The path from your Elden Ring deaths to self-driving cars is shorter than you think.
The Summary
- British startup Rerun is converting human gameplay into training data for embodied AI — teaching robots and autonomous systems to navigate physical space by watching how humans solve virtual worlds
- Games provide structured environments where millions of people already demonstrate spatial reasoning, decision-making under uncertainty, and adaptive problem-solving at scale
- The approach sidesteps expensive human labeling and addresses the core challenge of embodied AI: translating digital intelligence into physical-world competence
The Signal
Rerun's thesis is simple but non-obvious. Language models learned from text. Image models learned from photos. Embodied AI needs to learn from actions in space, and gaming is where billions of hours of those actions already exist.
The startup is building tools to capture gameplay data — controller inputs, visual feeds, spatial decisions — and convert it into training datasets for models that need to understand navigation, manipulation, and physical causality. Not watching gameplay videos. Capturing the input-output loop: player sees obstacle, player decides action, world responds.
"Gaming generates more structured behavioral data about spatial reasoning than any other human activity at scale."
This matters because embodied AI is having a data problem. Self-driving cars need millions of miles. Warehouse robots need thousands of labeled pick-and-place sequences. Humanoid robots need even more varied interaction data. All of it is expensive, slow, and brittle. Gaming environments, by contrast, are:
- Already instrumented with perfect ground truth data
- Diverse enough to cover edge cases human labelers wouldn't think to create
- Constantly generating new scenarios as players explore
The technical bridge Rerun is building is the translation layer. Games run on physics engines that approximate real-world dynamics. The challenge is converting those approximations into training signal that transfers to actual robots operating in uncontrolled environments. Early results suggest it works better than expected for high-level spatial reasoning, worse for fine motor control.
Why this approach could scale:
- Gaming generates 3 billion hours of spatial problem-solving weekly, mostly unpaid
- Game engines are converging with robotics simulation platforms (Unity, Unreal already used for both)
- Synthetic data from games is orders of magnitude cheaper than real-world collection
The business model flips traditional data labeling. Instead of paying humans to generate training data, Rerun pays game developers for API access to player telemetry, then sells the processed datasets to AI labs and robotics companies. Developers get revenue from existing player activity. AI companies get behavioral data at gaming scale.
The Implication
If gameplay becomes a training corpus for embodied intelligence, two things follow. First, game design becomes AI infrastructure design. The games that generate the richest spatial and decision-making data become more valuable, not just as entertainment but as simulation environments for the next generation of AI.
Second, watch for gaming companies to start thinking about their player bases as distributed data generation networks. Your Cyberpunk playthrough might train a delivery robot. Your Minecraft builds might teach construction drones spatial planning. The value you create playing games stops being purely entertainment and starts contributing to the agent economy, whether you know it or not.