ICRA 2026 | 2nd Workshop on Dexterity with Multifingered Hands:Hardware, Sensing, and Skills
Hongzhu Li, researcher at Amazon and an incoming University of Chicago faculty member, opens an ICRA 2026 workshop on multifingered-hand hardware, sensing, and skills, then introduces talks and discussion with researchers and industry speakers.
In this talk
- 0:01
Dexterity measured by throughput
Hongzhu Li welcomes attendees and frames the workshop around the fast-growing interest in dexterous hands. Howie argues that hand performance should be judged by throughput, which combines speed and reliability, rather than success rate alone. He presents evidence that tactile sensing can increase task throughput, and that additional finger degrees of freedom can reduce the time spent on regrasping and repositioning.
- 14:26
Designing hands and policies together
Howie describes a wrist with collocated joints and quasi-direct-drive actuators that can reach into constrained spaces and handle forceful contact. A compliant wrist helps absorb position errors during tasks such as wiping, while high-resolution tactile sensing supports faster hand interaction. He also introduces KARMA, a metric that estimates how well a hand can translate and rotate objects across different initial grasps, and argues that reinforcement learning and careful attention to controller gains matter for robust physical interaction.
- 37:56
Optimizing fingertip shape and learning contact skills
Rica Antonova presents a co-design pipeline that combines fingertip morphology optimization with reinforcement learning. The team varies fingertip shape and silicone properties, then manufactures low-cost tips for tasks that include grasping small screws, LEDs, and shallow surgical objects. Jia Song describes simulation-trained hand policies for grasping, rotation, and assembly, while emphasizing the difficulty of transferring contact-rich manipulation to real hardware.
- 1:48:12
Touch, structure, and scalable manipulation
Georgia Chalvatzaki argues that dexterity depends on how contact is structured, not simply on finger count. Her examples include planning contact switches, detecting slip with high-frequency tactile sensing, and combining sensors such as GelSight and Tac for dynamic tasks. Other presentations cover tactile-driven gentle grasping, drumstick manipulation, tactile calibration, and hands that use both fingernails and finger pads to handle thin objects.
- 3:00:16
From hand hardware to learning systems
Later speakers discuss tendon-driven and underactuated hands, compliant muscle-like actuators, open-source platforms such as ORCA, and simulation for soft interactions. Jeannette Bohg presents a play-trained policy that can be fine-tuned for assembly tasks, while other work explores real-world reinforcement learning and tactile feedback. The closing panel considers embodiment, data collection, classical control, and how general-purpose learning can support task-specific policies.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.