World Engine

1. Executive summary

World Engine (WorldEngine AI) is a physical-AI infrastructure company about ten months old. It was founded by Yue Wang, USC Assistant Professor, director of the Physical Superintelligence (PSI) Lab and former NVIDIA Research faculty researcher, and Di Huang, co-founder of Google Cloud Vertex AI and co-lead of Waymo's first-generation Foundation Model. The company sells the data, evaluation and infrastructure that frontier labs need to train humanoid foundation models, and publishes an open humanoid model (Ψ₀) that acts as a reference design for the ecosystem.

• Traction (company-reported): ~$100M 2026 run-rate, 8 frontier-lab customers, 30 employees, $15M to deliver within two months, 200K hours of human data processed to date. Run-rate and customer count are echoed publicly by investor UpScaleX.
• Insight: embodied AI is data-starved. Humanoid models train on roughly 1–2T tokens versus 100–1,000T for autonomous driving and 50–100T for LLMs, with almost no deployment surface to learn from.
• Edge: a full-stack position across capture hardware, post-processing algorithms, operations and a reference model, covering three data tiers (ego-centric, sensorized glove, teleoperation).
• Credibility: three CVPR 2026 honors (two best-paper, one runner-up), an open Ψ₀ model with ~2.9k GitHub stars, and a +40% success lift over strong humanoid baselines using ~8% of the data.
• Principal risks: unverified revenue quality and concentration, labor-intensive delivery and margins, insourcing by labs, well-funded rivals (XDOF at ~$1.2B in talks, Mecka at ~$500M reported), and data-rights exposure.

2. Company overview

3. The problem: a physical-AI data wall

LLM agents were built on internet-scale text and serve billions of users. Autonomous driving, with 6–6.5M vehicles on the road, generates 100–1,000T tokens of training data. Humanoid foundation models, in contrast, are 3B-class models trained on 1–2T tokens with only sparse robot deployments to learn from. Digital intelligence therefore compounds quickly while physical intelligence is flat, which the deck frames as a wall that new data sources must break.

Three properties make humanoid data hard. It is expensive (hardware, operators, annotation), embodiment-specific (human motion must be retargeted to a robot's body), and curriculum-sensitive (the right mix of cheap broad data and costly precise data matters more than raw volume, as Ψ₀ demonstrates).

4. Solution and product

World Engine positions itself as a "physical API": labs submit data specifications, a global workforce wearing capture devices returns experience, and World Engine's pipeline converts it into humanoid-ready training data.

• Tier 1, large-scale ego-centric data (pre-training): head-mounted capture; 200K → 1M hours targeted.
• Tier 2, sensorized dexterity (post-training): instrumented gloves capturing contact-rich hand motion; 20K → 200K hours.
• Tier 3, teleoperation (alignment): highest-fidelity robot-in-the-loop data; 5K → 50K hours.
• Model layer: an autoregressive transformer over interleaved language, vision, world-state and action tokens, used to learn which data the model needs (pre-training, post-training, real-world RL, agent orchestration).
• Ownership of the stack end to end (capture hardware, post-processing algorithms, operations) so each hour is engineered to "speak humanoid".

5. Technology and research credibility

The research output is the strongest evidence of technical depth. Ψ₀ (RSS 2026; arXiv 2603.12263; co-authors include NVIDIA's Marco Pavone) predicts a 36-dimensional whole-body action chunk covering hands, arms, torso and locomotion for humanoid loco-manipulation. It was trained on about 800 hours of human video plus 30 hours of robot data and reports 40%+ higher overall success than strong baselines (GR00T N1.6, π0.5, InternVLA-M1, H-RDT, ACT, Diffusion Policy) on eight real-world tasks.

Caveat: results are from the founders' own lab on eight tasks. Independent replication, and transfer to customers' embodiments at production scale, is not yet demonstrated. The open release lowers this risk because third parties can test the claims.

6. Team

Yue Wang (MIT PhD) brings 21K+ citations, 80+ papers, eight best-paper awards, an NVIDIA Fellowship and a Toyota Young Faculty Fellowship. He raised $3.6M for a 35-person lab within 2.5 years, evidence of fundraising and team-building ability. Di Huang scaled Waymo's foundation model from 7B to 140B parameters, the first robotic foundation model in production, on 1,000+ robotaxis, and co-founded Google Cloud Vertex AI. He supplies the production, infrastructure and operations credibility that pure researchers lack.

The wider team of 30 spans foundation-model, infrastructure and data, and product functions with alumni from Google Gemini, NVIDIA, Meta Superintelligence, ByteDance Seed, Amazon, Microsoft and AI2, plus an operations lead who has managed 5,000-person operations and a Stanford physical-AI advisor. Observations: the bench is young and heavily USC-sourced, which is a strength for speed and alignment but a risk for management depth as headcount scales toward the 12.5M-hour plan.

7. Traction and unit economics

Reported traction is exceptional for a ten-month-old company: ~$100M run-rate from eight frontier labs implies an average of ~$12.5M per customer, and the $15M-in-two-months delivery commitment is consistent with a ~$90M annualized pace.

Public posts by UpScaleX cite about 250,000 hours of humanoid data processed per month, while the deck says 200K hours processed in total. These are very different measures: the first implies a ~$33 per hour blended price at a $100M run-rate, the second would imply a much higher price per hour. Reconciling definitions (hours captured vs. processed vs. delivered, and by tier) is the single most important diligence item.

8. Market

The direct market is training data and infrastructure for humanoid and dexterous-manipulation models. It sits upstream of humanoid hardware, whose forecasts vary widely: roughly 13,000–16,000 humanoids shipped in 2025 with sector revenue below $1B, Goldman Sachs now sees 6.5M units and a ~$138B market by 2035 (up from $38B), Morgan Stanley $4.7T by 2050 and Citi $7T by 2050.

Capital is moving ahead of revenue: physical-AI funding reached ~$50B across ~500 deals in H1 2026 (Crunchbase). Data buyers are the best-funded labs, so near-term demand does not depend on humanoid unit sales, but long-term pricing power does depend on robots reaching real deployment.

• Bull view: every humanoid program needs orders of magnitude more data; data is the scarce input, as in LLMs where Scale AI, Surge and Mercor became multi-billion-dollar vendors.
• Bear view: labs insource capture once recipes stabilize, simulation and synthetic data substitute for part of the need, and hardware timelines slip.

9. Competitive landscape

World Engine's argument is that the sensorized-glove tier for dexterous post-training has no direct competitor, and that being both a data vendor and a model builder gives it feedback no pure-collection company has. The counter is that rivals have raised at higher valuations on far less disclosed revenue, so category enthusiasm may compress margins and invite well-funded entrants.

10. Business model and roadmap

• Data as a Service (today): sell ego-centric, glove and teleop data and evaluation to frontier labs.
• Infrastructure as a Service (mid term): license the capture, processing and deployment stack; VinMotion is the first customer.
• Deployment pilots (long term): vertical robot deployments such as the Becton Dickinson pilot, which generate real-world RL feedback.

The plan implies a 10x hours increase in 12 months and 50x in 24 months. That requires a step-change in recruiting, device logistics, quality control and tooling, and it is the main execution test for the next two quarters.

Sources

1. Ψ₀: An Open Foundation Model Towards Universal Humanoid Loco-Manipulation, arXiv 2603.12263
2. Ψ₀ project page and code (PSI Lab)
3. Yue Wang, USC
4. UpScaleX LinkedIn post on WorldEngine run-rate
5. PitchBook: WorldEngine AI profile
6. TechCrunch: XDOF in talks for Series B at ~$1.2B (Sep 4, 2026)
7. TechCrunch via Yahoo: Mecka AI nears $500M valuation (Sep 11, 2026)
8. The American Bazaar: Mecka AI raises $60M (Oct 8, 2026)
9. Motley Fool via Yahoo: Goldman Sachs raises humanoid forecast (Sep 14, 2026)
10. RoboZaps: humanoid robot market size compilation
11. Forbes citing Crunchbase: physical-AI funding H1 2026