Fast-LeWM Cuts Robot Planning Time by 48%

The existing approach for JEPA-based Visual World Models uses an autoregressive loop that sequentially predicts each state one by one when evaluating candidate action sequences, which has the drawback of very high computational cost.

The Fast LeWorldModel (Fast-LeWM) proposed in a new paper replaces this sequential rollout with Action-prefix prediction. This method encodes the action prefix to predict future latent states in parallel rather than step by step.

The key performance metrics are as follows:

• Reduced planning time: CEM planning solve time decreased from 54.4 seconds to 28.3 seconds, a reduction of about 48%.
• Model call efficiency: The number of model calls per planning cycle dropped significantly from 55 to 11.
• Improved accuracy: Across 4 simulation environments, the average task success rate rose from 85.8% to 90.5%, reaching 92.0% when the self-consistency mechanism was applied.
• Parameter efficiency: The model size is about 17.9 million parameters, nearly identical to the existing LeWM (18 million), showing that this is an algorithmic upgrade that can be adopted immediately without large-scale retraining.

This performance improvement is analyzed as a result of reducing the phenomenon where single-step errors accumulate during long-term prediction.