Overview
- This week, July 7–8, 2026, research teams released a concentrated set of arXiv preprints that shift VLA work from monolithic fine‑tuning to modular, deployment‑aware designs intended to close the lab‑to‑field gap.
- CamVLA introduces a calibration‑free approach that predicts end‑effector motions in the camera frame plus a 6‑DoF hand‑eye transform so a single monocular RGB view can drive robots without explicit camera extrinsics or depth.
- SVA (Search, Value, Act) separates action proposal from action evaluation by using tree search in simulation to label trajectories and distilling those returns into a lightweight Q evaluator, which lets smaller frozen VLAs match or beat larger models at lower latency.
- Other papers target complementary problems: CoRE‑VLA routes modality‑ and task‑specialized experts to handle missing sensors, Lift3D‑VLA adds explicit 3D point‑cloud and temporal modeling, CAC‑VLA and InternVLA‑A1.5 strengthen latent action and foresight interfaces, XS‑VLA compresses spatial knowledge for edge use, and Cortex provides canonical skill primitives for long‑horizon planning.
- Teams report substantial gains on benchmarks such as LIBERO, MetaWorld, and RLBench and show some matched real‑robot tests, but authors note that head‑to‑head real‑world comparisons and standardized large‑scale benchmarks remain needed before field reliability can be judged.