50K+ demos; Ο0.5 to 88.8 on LIBERO-Plus |
| 2 β Pipeline & Global Scale | Q3 2026 π§ | V2 closed loop + global rollout | Arctivar V2 launched β post-training tasks live |
| Next 6-12 months | Aug 2026 β mid-2027 | Data engine Γ ecosystem Γ commercialization | World's first large-scale DAgger dataset Β· $1M+ ARR |
| Long term | β | Open infrastructure | Foundation models, interoperable data markets, verifiable coordination |
Phase 0 β Foundation (Q1 2026 β )
Platform
- Web-based teleoperation via MuJoCo WASM β single-player, keyboard control
- Task lobby, account system (Privy), and data upload pipeline
- Initial task set: tabletop pick-and-place, push, sort
Technology
- Unified trajectory format and scene configuration schema
- Basic data cleaning pipeline (filtering, smoothing, upsampling)
- First model trained on collected data; published Arctivar V1 paper
Ecosystem
- Internal data collection and testing
- On-chain data validation and provenance live on Base
Phase 1 β Engine & Dataset (Q2 2026 β )
Product β Task Generation Engine (API / Web Service)
We will launch our second core module by opening the Task Generation Engine to the public. As the driving force behind data diversity and our core commercial unit, it generates tasks across 4 dimensions: scenarios, atomic capabilities, intra-task layouts, and asset randomization. The engine supports a vast library of mainstream robotic assets, rigid bodies, and articulated objects β ensuring unmatched representativeness, scalability, and real-world applicability.
- Multi-embodiment coverage: single-arm, dual-arm, dexterous-hand, wheeled bimanual, and humanoid embodiments unified under one asset library
Business β Franka Sim Dataset Release
We published a simulation-based Franka Research 3 dataset built entirely from browser-collected demonstrations: 207 manipulation tasks, 50,000+ human demonstrations, and 60,000+ task and scene variants across seven scene categories β a showcase of our "Browser + Retail User" crowdsourcing model, and evidence the platform reaches scale, quality, and diversity together.
Phase 2 β Pipeline & Global Scale (Q3 2026 π§)
Product β Automated Training & Validation Pipeline (Q3)
We are building out our end-to-end Data-to-Model pipeline. Users will be able to automatically train models using our platform's data and seamlessly validate the real-world performance of their trained policies β fully closing the loop from raw demonstration to deployable model. The first half of that loop shipped with V2 (below); automated training and validation is the piece still landing.
Operations β Global Market Expansion (Q3)
We will establish a strong presence in 5β7 core regional markets: North America, Europe, SEA, CIS, Africa, and East Asia. The goal is to build scalable, localized best practices β including massive distributed data collection communities, robust data partnership networks, and high-quality academic/enterprise collaborations.
Academic & Commercial (from August)
- Publish 1β2 academic papers based on our proprietary pipelines, focusing on 3D-ready asset generation and novel simulators addressing the sim-to-real gap
- Begin converting early POC customers into recurring revenue streams
The Next 6-12 Months
Fueling the next-gen Physical AI β from data scale to general-purpose robots. Three tracks run in parallel: scale the data engine, deepen the ecosystem moat, and close the commercial loop.

1 Β· Data Engine β Scaling the Proven Paradigm
- Ego-Centric Pipeline (in development). First-person capture in the mobile app, to maximize in-the-wild diversity.
- Oct 2026 β Sim Dataset V2 (in progress). 10,000+ hours and millions of data trajectories across multiple embodiments and a broader set of atomic skills. Building on V1's success (Ο0.5 to 88.8 on LIBERO-Plus), V2 further proves that massive, noisy crowdsourced data exponentially improves policy training.
- EoY 2026 β World's 1st DAgger dataset. Releasing massive human-gated, failure-recovery data β filling a critical industry gap for model correction.
- Q1 2027 onward β Continuous scaling of all three data streams.
2 Β· Global Ecosystem β The Uncopyable Moat
- Massive network (SepβOct). 100k+ global contributors reached β (229,000+ as of 10 October 2026). Expanding across LATAM & Eastern Europe.
- High-velocity yield. Achieving daily generation of 500+ hours (ego) and 50+ hours (sim).
- Academic influence. Deepening strategic partnerships with 5+ top-tier AI labs.
3 Β· Commercialization β Closing the Loop
- EoY 2026 β $1M+ exit ARR. Reaching a $1M Annual Recurring Revenue run-rate through expanded data subscriptions and learning-based sorting solutions.
- EoY 2026 β Paid pilots. Securing 2β3 additional paid pilot contracts with leading hardware and model companies.
- Q1 2027 β Tier-1 vendor. Officially entering the data vendor list of at least one Tier-1 Physical AI foundation model company.
- Mid-2027 β 3Γ growth. Achieving 3Γ YoY revenue growth.
Long Term β Open Infrastructure
The long-term vision for Arctivar is to become the open infrastructure for robotic intelligence β a network where data, tasks, and models flow freely with verifiable provenance and clear participation rules.
Foundation Models for Embodied Intelligence Cross-embodiment foundation models trained on the full community dataset, validated for transfer across heterogeneous hardware, and deployed via a continuous sim-to-real pipeline.
Open Data & Model Markets A data marketplace and a model arena where qualified participants can publish datasets or policies under defined licensing terms. Simulation-as-a-Service and Augmentation-as-a-Service productize our internal capabilities for the wider industry.
Verifiable Coordination Layer On-chain contribution provenance gives task creators, data contributors, and model trainers a shared audit trail. Any future participation or coordination mechanisms will be described separately if and when they are finalized.
Next-Generation Interfaces VR teleoperation with full hand and body tracking unlocks dexterous and whole-body behaviors that keyboard and mobile interfaces cannot capture.
Next: Getting Started β