A new method called Time-to-Goal (T2G) aims to help robots plan longer routes by giving lower costs to states that are closer in time to a goal. The researchers train it on offline trajectory data, without reward or policy labels, and use it in place of the usual Euclidean goal cost while keeping the world-model encoder and predictor frozen.
In results reported in the paper, T2G raised average planning success from 32.7% to 63.0% across four navigation and manipulation benchmarks, approaching or sometimes exceeding hierarchical JEPA baselines. The paper was submitted to arXiv on November 4, 2026.
Why it matters
For robotics researchers using flat world models, the reported results suggest a learned goal-cost metric could improve benchmark planning success without retraining the model's encoder or predictor; performance on real robots is not established here.
