Fictitious-world training improves robot planning in simulation

Researchers introduced a benchmark built around invented physical rules to test whether AI models can learn an unfamiliar operating environment. After continued training on its 342,069-token rule set and a separate stage for planning skills, small offline models outperformed GPT-4.1 with access to the same rules in a simulated robot task. The task required models to order objects for safe handling in familiar and novel scenes. The trained models scored 0.848 on the reported ranking measure, compared with 0.606 for GPT-4.1. The researchers found that answering questions about the rules alone did not reliably produce executable plans; the robot demonstration used a simulated arm with human correction.