We extract millions of multi-modal, action-conditioned trajectories from video games — safe and controlled environments where every interaction is captured — and deliver them as training-ready datasets for the next generation of AI: world models and embodied AI.
Temporally aligned video, telemetry inputs, and ground-truth 3D state — all synchronized and training-ready.
Three properties decide whether a dataset can train a frontier world model. We engineer for all three at once.
Petabytes of synchronized capture and billions of ticks of gameplay. Today's largest public datasets contain tens of thousands of hours. We are building toward one million hours.
Hundreds of titles, thousands of maps, physical environments, lighting conditions, and edge-case scenarios. Reduce the risk of overfitting and train your model across environments.
Zero-loss engine output under formal licensing agreements, not scraped from the web. No hallucinations and no alignment drift — exact temporal alignment on every frame. Every frame is the same source of truth the game itself runs on.
High-tick, physics-rich, multi-agent environments with skilled players generating diverse, intent-driven trajectories every hour.
200+ structured properties extracted per frame. Every chip below is a real column in your dataset.
Framework-agreement customers can request additional AAA titles, custom property extraction, or live engine instrumentation for novel research environments.
Generic video corpora teach a model what scenes look like. Our data teaches a model what a world is: a coupled system of state, motion, intent, and consequence, frame by frame, provably aligned.
Straight from the engine, not inferred from pixels: exact per-object position, rotation, and velocity for every entity in the scene. No perception noise, no labeling error, no occlusion guesswork. Your model learns the true geometry of the world instead of a noisy reconstruction of it.
The next state depends on the action taken — and that link is the hardest thing to source. We record raw controller and keystroke inputs aligned to the exact frame and state they produced. We don't infer it. We record it and align it to each single frame.
Hundreds of ground-truth properties, multiple ticks per frame. Not just highlights — the full trajectory. A dataset of a consistent world evolving rather than isolated moments.
In 2025, we published a position paper discussing game-generated data as an untapped resource for advanced AI training, and we are thrilled by the interest it has received from the research community. A huge thank you to everyone who has shared feedback and ideas so far, and we continue to push the boundaries and pursue research collaborations with academic partners.
Please reach out if you are interested in what we are building. We look forward to collaborating with researchers working on world models, physical AI, and self-supervised learning.
The paper explores how game-generated data can open up new possibilities for advancing AI training. In particular, we dived into the Joint-Embedding-Predictive-Architecture (JEPA), a state-of-the-art architecture for world model building, with a vision for achieving human-level intelligence.
We do not ship raw video clips. We ship a coherent dataset: time-aligned video, telemetry, structured game state, perfect action labels, and controlled environments — all generated by the engine itself, all immediately usable for training. Our team combines deep expertise across both gaming and AI, and we curate the best dataset for your specific needs.
1080p multi-view game capture, every frame timestamped to engine ticks.
Full positional, velocity, orientation, and physics vectors for every entity per tick. Semantic context, events, HP, inventory, weapon state, score, camera pose, depth — over 200+ annotations per tick.
Every stream synchronized from tick to frame. Precise alignment guarantees no drift.
Labeled controller, mouse, and keyboard inputs aligned to frames — recorded, not inferred.
Bespoke scenarios and gymnasiums built from commercial titles with rich physics.
Player intention and tasks, event annotations, human-annotated planning of goals and sub-goal achievements, causal relationships.
Delivered in Parquet, HDF5, and JSON-L by default. Custom export pipelines available for framework-agreement customers.
The same capture stack we ship to labs, miniaturized into the browser. Run an agent (you), generate frame-aligned ground-truth state, watch a world model style browser learner predict the next frame in real time.
A 60-second mini-game running on a live capture pipeline. Frame-aligned telemetry, ground-truth 3D state, and action labels stream into a world model style browser learner preview as you play.
For pricing inquiries and commercial contracts, send your request below and our team will be in touch.
Company announcements, research notes, and dataset releases. Newest first.
Research partnership seeks to establish a predictable commercial roadmap for next-gen AI that understands the physical world.
Cambridge, 4 September 2026
Worldmodeldata, a Cambridge-based startup building the world's largest database of video game-generated training data for next-generation AI, has announced a strategic partnership with Mila, one of the world's leading artificial intelligence research institutes. The collaboration will seek to advance the fundamental science behind next-generation AI by establishing whether a data scaling law exists for world models, AI that understands how the physical world works.
In January 2020, OpenAI's landmark paper 'Scaling Laws for Neural Language Models' proved that LLMs follow a reliable pattern of improved performance as you scale up training data, compute and model size, making scaling a viable strategy for improving AI. This discovery laid the groundwork for record-breaking, industry-wide investment and growth.
No equivalent, universally established data scaling law exists for world models today. Building upon Worldmodeldata's pioneering 2025 research highlighting video games as a vast, cost-effective source of action-conditioned training data, this strategic partnership with Mila will strive to answer these fundamental questions in embodied AI and world models. Commencing in September 2026, the research will be conducted over a 12-month period.
Under this joint initiative, Worldmodeldata will supply extensive data resources to empower Mila's research teams as they investigate key open questions in AI scaling and domain transfer. To establish data scaling laws for world models, the researchers will measure how effectively embodied agents acquire capabilities in virtual environments and define precise metrics for transferring synthetic experience into real-world performance.
“The importance of scaling laws is not purely academic: they give the industry a predictable roadmap, showing that investment in larger and better datasets, and model as well as compute, can systematically improve model performance, thus justifying the billions of dollars being invested in the industry. If embodied AI systems can learn from unlimited virtual environments, transfer to real-world performance, and follow a similar scaling law, such a rule would be highly significant and could become a 'GPT moment' for the world model industry. Demonstrating that relationship is therefore our next major milestone.”
“Mila's mission is to push the fundamental science of AI forward for the benefit of all. Understanding how embodied AI systems learn from experience, and whether that learning transfers from virtual environments to real-world performance, is central to that mandate. Partnering with Worldmodeldata lets our researchers test one of the field's most consequential open questions with the scale and quality of data this research requires. We're proud to help build the evidence base the industry needs to move forward responsibly.”
Worldmodeldata delivers the most diverse, scalable, and high-quality training data to accelerate the next generation of AI - world models, embodied agents, and physical AI. We close an urgent data gap by turning modern video games into a source of rich, structured training data captured directly from the engine with humans in the loop.
Founded by Professor Yoshua Bengio, Mila – Quebec Artificial Intelligence Institute is one of the world's leading artificial intelligence research institutes. Based in Montreal, Mila was created out of a unique partnership between Université de Montréal and McGill University, dedicated to advancing scientific breakthroughs that drive innovation and ensure AI benefits everyone. A non-profit organization, Mila is strongly supported by the Government of Canada and by the Government of Quebec. Internationally recognized for its influential scientific contributions, global innovation partnerships, and leadership in multilateral efforts on responsible AI, Mila continues to shape the future of AI worldwide. For more information, visit mila.quebec.
British startup emerges from stealth and seeks to generate 1 million hours of training data by the end of 2026.
Cambridge, July 2026
Worldmodeldata, a Cambridge-based startup building the world's largest database of video game-generated training data for next-generation AI, has raised £7m in seed funding as it emerges from stealth. The round was led by Iona Star Capital, a London-based venture capital fund focused on early-stage companies operating at the intersection of AI, data and technology.
The company was founded by serial entrepreneur Rhea Loucas and is supported by Lord Richard Allan, UK technology policy specialist and Meta's former VP of Public Policy, who joins the board as Chairman.
World models, an AI system's internal understanding of how the world works, will form the backbone of the next generation of AI. Rather than simply reacting to inputs, world models learn how things look, interact, and change over time. This allows them to predict what will happen next and plan actions accordingly to operate safely in complex environments.
However, these models are only as capable and robust as the data they are trained on. Unlike generative models, which had a head start benefiting from the internet, the data world models need is even harder to find. This data scarcity represents a critical limitation for developing AI capabilities in high-stakes industries where trial-and-error training is not an option.
Worldmodeldata overcomes this bottleneck by aggregating and structuring rich datasets from modern video games that capture real human behaviour and interactions in complex, dynamic environments. Delivered as curated datasets, it gives customers, such as frontier labs building world models, physical AI systems, and robotic companies, a powerful foundation for training models that need to understand dynamics, predict outcomes, and make safer decisions in the real world. For instance, enabling world models to act as an internal simulator for self-driving cars as they navigate traffic and predict pedestrian movement.
This data is sourced directly from real gameplay in titles built on engines such as Unreal and Unity. It is acquired via formal licensing agreements that allow the gaming community and developers to monetise their gameplay and assets built with Worldmodeldata, rather than using web scrapers.
The company aims to build a data library of 1 million hours by the end of 2026, compared to the current largest database, which has just 40,000 hours. The funding will help advance this goal by fuelling product development, team expansion, and the securing of crucial data-sourcing agreements.
“World models represent a fundamental paradigm shift in AI, but progress like this needs fuel: internet-scale data to help AI systems make predictions and reason in physical environments. Such a comprehensive dataset does not yet exist; however, video games, as safe, controlled environments, are the perfect setting for generating the action-conditioned data needed to train the next generation of AI at the required scale. Worldmodeldata is built to bridge that gap.”
“We are proud to anchor ourselves in the UK's AI ecosystem, a strategic choice driven by the urgent push for sovereign AI capabilities and the robust infrastructure that powers them. The deep expertise of this team is hugely exciting and positions them perfectly to lead this charge. In tackling critical gaps that LLMs cannot bridge, this isn't about improving AI model training, but building an essential foundation for deploying AI in sectors where the demand is vast but the solutions remain limited.”
“AI spent the last few years learning to describe the world, constrained primarily by compute and architecture. The next era is about acting in it, and you cannot learn to act from passive video or text, because acting requires having seen how the world answers back. That coupling between action and consequence is the scarcest resource in AI today. Worldmodeldata is manufacturing that missing ingredient, and every company building Physical AI or Digital AI will eventually face the same challenge. That conviction is what led IonaStar to lead the investment in Worldmodeldata, and that is why I joined the board.”
Worldmodeldata is a Cambridge-based company building the world's largest database of video game-generated training data for next-generation AI. The company delivers curated, action-conditioned datasets to frontier labs, physical AI systems, and robotics companies. For more information, visit worldmodeldata.com. Media inquiries: worldmodeldata@fieldhouseassociates.com.
We envision a world where next-generation AI, deeply grounded in reality and aligned with human values, improves life everywhere. The new AI paradigm transcends static pattern recognition and truly understands the dynamic mechanics of the world.
Our vision is to be the trusted data layer and intelligence backbone that powers this transition — enabling AI systems to operate safely, predictably, and intelligently within the complex real world.
We’re hiring across research, engineering, and data. Send a note with your CV and what you want to build.
EMAIL_JOBS ▸jobs@worldmodeldata.com
For press, interviews, and media requests, contact our communications team at Fieldhouse Associates.
EMAIL_PRESS ▸worldmodeldata@fieldhouseassociates.com
The people behind the data layer.
Worldmodeldata is built by a team that sits at the exact intersection the problem demands. We bring together serial entrepreneurs who have built and scaled companies, executives from the game industry who know how these worlds are made and how to access them at scale, world model specialists and ML engineers who understand precisely what the next generation of AI needs to learn from, and senior voices in AI and technology regulation who ensure that data is sourced the right way.
We use essential cookies to run the site and an optional functional cookie to remember your AI assistant conversation between visits. No analytics, no tracking, no ad networks. Change your choice any time from the footer.