Robot VLA papers are optimizing for deployable foresight

World-model work centers on smaller latent state and better planning signals. OneWM-VLA compresses each camera view and frame into one semantic token, then generates future latent tokens and action chunks together. It reports MetaWorld average success of 61.3% against 47.9% for π0, 98.1% average success on LIBERO, and 71.7% real Piper-arm success under clean conditions against 50.0% for π0.

RLA-WM uses Residual Latent Action (RLA), a compact code for DINO feature changes. It predicts future visual features with 3.5T FLOPs per inference and beats listed feature and flow baselines on ManiSkill and IWS prediction metrics. RC-aux adds reachability supervision to latent world models, showing that accurate short-horizon prediction can still mislead a planner. On Wall, it raises success to 83.6 ± 3.6 compared with 50.4 ± 6.5 for the LeWM control.

Several papers add feedback channels that only appear during execution. AT-VLA adds a tactile gate and a dual stream: slower vision-language reasoning plus faster tactile correction. The paper reports closed-loop tactile response within 0.04 s. In real-robot contact-rich tasks, AT-VLA improves Unzip Bag to 0.33 success against 0.20 for GO-1 and 0.00 for π0.5, and Wipe Vase to 0.67 against 0.07 and 0.33.

AFIL uses failed rollouts as negative guidance for diffusion- and flow-based VLA policies. It trains separate success and failure action generators while sharing one vision-language backbone. The available summary gives no table values, so the grounded point is architectural: failure data becomes an online training signal and a sampling-time repulsion term.

The training papers focus on data that is private, sparse, or domain-specific. ForgeVLA trains on distributed vision-action logs without centralizing raw data or adding manual language labels. Each client assigns pseudo instructions locally, while a contrastive planning loss and server aggregation handle non-i.i.d. robot data. On LIBERO-Goal, it reaches 55.2% success and 100% Pass@50 compared with 28.8% and 80% for FedAvg.

ACA addresses small real-robot adaptation budgets by repeating demonstrations at selected anchor conditions, then collecting boundary data where the policy deviates. With 100 trajectories on a Franka Panda setup, π0.5 plus ACA reaches 72.5% mean success against 31.7% for π0.5. BioProVLA-Agent applies the same deployment pressure in wet labs: it parses biology protocols into verified subtasks, uses visual checks before and after actions, and runs on a reported 800–850 USD hardware platform, though the excerpt gives no exact success rates.

LaWM and Sword make rollout quality a control requirement. LaWM defines the next latent state through a learned least-action objective and an unrolled solver. Its strongest numbers are on controlled physical motion: PIS-vx reaches 0.9938 ± 0.0045, and acceleration consistency PIS-ax reaches 0.8964 ± 0.0275 against 0.6568 ± 0.013 for NewtonGen.

Sword trains an action-conditioned world model for LIBERO rollouts under visual style changes. Dynamic Latent Bootstrapping caches predicted VAE latents and gradually feeds them into training, cutting context-frame storage to under 20 GB. On LIBERO-Mixed, Sword reports FID 32.59 and FVD 111.19 against 119.62 and 198.84 for WoVR.