Time-to-Goal: Learned Temporal Metrics for Long-Horizon Planning with JEPA World Models
Time-to-Goal: Learned Temporal Metrics for Long-Horizon Planning with JEPA World Models
This research paper presents Time-to-Goal (T2G), a method that learns temporal ordering metrics for goal costs in Joint-Embedding Predictive Architectures (JEPAs) to improve long-horizon planning. By training on latent triples from offline trajectories and using a margin ranking loss, T2G enhances temporal contrast without requiring rewards or policy labels. Evaluated across four navigation and manipulation benchmarks, the approach raises average planning success rates from 32.7% to 63.0%, outperforming flat LeWorldModel baselines and approaching hierarchical H-JEPA performance.
The study introduces Time-to-Goal (T2G), a lightweight metric trained on latent triples from offline trajectories to enforce temporal ordering in Joint-Embedding Predictive Architecture (JEPA) world models, improving planning success rates from 32.7% to 63.0% across four benchmarks without modifying the frozen encoder or predictor.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ves the average planning success rate from 32.7\% (flat LeWM) to 63.0\% and approaches or even surpasses the performance of the hierarchical H-JEPA baselines. These results show that learning temporal ordering for the goal cost can substantially improve latent planning even with a flat world model.
Open source S5
Source：arXiv Robotics — research abstracts · arxiv.org