Recurrent state-space world-model baseline under the shared clean-image boat drift benchmark.
This method uses an image encoder, recurrent latent dynamics, action-conditioned rollout, and decoded pose prediction under the same data, optimizer, and evaluation protocol as FlowMo.
