Active

Patching deficiencies in general-purpose planners using learned components

Learning to Patch Plans Based on Inaccurate Models - IROS 2020

Learning primitive manipulation skills for using in multi-step Task and Motion Planning problems

I continue investigating this problem in my PhD with other students in my lab: Mohit Sharma, Jacky Liang, and Alan Zhao

Touch corresponds to important semantically useful events such as coming into contact and sliding. Thus it can be a useful tool for reliable manipulation. Even with pose uncertainty, a detect contact provides reliable feedback.

My previous work attempted to use a simple force controller while executing a position-based DMP as an additional signal. More updated work with Jacky Liang, Mohit Sharma, and Alan Zhao learns how to compose position and force controllers that are orthogonal to each other so progress made in one dimension does not affect progress many in another dimension

If a human asks for something like the "half-empty water bottle in the cabinet" how can the robot determine what the human is talking about while also accomplishing the task?

We address this problem using Hierarchical Planning in the Now to combine information gathering actions as a prerequisite for achieving the goal, planning with placeholders until the object is known.