Maria Bauza, an MIT mechanical engineering PhD student, discusses tactile sensing for dexterous robot manipulation in a seminar hosted by the AI Robotics Seminar at the University of Toronto.
Bauza contrasts stable bin picking with more demanding tasks such as placing or using an object. A robot may need to choose a grasp based on stability, what it can perceive, whether its motion is feasible, and whether it will need to regrasp. For example, grasping a crayon at its center can be stable but make its pose hard to identify.
Bauza describes a kitting system that picks objects, estimates their poses through touch, and places them, sometimes after a regrasp. A digital twin uses object geometry to simulate candidate grasps and represent tactile input as contact locations. A neural network scores grasps against objectives such as stability, localization quality, and motion feasibility, with the aim of transferring the selected actions to a real robot.
For tactile localization, Bauza assumes the object geometry is known and seeks to estimate pose from the first contact. Simulation generates a dense set of possible contact shapes and their corresponding poses; a neural network maps sensor readings into a representation that can be matched against that set. Point-cloud registration can refine a match, while a pose distribution preserves ambiguity when different poses produce similar contacts.
Bauza explains that multiple contacts and prior pose constraints can be combined as probabilities. In tests, single-touch estimates were often within a few millimeters, though non-unique contacts can make a single pose impossible to distinguish. Keeping the ten most likely matches helped retain alternative poses, and additional contacts improved estimates for more ambiguous objects.
The tactile sensors use a camera to record deformation in an elastic membrane; Bauza says the lab builds them in-house, with component costs around $100 and $20 as an achievable target. She describes combining touch with vision, tracking contacts over time, and estimating object models when geometry is unknown. She also points to system-level optimization that identifies which task or system constraints limit accuracy.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
MIT PhD candidate Maria Bauza presents research on precise robot pick-and-place for novel objects at the UW Robotics Colloquium, introduced by Dieter Fox of the Paul G. Allen School.
