Grasp Planning with Samarth Brahmbhatt | AI in Robotics Seminar
Samarth Brahmbhatt, a researcher at Intel and former Georgia Tech robotics and computer vision PhD researcher, presents work on hand-object contact at the University of Toronto’s AI in Robotics Reading Group.
In this talk
- 2:19
Why hand-object contact matters
Brahmbhatt argues that contact and hand pose can reveal how someone intends to use an object, such as pressing a camera button. He sees potential for turning everyday objects into computer interfaces and for informing robotic hand design, though his work focuses on understanding hands in action.
- 5:02
ContactPose data and collection
ContactPose combines 3D object models, hand joints, fitted MANO hand models, images, and surface contact maps. Thermal cameras capture heat left on 3D-printed PLA objects after participants grasp them; multiple camera views and object tracking help reconstruct hand joints. Fifty participants grasped 25 household objects in use and handoff tasks, producing more than 2,300 grasps and nearly 3 million RGB images.
- 20:38
What human grasp patterns reveal
Across the dataset, thumb and index finger contact is common, but contact also frequently involves proximal finger sections and the palm. Little-finger contact is less likely than index-finger contact. Handoff grasps show more concentrated contact, while use grasps spread contact across more of the hand; object contact maps also highlight functional areas such as mouse and camera buttons.
- 26:51
Predicting and fitting contact
Brahmbhatt describes methods that predict object-surface contact from hand geometry or grasp images, associating learned features with points in a point cloud. Contact fitting optimizes hand parameters to attract the hand to contacted regions, repel it from non-contacted regions, and avoid object penetration. ContactOpt combines DeepContact predictions with differentiable contact fitting to refine estimated hand poses, including poses with small errors that can change the inferred contact pattern.
- 50:25
Robotics implications and open questions
In questions, Brahmbhatt says the dataset has not yet been used for robot control; his robotics work has focused on configuring high-dimensional hands from contact maps. He notes that human grasp data transfers most directly to robot hands with similar shapes, and that planning whole-hand contact remains difficult. He hopes the dataset encourages more attention to palm and other non-fingertip contact.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.