EWAM: Enhanced World Action Model for Zero-Shot Closed-Loop Adaptation in Embodied Intelligence

Researchers have published EWAM (Enhanced World Action Model), a closed-loop online adaptation architecture built on a pretrained and fully frozen Cosmos3 backbone network. EWAM is focused on reducing deployment data required to adapt to new task layouts, with no fine-tuning performed on the backbone network. Performance gains stem from an inference-time co-reasoning mechanism composed of four neural layers: Neural Experience Memory Layer (DiT context), Neural Anomaly Detection Layer (state divergence monitoring), Neural Policy Routing Layer (execution/replanning/rollback selection), and Neural Action Correction Layer (action refinement).