Event

RealtimeWAM: A World Action Model with One-Step Generation and Asynchronous Inference

1 reports · 1 sources

Chengtao Lv and other researchers propose RealtimeWAM, an efficient variant of a world action model (WAM) that uses one-step action generation and asynchronous inference to address two major efficiency bottlenecks: multi-step denoising and serial waiting among experts. The paper reports that experiments on multiple benchmarks including LIBERO, LIBERO-Plus, and RoboTino, as well as model variants such as Fast-WAM and Faster-WAM, show that RealtimeWAM has a performance drop of less than 1% while achieving about 25x end-to-end speedup on H100; the code and checkpoints have been made public.