What it is
A world-action model for mobile manipulation that aligns video prediction, action structure, and rollout training via latent actions, disentangled transformers, and dream forcing. Unlike prior WAMs that use coarse video chunks, entangled nav+manip actions, and GT-conditioned inverse dynamics, it inserts frame-level latent actions, disentangles base/arm control with dual-level MoT, and trains on self-predicted videos via dream forcing.
Key results
- Claims SOTA on long-horizon mobile manipulation success rate benchmarks
- Claims SOTA fine-grained control accuracy on manipulation benchmarks
- Real-robot demos succeed on tasks like Find Toaster and Arrange Flower
Mobile manipulation benefits from aligning temporal granularity, disentangling heterogeneous actions, and matching training to autoregressive rollout conditions.