GlanceWAM Moves Visual Foresight Off the Control-Critical Path, Cutting Action Decoding to 48 Milliseconds per Chunk

The research team released GlanceWAM on August 25 in an attempt to resolve the latency tradeoff in world-action models: using a video diffusion model to generate future frames during every control cycle blocks the robot during sampling, while eliminating test-time foresight altogether removes explicit visual subgoals. The new method separates two clocks within the same 1.3B-parameter SkyReels-V2-DF video DiT. A background proposer predicts a future state roughly every three seconds, while the action head produces a 0.8-second action chunk containing 16 control steps.

The key is that the predicted future is not decoded into RGB. Instead, it remains in the video VAE’s latent space and directly conditions subsequent action-head outputs. The model uses three types of prefix-LM attention masks to prevent clean foresight tokens from contaminating observation and video-generation representations. During training, it also randomly samples a temporal offset from the full foresight horizon and explicitly passes the remaining offset to the action head, allowing the policy to adapt as the same foresight representation gradually becomes stale. The authors also drop foresight tokens with a 10% probability, reducing vulnerability when predictions are missing or inaccurate.

The paper reports a 72.2% success rate across 24 RoboCasa kitchen tasks, exceeding the synchronous Cosmos Policy baseline at 67.1% and a variant without test-time foresight at 64.4%. Its LIBERO score is 99.0%. On a single A100, action decoding takes 48 milliseconds per chunk, which the authors claim is 24× faster than the synchronous baseline. The code, data, and checkpoints have been released under the MIT License, with the reproduction package totaling approximately 21 GB. The repository also notes that RoboCasa does not fix the policy random seed, resulting in rerun variance of approximately ±0.02.

The 48-millisecond figure covers only the action-critical path; background diffusion still consumes compute, and deployment may require parallel workers or an additional GPU. All evaluations were conducted in simulators, so the system’s safety on physical robots under network latency, sensor noise, and sudden environmental changes remains unproven. The next step should be to compare total power consumption and hardware requirements, including the background proposer, and test whether the system can fall back quickly when its foresight is severely inaccurate.