GlanceWAM separates imagination and control in a single video DiT for robots

Researchers presented GlanceWAM, an approach to world-action models that moves video generation off the critical path of control. An asynchronous proposer running at a slow cadence predicts a single frame several seconds ahead in the background, while an action head decodes actions at a control frequency of 48 ms without blocking. On the RoboCasa benchmark, the method achieves 72.2% success, outperforming the synchronous Cosmos Policy (67.1%) and learning without imagination (64.4%). Details are in the paper on arXiv.

What it means

This news concerns robotics, not video generators for users. In our video generators section, there are no direct analogs to GlanceWAM: it is not a text-to-video or image-to-video tool, but a robot control method where video is used as an internal world representation. A comparison with our fact sheets shows the difference in tasks. LTX-Video handles image-to-video and runs on consumer hardware from 1 GB of VRAM. CogVideoX is text-to-video, requires CUDA, and is aimed at generation from a description. Pyramid Flow is text-to-video with support for Multi-GPU and CPU offloading. All three generate video as a final product for humans, whereas GlanceWAM consumes video in latent space to select robot actions. The technical idea — asynchronous generation of a single frame instead of a synchronous video stream — may be of interest to those tracking latency reductions in video models, but our tool fact sheets contain no such mechanisms. The developers do not indicate whether the approach is applicable outside robotics.