DreamWAM: Beyond RGB Future Prediction for World Action Models

DreamWAM argues that World Action Models should predict action-relevant future states beyond RGB, since texture, illumination, background, and viewpoint changes can obscure useful transitions. It represents the future through complementary appearance, motion, geometry, and semantic signals. Training combines joint latent denoising of RGB and motion with gated residual branches for geometry and semantics, while inference remains RGB-only because the auxiliary supervision branches are disabled. On LIBERO, success improves from 97.30% to 98.40% without rollout and from 98.00% to 98.90% with joint video-action inference. Under unseen LIBERO-Plus perturbations, gains are larger, and real-world manipulation reaches 74.4% average success versus 55.6% for Fast-WAM-Joint.