Researchers have introduced DeltaWAM, a world-action model (WAM) for bimanual robot manipulation that rethinks how visual dynamics are predicted during training and inference. Existing WAMs transfer visual and motion priors from pretrained video generators by jointly modeling visual dynamics and actions, but they predict dense future frames — repeatedly modeling content that barely changes and entangling action-conditioned dynamics with irrelevant appearance variation.
DeltaWAM instead jointly predicts visual deltas and actions using three streams: a dense-anchor stream, a sparse-delta stream, and an action stream. The authors present three architectures that differ in how representations and computation are shared across these streams.
To address inference bottlenecks, the team developed Streaming Delta Memory (SDM), which updates cached anchor context with compact observed deltas, avoiding repeated processing by the heavy video expert that otherwise slows few-step action generation.
On the RoboTwin benchmark, DeltaWAM with SDM raised 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 three architectures reduced training FLOPs by 17.78–23.77%, while SDM cut one-step inference latency and FLOPs by 36.57% and 31.55% respectively. Real-world evaluations reportedly showed the highest overall success rate and normalized progress among the evaluated policies.
The work is described as in progress, with code being prepared at the project's GitHub repository and an accompanying website.