David Held, faculty at Carnegie Mellon University, presents research on precise, generalizable robot manipulation at the 2026 Spring Robotics Colloquium, hosted by the Paul G. Allen School and introduced by Abhishek Gupta.
Held contrasts broad knowledge in large pretrained robot policies with their difficulty executing precise actions. In a 2023 example, a robot recognized Taylor Swift but dropped a Coke can on the table instead of placing it beside her. He argues that combining generality with precision calls for policies that use large datasets while adding structure to guide action.
ArticuBot learns to open unfamiliar articulated objects using more than 40,000 simulated demonstrations, then transfers to three robot setups and varied real-world locations. Its high-level policy predicts gripper subgoals, while a low-level policy produces the motions. An expanded system trained on over 180,000 trajectories handles opening, closing, grasping, and pick-and-place, and can be fine-tuned for laundry folding with 33 demonstrations.
Held describes using third-person human videos to train high-level predictions of shared task subgoals, while training the robot's motion policy on robot data. This separates what a task requires, such as grasping and placing a mug, from the hand movements specific to a human or robot. The method outperforms the presented Diffusion Policy and MimicPlay baselines on the evaluated tasks.
For placement tasks, the team combines point-cloud diffusion with a dense Gaussian mixture model that proposes likely target locations before diffusion refines the placement. Separating object shape and orientation from translation helps with precision, and the method reports 80% to 100% success on industrial insertion tasks requiring about one millimeter or one degree of accuracy. It also generalizes across object shapes and supports a deformable cloth-hanging task.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
David Held, associate professor at Carnegie Mellon University and head of the RPAD Lab, presents research on precise, generalizable robot manipulation at a GRASP on Robotics seminar hosted by a Penn faculty member.
David Held, a Carnegie Mellon University faculty member, presents research on perceptual robot learning at the UW Robotics Colloquium, introduced by Dieter Fox.
David Held, an assistant professor at Carnegie Mellon University's Robotics Institute, discusses perceptual robot learning and manipulation research with the umrobotics audience.
