Frame rate is not the same as action response
The CADRE paper, first posted on October 21, 2026, revisits the use of video-generation autoencoders inside interactive world models. Compressing several frames into one temporal block saves computation, but can make an action wait for that block before its effect appears. High throughput therefore does not automatically mean immediate control. The researchers compare causal autoencoders with different spatial and temporal compression settings.
Different cameras need different compromises
They train the same world-model architecture across four manipulation and navigation environments. The authors report that temporal compression is relatively less costly for manipulation with a still camera, whereas spatial compression is more attractive when the camera moves. Their study argues for evaluating motion and action following alongside image quality, rather than selecting a representation by reconstruction sharpness alone.
These are environment-specific research results, not a general-purpose simulator release. Transfer to generalist models, policy training and planning remains unverified, and the paper should be identified as a preprint without established peer-review acceptance. A practical evaluation sheet should separate visual quality, throughput, input-to-effect latency and action-following failures, while recording camera conditions. That would expose trade-offs a single frame-rate number can hide.