Vision-language-action (VLA) models built on pretrained vision-language backbones achieve strong results on diverse robotic manipulation tasks. Directly mapping the current 2D observation to actions, however, often lacks spatial‑temporal understanding, limiting precision and long‑horizon performance. Recent works add geometric supervision and full‑scene future‑state prediction, but the extra scene information can distract the model from the geometry and dynamics relevant to the ongoing interaction. FOCAL-VLA addresses this by uniting subtask‑guided geometry distillation with implicit world modeling to capture the present spatial layout and future interaction dynamics. In the geometry distillation stage, geometric latent vectors from the teacher model VGGT are aligned with features extracted from image regions pertinent to the current subtask, transferring geometric knowledge to the VLA model. In the implicit world modeling stage, Track4World features from both current and future demonstration frames are used to encode the 3D evolution of the interaction. The two complementary representations jointly steer action generation at inference time without invoking VGGT or Track4World. Experiments on simulation benchmarks and real‑world manipulation tasks show consistent gains over strong baselines. Project page: https://zhiyuan-gao.github.io/FOCAL-VLA/.
Review