Alberto Rodriguez, a mechanical engineering faculty member at MIT, discusses manipulation primitives for building dexterous robotic systems at the CMU Robotics Institute Seminar, introduced by his former advisor Matt.
Rodriguez contrasts industrial robots, which mostly move objects from place to place, with the flexible hand use seen at his daughter’s crafting table. He frames that flexibility as a long-term goal for robotics and outlines four perspectives on manipulation: mechanics and control, haptics, planning, and visual feedback.
Rodriguez describes the four perspectives as complementary: some emphasize feedback from local contact, while others emphasize prediction from global visual information. A manipulation primitive such as pulling makes assumptions about contact and object motion. Those assumptions simplify mechanics and planning, and help clarify what tactile and visual feedback should do.
Rodriguez separates tactile control into maintaining a desired contact state and moving the object while that state holds. In a pulling example, tactile sensing detects slip so a controller can adjust downward force; another controller moves the object toward its goal. In a cable task, a gripper adjusts force at about 60 Hz to regulate friction, while an LQR controller steers the cable.
Rodriguez describes learning visual affordances for actions such as grasping and suctioning, then extending them to pushing, pulling, and multi-step manipulation. The system uses segmented point clouds and learned primitive models to sample contact locations and predict object transformations for a sampling-based planner. It can plan actions on unfamiliar objects, though the demonstrated execution has no tactile feedback.
Rodriguez says primitives help maintain intended contact modes and make object motion predictable enough for planning. He argues that manipulation systems should combine mechanics, haptics, planning, and vision rather than treating them separately. In the discussion, he notes that parallel tactile and task controllers can conflict, and that selective sticking helps make manipulation more reliable and predictable.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Ya Durban, Atlas product strategy lead at Boston Dynamics, and Alberto Rodriguez, Atlas behavior team director, discuss manufacturing applications, robot training, reliability, and the path to deploying humanoids.
