Vision Language Actions using raw action as action tokens

Vision–Language–Action (VLA) models treat raw, continuous actions (e.g., joint torques, velocities, gripper commands) directly as action tokens, embedding them into the same sequence modeling framework as vision and language. This avoids learned action abstractions, preserves low-level control fidelity, and enables end-to-end training from perception and language to precise motor execution, at the cost of higher dimensionality and sensitivity to embodiment.