New SafePBDS framework enables dexterous robot hands with 92.5% grasp success

Robotic dexterous manipulation has long struggled with reconciling objectives and constraints across heterogeneous geometric spaces—a robot's configuration manifold (ℝ⁷) must simultaneously track end effector poses on SE(3) while satisfying obstacle avoidance in ℝ. The new Safe Pullback Bundle Dynamical Systems (SafePBDS) framework, developed by Albert Wu and colleagues at Stanford, elegantly solves this by computing optimal, certifiably safe configuration manifold accelerations from objectives defined on arbitrary task manifolds. SafePBDS builds on prior work combining predefined task manifold dynamical systems for autonomous motion, adding two key innovations: a pullback control barrier function that converts task manifold safety conditions into linear constraints, and a task manifold action interface that allows high-level policies to inject low-dimensional residual motions while preserving safety. This means zero input recovers autonomous behavior, while arbitrary inputs remain safe—enabling efficient exploration and precise motion.

Validated on a 23-DOF Franka Panda-Allegro Hand platform, SafePBDS achieved a 92.5% success rate across 20 household objects and 120 trials during dexterous grasping. By using the action interface to exclude any one finger via a one-dimensional action, the method reached 94.4% 3-finger grasp success across 3 objects and 36 trials. Most impressively, SafePBDS enabled the first model-based, fully actuated palm-down in-hand reorientation, exceeding 360° of yaw rotation in both directions under varying object weight and wrist motion. This geometrically principled approach offers a path toward robust, safe, and steerable robotic manipulation in real-world settings, bridging the gap between high-level task planning and low-level motor control.