Robot VLA progress is being judged by execution under real control pressure

World models are now used as training spaces and control priors, with clear pressure to reduce task-specific robot data. HarmoWAM combines a video world model with predictive and reactive action experts, then routes between transit and interaction phases. Its reported out-of-domain gains are 33 percentage points over prior VLA models and 29 points over prior World Action Models (WAMs) across six real-world tasks.

RAW-Dream gives the week’s clearest data-efficiency claim. It trains OpenVLA-OFT with reinforcement learning inside a task-agnostic video world model and uses Qwen3-VL for reward judgments. On LIBERO, it raises a 1-shot supervised fine-tuning baseline from 43.4% to 52.3% with 10 target demonstrations and no target rollouts for world-model training; with in-domain world-model tuning, it reaches 66.0%.

Several papers target the exact moments and scene features that decide whether a VLA policy succeeds. GuidedVLA assigns action-decoder attention heads to object grounding, skill phase recognition, and depth-based geometry. On LIBERO-Plus, the full model reaches 75.4% average success, above its π0 base at 68.2%, and its real-world trials improve over the base policy across in-domain, scene-change, and lighting-change settings.

FrameSkip treats dense demonstrations as uneven supervision. It keeps 20% of unique frames, protects gripper transitions and high-action-change frames, and raises the macro average across RoboCasa-GR1, SimplerEnv, and LIBERO from 66.50% to 76.15%. Evo-Depth and AffordVLA add complementary spatial signals: RGB-derived depth for placement and grasping, and training-time affordance alignment for contact regions without an added inference module.

Long execution is treated as a control problem with shared state and protected task knowledge. Pelican-Unified trains one embodied model to reason in language, predict future video, and output action chunks through a common latent state. It reports 93.5% average success on the 50-task RoboTwin dual-arm benchmark and an EWM Score of 66.03 on WorldArena.

DyGRO-VLA focuses on reinforcement learning after imitation training. It freezes a base VLA and learns routed residual experts that correct action chunks, reducing damage to shared representations during multi-task tuning. On LIBERO it reports 97.1% average success, and on LIBERO-Long it adds 9.8 percentage points over its offline base.

Evaluation is adding checks for what happens during a rollout, not only the final task state. SafeManip defines temporal safety properties with finite-trace Linear Temporal Logic monitors over predicates such as collision, stable grasp, containment, sanitation, and fixture access. In 50 RoboCasa365 tasks, π0.5 improves task success over π0 from 8.1% to 9.3%, while its safety violation rate rises from 69.7% to 82.8%.

The same concern appears in the week’s contact-aware VLA work. AffordVLA trains policies to attend to functional contact regions, and the May 17 trend note highlights behavior-level checks as part of VLA interpretability and safety. The evidence points to a stricter deployment bar: success rates need to be paired with contact, order, recovery, and unsafe-state exposure measurements.