Robot world models now have to prove control value under latency and data limits

World-action models are being evaluated as control systems, not only as predictors. MotuBrain joins future video latents and action tokens in one diffusion model, then cuts reported end-to-end latency from 4.90 seconds to 0.09 seconds. Its RoboTwin 2.0 scores stay above 95% in both clean and randomized settings, which makes the latency result central to the claim.

Being-H0.7 takes a lighter route. It trains latent tokens with future observations, then removes the future-aware branch at inference. The policy keeps the action-oriented latent state and avoids test-time video rollout, with reported deployment at 3–4 ms per step. The survey paper gives the common definition behind these systems: a robot world model predicts future states or observations conditioned on current state, actions, and optional language.

Vision-Language-Action (VLA) policies in this set add internal signals for physical progress before emitting actions. LaST-R1 uses reinforcement learning (RL) to update both latent reasoning embeddings and action tokens. The reported result is 99.9% average success on LIBERO after one demonstration per task, plus up to a 22.5% improvement over supervised fine-tuning in real robot deployments.

PRTS adds a different progress signal. It trains a 4B VLA model to score whether a state-action pair can reach a language goal, using contrastive learning over offline trajectories. The excerpt does not provide success-rate tables, but it does ground the scale of the run: 167B tokens, 64 H100 GPUs for one week, and evaluations across LIBERO-family benchmarks, SimplerEnv, and 14 real-world manipulation tasks.

Lucid-XR treats extended reality (XR) as a data collection interface for robot manipulation. Demonstrators act in a browser-based MuJoCo simulator running on an XR headset. The system records motion, retargets it to robot hands or grippers, and turns simple virtual scenes into realistic multi-view training images with a generative image pipeline.

The reported numbers point to data throughput and visual coverage. The simulator runs at 90 fps on Apple Vision Pro and records at 25 fps. In 30-minute sessions across three tasks, users collected about twice as many demonstrations as real teleoperation. With augmentation, the effective dataset reached about five times the real teleoperation baseline. On kitchen clearing with unseen meshes, ACT plus LucidSim retained 90% success in low clutter where ACT alone fell to 0%.

Two survey papers give this period a stronger taxonomy layer. The robot world-model survey connects policy learning, learned simulators, controllable video prediction, datasets, and benchmarks under action-conditioned prediction. It names three core capabilities for actionable models: foresight, planning through imagined outcomes, and data amplification.

The graph world-model survey narrows the lens to entities and relations. It defines graph world models with structural abstraction and relational transition operations, then groups prior work by spatial, physical, and logical biases. This matters for manipulation and navigation because many failures come from missed object relations, contact relations, or long-horizon state updates, not only weak visual features.