Researchers have a fix for robot brains that daydream without finishing the chore.
World Action Models (WAMs) work by predicting a visual future alongside the robot actions needed to get there. The problem: when these models get adapted for short-chunk control, they tend to generate the next plausible-looking moment rather than the moment that actually completes the task. Researchers call this task-incomplete imagination, and a new paper identifies it as a byproduct of how the models are adapted, not a flaw baked into the underlying world model. Their fix, called Completion Aware Guidance (CAG), is a training-free sampling method that steers generation toward completion rather than just plausibility. On a RoboTwin 2.0 benchmark subset, it pushed task success from 64% to 70%, and in zero-shot simulation from 69% to 75%, while cutting task-incomplete imagination from 79% down to 40%.
This matters because it is a cheap fix for an expensive problem. Training-free means no retraining, no new data, no extra compute budget for a fresh model run - just a different sampling strategy layered on top of existing WAMs. For a field where every robotics lab is racing to make prediction-based planning reliable enough to trust outside a lab, a drop-in method that meaningfully cuts failure rates without touching the model weights is the kind of unglamorous progress that actually ships.
Still, a 40% incomplete-imagination rate after the fix means robots guided by CAG will still stall out on the finish line nearly half the time. That is better than roughly 4 in 5 failures before, but it is a reminder that teaching a model to picture the right future is still a long way from teaching it to reliably deliver one.