RealtimeWAM speeds robot action prediction while keeping success

When a robot acts on a prediction, has the scene already changed? World action models can take long enough to make that a problem for moving objects. The research asks whether their inference can be made faster while preserving task performance.

RealtimeWAM overlaps model work, selectively reuses observation features, and adjusts prediction refinement to reduce inference latency in existing robot models.

RealtimeWAM overlaps observation processing with action prediction, selectively reuses image features, and varies later refinement with predicted motion.

The authors compared it with each model’s original inference on the same four simulation settings, measuring request latency and whether each rollout met its task goal.

Across those settings, both models had much lower average latency and similar average success. This supports faster inference without a large change in average task success in the tested simulations.

The simulation averages cover four settings for two model architectures, while the real-robot results cover five tasks on one robot platform. They do not establish how RealtimeWAM would perform on other models, tasks, or robots.