Prediction enters the control loop, while execution remains the hard test

Prediction is becoming part of the policy state rather than a separate planning output. RoboTTT compresses up to 8K timesteps into fast weights without latency growing with context length. It reports 79% average completion on three bimanual assembly tasks, versus 42% for its single-step baseline. FoMoVLA instead predicts a future feature state and sparse point trajectories; it reaches 97.6% on LIBERO-Long with 9.4 ms median overhead. Lumo-2 provides a third formulation by aligning latent world dynamics with action, vision, and language, though its available evidence does not include numerical task margins.

Deployment work increasingly models stale observations and discontinuous actions as policy failures, not just systems overhead. Jetson-PI predicts the future representation corresponding to already committed actions and reaches 6.06 Hz on Jetson Orin, versus 0.70 Hz for naive π₀.₅. ChunkFlow trains overlapping action chunks to agree at their seams and reports 93.4% success on LIBERO-Long with 4.43 ms average reasoning latency. In driving, a slow 7B scene model paired with a fast action expert raises short-route completion from 37.0% to 94.0% by issuing fresh control every tick. Its long-route driving score remains only 2.96, showing that higher control frequency does not establish safety.

Two training studies target shortcuts that remain hidden by strong in-distribution scores. Semantic anchoring preserves pretrained task structure during robot fine-tuning; on a bimanual robot, it raises out-of-distribution success from 49.5% to 71.0%. AC-VLA decomposes demonstrations into reusable sub-tasks and suppresses wrist-view shortcuts during placement, improving real-world out-of-distribution success from 35.0% to 82.5%. IMBench shows why these interventions matter: leading vision-language models recognize constraints about 74% of the time, yet closed-loop GPT-5.5 succeeds on only 11.3% of tasks with vision-only input. The benchmark is simulation-only, so it establishes an evaluation gap rather than a universal capability ceiling.