Robotics papers judge models by how well training signals survive contact with execution

EmbodiedMidtrain makes data selection a first-class part of VLA training. The paper measures a real mismatch between generic vision-language model data and robot trajectories, then mid-trains on samples scored as closer to robot data. The gains are large for small backbones: InternVL3.5-1B rises from 36.5 to 56.3 success on SimplerEnv-Bridge and from 39.0 to 54.2 on Libero-10. Qwen3VL-2B also improves across Calvin, SimplerEnv-Bridge, and Libero-10. This gives the day a concrete message: better robot policies are coming from better pre-action data alignment, not only larger action heads or more robot fine-tuning.

World-model papers in this period care more about executable structure than photorealistic prediction. Mask World Model trains on future semantic masks, which keeps object layout and contact cues while dropping texture noise; it reports 98.3% average success on LIBERO and 68.3% on RLBench, ahead of several strong baselines. RoboWM-Bench tests the opposite side of the story: generated videos may look plausible yet still fail when converted to actions. Its robot evaluations stay low even for strong generators, and step-level analysis shows many systems can reach contact but fail to complete the task sequence. Taken together, the signal is clear: a useful world model for robotics is judged by execution fidelity.

Two papers widen the training scope around robot control. UniT introduces a shared discrete action language for human and humanoid behavior, with policy and world-model training built on the same tokens. The paper uses 27,419 human trajectories, few-shot robot data, and real-humanoid tests, though the excerpt does not expose the main margins. VLA Foundry tackles a different bottleneck: it puts language pretraining, vision-language training, and action training in one stack, with support for distributed runs up to 128 GPUs and open model releases. The common theme is infrastructure for transfer. One line tries to reuse human motion across bodies; the other makes full-pipeline VLA experiments easier to reproduce and compare.