Dexterous Decision-Making for Real-World Robotic Manipulation

Rachel Holladay

MIT

For a robot to prepare a meal or clean a room, it must make a large array of decisions, such as what objects to clean first, where to grasp each ingredient and tool, how to open a heavy, overstuffed cabinet, and so on. To enable robots to tackle these tasks, I decompose the problem into two interdependent layers: generating a series of subgoals (i.e., a strategy) and solving for the robot behavior that achieves each of these subgoals. Critically, these subgoal solvers must account for force, motion, deformation, contact, uncertainty, and partial observability to accomplish a rich set of manipulation tasks.

My research contributes models and algorithms for generating this kind of robot behavior such that it both generalizes to new environments and can be composed into long-horizon strategies. In this talk, I will first discuss how this approach has enabled robots to perform tasks that require reasoning over and exerting force, like opening a childproof medicine bottle with a single arm. Next, I will illustrate different ways robots can make robust choices in the face of uncertainty. For example, this empowers robots to reliably chop up fruit of unknown ripeness! Finally, I propose how robots operating with uncertain dynamics can generate cautious behavior, such as shoving an object near the edge of the table without it falling.

Rachel Holladay is a Ph.D. student in the Electrical Engineering and Computer Science Department at the Massachusetts Institute of Technology.  Her research focuses on developing algorithms and models that enable robots to robustly perform long-horizon, contact-rich manipulation tasks in everyday environments. She received her B.S. in Computer Science and Robotics from Carnegie Mellon University.

https://columbiauniversity.zoom.us/j/99497834307?pwd=OGsvQ2RkcGZRV0dBNUhlMm5sbld2Zz09

Meeting ID: 994 9783 4307
Passcode: 59756