Ken Goldberg, a roboticist at UC Berkeley, presents research on robot grasping for the University of Toronto Robotics Institute, hosted by his former PhD advisor Animesh.
Goldberg argues that robots remain unreliable at picking up everyday objects because perception, motion control, and contact physics all involve uncertainty. Even a tiny shift or a patch of different surface friction can change how an object moves. He points to homes and e-commerce warehouses as settings where better grasping could help.
Goldberg describes an early approach that uses precise object geometry and force models to identify stable grasps, but says it assumes more certainty than robots have. Dex-Net treats pose, shape, friction, and other factors as uncertain, then samples possible outcomes to estimate a grasp’s chance of success. Real robot data can teach grasping too, but he notes that one cited system still failed about 20% of the time after 10 million examples.
To generate training data at scale, the lab built a collection of 3D models and simulated candidate grasps under varied conditions. A neural network then scores grasps from depth images, with structured sensor and control variation intended to help transfer results to physical robots. Goldberg reports that the system handled unfamiliar objects and was extended to picking from cluttered bins.
The lab added suction grasping to its parallel-jaw work and developed a shared success estimate to choose between the two tools. The system can route porous objects to the jaws when suction fails, though small, transparent, heavy, or awkward items remain difficult. Goldberg also describes projects on fast object movement, cable untangling, surgery, and tending a mixed-species garden.
In discussion, Goldberg weighs soft grippers’ ability to conform to objects against wear and the risk of covering labels or other useful surfaces. He calls tactile sensing a major gap and notes that human touch can detect very fine surface features. For warehouse use, he says a success rate around 95 to 98 percent may be practical if robots can also work quickly.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Bloomberg Originals host and narrator, with 1X founder Bernt Børnich, UC Berkeley professor Ken Goldberg, GXO’s Patrick Kelleher, and other robotics, industry, and labor experts, examines humanoid robots’ potential and the hurdles they face in factories, warehouses, homes, and hospitals.
Ken Goldberg, a professor at UC Berkeley, discusses how robotics can combine learned models with engineering at an IEEE Robotics and Automation Society event.
Ken Goldberg, a robotics researcher at UC Berkeley and co-founder of Ambi Robotics, explains at TEDxMarin why household robots struggle with manipulation and describes progress in warehouse sorting and home tasks.
