Modern vision‑language‑action (VLA) policies output an entire action chunk—typically one to two seconds of coordinated motion—in a single forward pass. An action chunk is essentially a short multivariate trajectory, yet current models treat it as a sequence of generic per‑timestep hidden tokens decoded by a linear head. This design overlooks two motion structures. First, frequency: a chunk combines a smooth global trend with fine‑grained corrective motions across time scales, and a single token entangles both. Second, cross‑phase geometry: different phases (reach, contact, grasp adjustment, settling) evolve along nearly orthogonal directions in representation space, yet they are tightly coupled for the task and appear along the time axis. Dot‑product attention scores similarity via an inner product, favoring aligned tokens and being insensitive to near‑orthogonal relationships, forcing the network to recover them indirectly. To address these blind spots we introduce Time‑Frequency Geometric Cross‑Attention (TFGCA), a drop‑in module for existing VLA models. TFGCA applies a learnable stationary wavelet transform per dimension to decompose the action chunk into time‑frequency tokens. Each time token then retrieves information from these tokens via cross‑attention that fuses the dot‑product similarity with the wedge‑product magnitude (sensitive to near‑orthogonality) using a learnable weight. A zero‑initialized residual reproduces the base behavior at initialization, allowing TFGCA to be attached to a pretrained VLA and fine‑tuned jointly. Compared with the same‑source baseline, TFGCA improves in‑distribution LIBERO by +1.5 on average, OOD LIBERO‑Plus by +6.3, the randomized average under RoboTwin domain randomization by +28.5, and the overall success rate on three real‑robot AgiBot A2 tasks by +11.67 points, with especially large gains out of distribution.
Review