Steve Levine

Steve Levine

Ph.D Research

My Ph.D research focused on execution for teams of humans and robots working together on a shared task. I examined this problem from several different perspectives, developed algorithms for each, and demonstrated on robotic hardware testbeds. I began working on my Ph.D immediately after finishing my M.Eng in 2012 and was advised by Prof. Brian C. Williams.

Concurrent Intent Recognition & Adaptation

My first work in this area can be seen as an extension of my M.Eng, which focused on robotic execution of temporal plans. A key part of M.Eng thesis was the use and derivation of causal links for these temporal plans, and their use online to do proactive execution monitoring (i.e., predicting failure before it occurs).

My initial Ph.D work extended this area, by making the following changes to the temporal plan representation:

  • Plans now target humans and robots. That is, these are “team plans” – some actions are for the robots, and other actions are for the humans
  • Plans now have choices. Some of these choices are made by the robots (“controllable”), while others are reserved for the human (“uncontrollable”).

These two changes are enough to enable a rich human-robot interaction. I developed a framework that concurrently (1) recognizes a human’s intentions as he/she makes choices and acts according to the team plan, and (2) selects adaptation actions for the robot teammates accordingly.

To do this, I extended the notion of causal links to the contingent setting as labeled causal links, which capture their dependence on choice. This allows Pike to reason over which choices can occur based on other choices that have been previously made in the plan. We can prune out certain choices that would never be feasible – pruning out uncontrollable choices amounts to intent recognition (as we are inferring options for what the human will do), and pruning out controllable choices amounts to honing in on suitable adaptations for the human’s intents.

For more details, please see my ICAPS 2014 paper (link below). Feel free to also watch my presentation at the conference.

Probabilistic Intent Recognition & Adaptation

The above work provided a framework for concurrent intent recognition and adaptation. I’ve since been extending this framework to a probabilistic setting, in which we now have a probability distribution over likely choices that the human may perform. This allows Pike to perform a chance-constrained execution, by making choices that are safe enough and better match the human’s intentions.

This section will be updated with more information and details later!

Testing in Hardware

Over the years, I’ve contributed to a number of hardware demonstrations of AI algorithms in my lab. Below is one scenario from several years ago where I am collaboratively building a (simplified!) frame of an airplane wing: two WAM arms hold a piece it in place, while the red Baxter robot recognizes my intent and hands me appropriate tools or parts for my actions.

Publications

Selected publications: