Professor Nancy Pollard of Carnegie Mellon University discusses practical robot dexterity at an MIT Robotics seminar, drawing on research with her team and collaborators.
Pollard defines dexterity as finding a correct, quick, economical way to solve a movement problem. She frames the challenge around tasks that are awkward to automate, such as stabilizing a loose automotive sensor while connecting it and holding a bell pepper while cutting it from a plant. The pepper-picking project took three years to move from a one-hand design to a two-handed robot modeled on how people do the job.
Pollard shows that everyday handling involves frequent bumps, and that compliant hands can absorb them and adjust. In a store demonstration, she observed roughly 100 milliseconds between collisions and motion corrections; a rigid prosthetic, by contrast, made picking up small objects difficult. She argues that contact often spreads across the hand rather than occurring at isolated points, and describes representing and transferring contact patches across 3D meshes.
A study by Sun Yu Wang examined hand poses used when sculptors carve, poke, and press clay. Wang evaluated about 100,000 hand designs against these recurring motions; roughly one in ten could reach the target contact arrangement, and nearly all of that subset could perform the tool motions. Pollard says this supports assessing many designs against task needs instead of searching for one universally optimal hand.
Pollard describes using mesh geometry to compute rolling and sliding interactions on real hand and tool shapes, including forms that are hard to represent as simple primitives. Her education kit combines a soft, tendon-driven hand with 12 degrees of freedom and a four-degree-of-freedom gantry, for about $2,000. The foam hand costs less than a dollar, and the team plans to open-source the kit.
Pollard's group is developing a process for transferring demonstrated tasks to robot hands, with a best-case target of about five days when hand design is already in the pipeline. She notes that a single human-like hand is not necessary for every job: simpler grippers and task-specific robots already handle many kinds of factory and food work.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Nancy Pollard, professor of robotics and computer science at Carnegie Mellon and co-founder and president of FutureHand Robotics, discusses building dexterous robot hands for real-world tasks at a CMU Robotics Institute seminar introduced by Henny Admoni.
