Xing Li (Science of Intelligence), “Learning to Manipulate Articulated Objects From Human Demonstrations”

Abstract:

Programming robots to manipulate articulated objects such as drawers, doors, or locks is a challenging task. One of the major reasons for this difficulty is that robots must physically interact with objects, and even minor errors during manipulation can result in significant internal forces that may cause damage.

While robots struggle with these manipulation tasks, humans can effortlessly operate complex mechanisms with great reliability. Moreover, humans can transfer their experience between objects of the same type, resulting in remarkable generalization. This raises the question of how we can transfer these robust and general manipulation skills from humans to robots.

In this presentation, we will introduce a viable solution to achieve this transfer in the context of manipulating articulated objects. Specifically, we will demonstrate that a robot can acquire a manipulation policy that reliably manipulates various instances of the same type based on a single demonstration of a human opening an articulated object.

Following the presentation, we invite those who are interested to participate in an interactive session where we can discuss and share our experiences with controlling the robot with the soft hand in the robotics lab on the second floor.

This talk will take place in person at SCIoI.