KineVLA: Towards Kinematics-Aware Vision-Language-Action Models with Bi-Level Action Decomposition

arXiv 2603.17524·medium signal

Introduces kinematics-dense language commands into VLA (Vision-Language-Action) model training — encoding joint angles, velocities, and torque constraints directly into the task specification rather than relying on high-level natural language alone. Bi-level action decomposition separates coarse motion planning from fine kinematic execution, addressing the gap between LLM spatial reasoning and physical robot precision. Demonstrates improved transfer from simulation to physical manipulation tasks versus standard VLA baselines.