Migrated from haga-core. Technical detail for diligence: Methodology report. Open/proprietary boundary: IP strategy.

Sources

Every figure and claim in the README's Market Context and Why Manipulation and Locomotion sections, traced to its origin. This section exists because Haga's entire premise is independent verification — an unsourced stats table would undercut that premise the moment someone checks it.

Market Sizing

• AI World Models Market (verified primary figure) — $5.8B (2025) → $28.6B (2034), 58.2% CAGR. "Interactive Physics Simulators" makes up ~31.5% of this market. marketintelo.com, AI World Models Market report. https://marketintelo.com/report/ai-world-models-market

• World Model Simulators Market — $170M (2026) → $3.06B (2036), 33.5% CAGR. Widely repeated in secondary sources; a primary report matching this figure exactly could not be independently located. Treat as directional context, not a verified data point, until a primary source is confirmed.

• Synthetic data market — $710M (2026) → $3.67B (2031). Mordor Intelligence, cited in: Synthetic Data Startup Statistics, Mean CEO Blog (May 2026). https://blog.mean.ceo/synthetic-data-startup-statistics/

• Synthetic data generation market — $791.34M (2026) → $6.91B (2034). Fortune Business Insights, cited in the same Mean CEO roundup above. https://blog.mean.ceo/synthetic-data-startup-statistics/

• Physical AI and robotics startups raised $27.6B across 1,009 deals in 2025 (>2x 2024). PitchBook data, cited in: Physical AI Startup Statistics, Mean CEO Blog (May 2026). https://blog.mean.ceo/physical-ai-startup-statistics/

Funding Rounds Referenced in the Market Context Table

• Antioch — $8.5M seed, simulation tools for robot builders without in-house sim capacity ("Cursor for physical AI"). TechCrunch, April 16, 2026. https://techcrunch.com/2026/04/16/this-simulation-startup-wants-to-be-the-cursor-for-physical-ai/

• Bifrost AI — $8M Series A (October 30, 2024), led by Carbide Ventures with Airbus Ventures, Peak XV's Surge, Wavemaker Partners, MD One, and Techstars participating. Total raised to date: $8.56M (not $13.7M as stated in an earlier draft of this document — corrected after verification). PRNewswire, Oct 31, 2024: https://www.prnewswire.com/news-releases/bifrost-raises-8m-to-accelerate-ais-ability-to-solve-problems-in-the-physical-world-302292864.html CB Insights (total raised figure): https://www.cbinsights.com/compare/bifrost-2-vs-sky-engine

• Patronus AI — $50M Series B, "Digital World Models" for simulation-based agent evaluation before deployment. Total raised to date: $70M. Investors include Lightspeed Venture Partners, Notable Capital, Datadog, Samsung. TechNews180, reporting on the raise. https://technews180.com/funding-news/patronus-ai-secures-50m-to-build-ai-training-simulations/

• Instance — physics-aware quality layer for AI-generated video, scanning synthetic video for physics violations to benchmark real-vs-fake at scale. YC-backed (specific batch not independently confirmed — do not cite a batch number without reconfirming). Y Combinator company directory, Robotics category. https://www.ycombinator.com/companies/industry/robotics

• Robocurve — open-source tools and independent, reproducible robotics benchmarks; ships an open-source framework called "Inspect Robots" (v1 shipped, first pilot scoring a frontier model on a real robot completed). YC-backed (specific batch not independently confirmed). Explicitly positions itself around real-hardware benchmarking, not simulation: "simulation only goes so far, since models that look strong in sim can show large performance gaps once deployed in the real world" — a direct, relevant tension with Haga's sim-only v0 scope (see Risks). Y Combinator company page: https://www.ycombinator.com/companies/robocurve

Prior Synthetic-Data Funding Cited as Category Validation

• MDClone ($63M Series C, 2022), Gretel ($50M Series B, 2021; later acquired by NVIDIA per public reporting), Datagen ($50M Series B, 2022), Tonic.ai ($35M Series B, 2021), Parallel Domain ($30M Series B, 2022), MOSTLY AI ($25M Series B, 2022), Synthesis AI ($17M Series A, 2022). Compiled in: Synthetic Data Startup Statistics, Mean CEO Blog (May 2026). https://blog.mean.ceo/synthetic-data-startup-statistics/

Technical Stack Claims

• MuJoCo is free, open-source (Apache 2.0), and runs on Windows/Linux/macOS. MuJoCo official documentation. https://mujoco.readthedocs.io/en/stable/programming/index.html

• MJX (JAX reimplementation of MuJoCo) targets NVIDIA/AMD GPUs, Apple Silicon, and Google Cloud TPUs via XLA — but on Apple Silicon, mjx.step() fails on the Metal backend with error: failed to legalize operation 'mhlo.cholesky'. Root cause: Apple's Metal/MPS JAX backend does not implement the Cholesky decomposition op that MJX's smooth.factor_m() requires (call chain: mjx.step() → forward() → fwd_position() → smooth.factor_m() → jax.scipy.linalg.cho_factor()). This is structural, not a version mismatch: jax-metal (the plugin needed for Metal support) was abandoned by Apple after its last release in Oct 2024, and the community replacement jax-mps is early-stage and likely to hit the same missing-op issue. CPU backend works correctly (confirmed: all tests pass on JAX_PLATFORMS=cpu). Do not expect Metal GPU acceleration for MJX to become available soon. Primary source, most detailed and current: https://github.com/talmolab/stac-mjx/issues/126 (filed Mar 3, 2026, tested on M4 Pro/macOS Sequoia) MuJoCo official documentation, MJX section: https://mujoco.readthedocs.io/en/stable/mjx.html jax-metal abandonment: https://github.com/jax-ml/jax/issues/34109 Earlier reports of the same Cholesky limitation: https://developer.apple.com/forums/thread/748554, https://github.com/jax-ml/jax/issues/16321

• Robosuite is a MuJoCo-based manipulation benchmarking framework with standardized tasks (Lift, Stack, PickPlace, Door, etc.), macOS-compatible (viewer requires mjpython). MuJoCo on macOS setup guide, Sangil Lee (March 2026). https://sangillee.com/2026-03-04-start-mujoco/

• Aerial Gym Simulator and Pegasus Simulator are built on the NVIDIA Isaac stack and require CUDA — not available on Apple Silicon. Best-of-list of robot simulators (community-maintained, weekly-regenerated). https://github.com/knmcguire/best-of-robot-simulators

Bottoms-up TAM / SAM / SOM inputs

Used by TAM / SAM / SOM. These are account-count and category proxies — not substitutes for the bottoms-up ASP math on that page.

• PitchBook physical AI / robotics VC (2025) — $27.6B across 1,009 deals. Deal count ≠ unique buyers; capital is concentrating into mega-rounds. Crunchbase’s narrower robotics startup set (~$13.8B in 2025) is a category-breadth cross-check only. PitchBook via Mean CEO / Embodied Global — see Market Sizing above.

• RAISE Observatory robotics company sample — 5,056 companies / 1,279 startups (investment-age filters through mid-2024). RAISE Observatory, Chronicles of Robotics Investments. https://www.raiseliguria.it/wp-content/uploads/2025/04/RAISE-Observatory-Chronicles-of-Robotics-Investments.pdf

• Crunchbase robotics hub — ~9,765 organizations. Breadth warning only; includes OEMs and integrators outside Haga ICP. https://www.crunchbase.com/hub/robotics-companies

• IFR service-robot producers — 944 worldwide (excludes prototyping services and system integrators); 80% SMEs in sampled producers. IFR, World Robotics 2025 — Service Robots executive summary. https://ifr.org/img/worldrobotics/Executive_Summary_WR_2025_Service_Robots.pdf

• Pure-play world-model financings — 23 unique companies, 30 disclosed deals, $5.88B (Aug 2024–Jul 2026, equity ≥$300K). New Market Pitch, World Model Startup Funding (2026). https://newmarketpitch.com/blogs/news/world-model-funding-analysis

• Synthetic data generation — ~60–63 companies (~59 active as of Apr 2026); 34 funded; ~$447M cumulative VC/PE in Tracxn’s set. Tracxn, Synthetic Data Generation. https://tracxn.com/d/trending-business-models/startups-in-synthetic-data-generation/__KY2JOSQTguaaId7w-qIRnX6sBRXFJWxyZ5tiWnw5nlM

• China embodied-data players — ~97 in an incomplete domestic survey (Jul 2025–Jul 2026 window); 15 independent data-only providers raised ~¥4.47B in that year. Qbit via 36Kr. https://eu.36kr.com/en/p/3892027841362694

• Robot software market (cross-check only) — ~$24–35B around 2025–2026 depending on vendor scope (Mordor / Fortune Business Insights / IMARC diverge sharply on CAGR). Do not treat as Haga TAM. Mordor Intelligence robot software report; Fortune Business Insights robot software market page.

Note on Methodology

Market-size figures differ across providers (Mordor Intelligence vs. Fortune Business Insights vs. Fact.MR) because base years, segmentation, and forecast methodology differ between firms. Where two figures are given for ostensibly the same market, both are preserved rather than blended, consistent with how each source frames its own scope. Readers verifying these claims should expect the same variance any market-research consumer would.

Bottoms-up TAM/SAM/SOM deliberately does not take a percent of the report markets above; it multiplies a constructed buyer universe by working ASP. See TAM / SAM / SOM.

Last verified: July 2026. Funding amounts and company statuses change quickly in this space — reconfirm before citing.