DeltaWAM: Delta World Action Models for Bimanual Manipulation
DeltaWAM reports higher bimanual manipulation success while cutting training and inference cost.
The paper says the method avoids repeatedly predicting mostly unchanged full future frames by modeling visual deltas alongside actions. Its Streaming Delta Memory caches anchor context and updates it with compact observed changes, reducing the heavy video-expert work at inference. On RoboTwin, DeltaWAM with SDM improves average success over Fast-WAM from 81.3% to 85.4% in the clean setting and from 75.8% to 83.9% under visual randomization. The authors also report lower training FLOPs, lower one-step inference latency and FLOPs, and the best overall real-world success among the evaluated policies. HF Daily Papers' note
The paper says the method avoids repeatedly predicting mostly unchanged full future frames by modeling visual deltas alongside actions. Its Streaming Delta Memory caches anchor context and updates it with compact observed changes, reducing the heavy video-expert work at inference. On RoboTwin, DeltaWAM with SDM improves average success over Fast-WAM from 81.3% to 85.4% in the clean setting and from 75.8% to 83.9% under visual randomization. The authors also report lower training FLOPs, lower one-step inference latency and FLOPs, and the best overall real-world success among the evaluated policies. HF Daily Papers' note
score 5