SPARK

SPARK is a training-free neurosymbolic manipulation framework that uses sequential planning via anchored robotic keypoints. It reaches 43.7% on six LIBERO-PRO position and task evaluation cells, more than doubling VLA baselines. Core innovation: a single Gemini call generates a typed behavior tree, five basic primitives compose 30+ typed skills, and heavy computation is invested in perception. Uses SAM3 for object grounding, and when primitives fail, the system re-perceives rather than replanning. Supports UR10e, Franka FR3, and dual-arm Franka across nine manipulation tasks. Can auto-generate annotated trajectories for VLA training.