Ctrl-World: A Controllable Generative World Model for Robot Manipulation

Researchers have developed Ctrl-World, a controllable generative world model that enables robot policies to be evaluated and improved through simulation rather than costly real-world testing. The model, trained on 95k trajectories, can generate consistent 20+ second simulations and improved policy success rates by 44.7% through synthetic data generation.

• →Ctrl-World enables robot policy evaluation without expensive real-world rollouts by using imagination space simulations.
• →The model supports multi-view prediction and maintains long-horizon consistency for over 20 seconds of simulation.
• →Training on the DROID dataset with 95k trajectories across 564 scenes provides robust generalization capabilities.
• →The approach can accurately rank policy performance without physical robot testing.
• →Synthetic trajectory generation for supervised fine-tuning improved policy success rates by 44.7%.