Contents

Bridging Vision‑Language‑Action Models and Whole-Body Controller via Structured Action Interfaces

Anonymous authors

Overview

Overview of QuadArm-VLA: IsaacLab training, whole-body teleoperation, and four real-world manipulation tasks. — Overview of QuadArm-VLA. We acquire embodiment-specific arm-body coordination through large-scale IsaacLab training and collect whole-body demonstrations through teleoperation. A pretrained vision-language-action model is then adapted to the resulting controller interface. QuadArm-VLA supports (A) mobile pick-and-place, (B) small-object transport and elevated placement, (C) contact-rich box pushing around an obstacle, and (D) drawer opening, object placement, and drawer closing.

Real-World Demos

Task 1: Mobile Pick-and-Place

Task 2: Open Drawer

Task 3: Carry Cube

Task 4: Push Box