Robot VLA work is measuring prediction, action timing, and safety under deployment pressure

World-model work is the main technical center of the day. RAW-Dream trains a VLA policy with reinforcement learning inside a task-agnostic video world model, then uses Qwen3-VL as a zero-shot reward judge. On LIBERO, its zero-shot world model raises average success to 52.3% against 43.4% for 1-shot supervised fine-tuning, using 10 target demonstrations and no target rollouts for world-model training.

OrbiSim takes a different route. It predicts explicit physical states and renders pixels from those states, so the learned simulator can expose gradients for policy and parameter optimization. On robosuite Push, it reports lower trajectory error than Vid2World and AdaWorld, and better long-horizon video metrics than its own version without dynamics-vision separation.

The WAM survey gives this activity a shared definition. It defines World Action Models (WAMs) as models that predict future observations and actions together, then separates cascaded and joint designs. CoME adds a memory angle for diffusion world models, using short-term, long-term, and spatial experts to keep generated futures consistent with prior observations.

GuidedVLA focuses on the action decoder itself. It assigns attention heads to object grounding, skill phase recognition, and depth-based geometry. The method keeps the pretrained policy stable at the start by adding a zero-initialized residual attention branch.

The reported gains are broad across perturbations. On LIBERO-Plus, the full model reaches 75.4% average success, compared with 68.2% for its π0 base. The ablation results also support the design choice: object, skill, and depth heads each improve success, while all heads together perform best. Real-world trials across six tasks show higher average success under in-domain, scene, and lighting settings.

Premover treats user input time as part of the control loop. It lets a frozen VLA policy begin acting before the full instruction is complete, using a focus map and readiness gate to avoid premature movement. On LIBERO, it reduces mean wall-clock time from 34.0s to 29.4s while keeping success nearly unchanged at 95.1% versus 95.0% for the full-prompt baseline. Naive early execution drops success to 66.4%, which makes the gate central to the result.

MindVLA-U1 brings the same deployment concern into autonomous driving. It uses one vision-language model backbone for language tokens, streaming memory, and continuous waypoint generation through flow matching. On WOD-E2E, it reports 8.20 RFS with two diffusion steps and about 16 FPS at roughly 1B parameters.

SafeManip adds temporal safety monitors to manipulation evaluation. It maps rollouts into symbolic predicate traces and checks Linear Temporal Logic over finite traces, covering collision, grasp stability, release stability, contamination, containment, and access rules.

The result is a clear warning for benchmark interpretation. Across 50 RoboCasa365 tasks and six VLA policies or variants, π0.5 raises task success over π0 from 8.1% to 9.3%, while its safety violation rate also rises from 69.7% to 82.8%. The paper identifies collision/contact and release stability as major failure sources, so final task completion is an incomplete measure for household manipulation.