RI Seminar: Tucker Hermans : Improving Multi-fingered Robot Manipulation...

In this talk

1. 6:38

   Plan grasps through learned success models

   Hermans describes grasp planning as probabilistic inference: a learned classifier estimates whether a hand configuration will succeed, and an optimizer searches for a high-probability, feasible grasp. In simulation, this approach outperformed heuristic grasps and selecting from sampled candidates; real-robot trials also improved on a heuristic baseline. The training set contained 1,500 simulated attempts, about 200 of them successful.

2. 19:33

   Guide grasp type and avoid collisions

   Separate classifiers for precision and power grasps let the planner target a grasp suited to the task. On tested objects, explicitly modeling grasp type substantially improved precision-grasp success, while power grasping reached 100% success. Hermans also describes adding a learned signed-distance representation of an object, joint limits, and collision constraints to grasp optimization.

3. 30:47

   Manipulate objects within the hand

   For in-grasp manipulation, Hermans plans joint-space trajectories without modeling object or contact dynamics. The method treats one finger as rigidly attached to the object and relaxes the constraint for other fingers to preserve a stable grasp. Across 500 trials on 10 objects, the median position error was 1.32 centimeters, with no drops reported.

4. 39:34

   Relocate fingertips and move object collections

   A fingertip-gaiting method alternates between moving fingers as far as their reachable workspace allows and repositioning the object to bring target contacts within reach. The resulting plans caused objects to fall on the robot, so Hermans describes adding slip-aware feedback as ongoing work. He also presents a graph neural network policy trained in simulation that pushes variable collections, such as cubes or silverware, from images.