Robots learn from their own failures — not human demonstrations — to build better world models

A new system called PlayWorld trains robots to predict physical interactions by letting them play unsupervised instead of learning from human-collected examples. This means robot simulators can now learn from messy, realistic data at scale, including the long-tail failures that actually matter in manipulation tasks.

Robot world models have been stuck on a problem: human-collected data is clean but limited, missing the weird edge cases where physics actually breaks down. PlayWorld solves this by letting robots generate their own training data through autonomous play, capturing the rare, complex interactions that make the difference between a simulator that works in the lab and one that works in reality. The paper shows 40% better prediction accuracy on contact-rich tasks and 65% better real-world policy performance than models trained on human data. This matters because scalable data collection has been the bottleneck in robot learning — you can't hire enough humans to demonstrate every manipulation task, but you can let robots play unsupervised.