Embodied models are judged by resilience and reusable visual understanding

STRONG-VLA centers this day on failure tolerance in embodied control. The paper argues that robustness training should be split into two steps: first learn under perturbations, then recover clean-task fidelity. The evidence is concrete. On LIBERO, gains reach +12.60% seen and +7.77% unseen for OpenVLA, +14.48% and +13.81% for OpenVLA-OFT, and +16.49% and +5.58% for pi0. Clean performance stays close to baseline. The benchmark also matters. It covers 28 perturbation types across text and vision, including held-out tests such as semantic drift and dynamic visual artifacts. This keeps the paper tied to deployment problems like occlusion, instruction corruption, and sensor noise, not just synthetic stress tests.

Zero-shot World Model, or ZWM, gives the day a second center of gravity: broad visual competence from sparse natural video. The setup is specific. The model sees one frame fully, about 10% of the next frame, and learns to predict the rest. At test time, small input interventions are used to read out flow, depth, segmentation, and simple physical reasoning without task-specific training. The paper claims this works with child egocentric video alone: 868 hours from 34 children for BabyZWM, and even a 132-hour single-child version stays close on most tasks. The strongest result is breadth. The same predictor is reported as competitive on TAP-Vid-DAVIS flow, above 90% on UniQA-3D depth, strong on SpelkeBench segmentation, and near 100% on the paper's short-timescale physics benchmark.