Michelle Zhao

Hi, I am a fifth-year PhD student in the Robotics Institute at Carnegie Mellon University, working on human-robot collaboration. I am advised by Reid Simmons and Henny Admoni, and am part of the Human and Robot Partners (HARP) Lab and Reliable Autonomous Systems Lab (RASL). I am supported by the National Defense Science and Engineering Graduate (NDSEG) Fellowship.

My research interests lie in the areas of interactive machine learning, reinforcement learning, and human robot interaction. See here for a research statement. I graduated from Caltech ('20) with a degree in Computer Science and a minor in Information and Data Science, where I worked with Yuxin Chen and Yisong Yue on machine teaching and Soon-Jo Chung on distributed control.

Email  /  CV  /  Bio  /  Google Scholar  /  Twitter  /  Github

Research

I'm interested in developing algorithms to enable fluent human-robot coordination and collaboration (2-page research statement). I currently work on projects focused on uncertainty quantification, human-robot collaboration, adaptation, and modeling human behavior.

Uncertainty quantification offers a way for the learner (i.e. robot) to contend with distribution shifts encountered during deployment by actively seeking additional feedback from an expert (i.e. human) online. From the conformal prediction side, we introduce a novel uncertainty quantification algorithm called intermittent quantile tracking (IQT) that leverages a probabilistic model of intermittent labels, maintains asymptotic coverage guarantees, and empirically achieves desired coverage levels. From the interactive IL side, we develop ConformalDAgger, a new approach wherein the robot uses prediction intervals calibrated by IQT as a reliable measure of deployment-time uncertainty to actively query for more expert feedback.

Assistive robotic arms often have more degrees-of-freedom than a human teleoperator can control with a low-dimensional input, like a joystick. To overcome this challenge, existing approaches use data-driven methods to learn a mapping from low-dimensional human inputs to high-dimensional robot actions. However, determining if such a black-box mapping can confidently infer a user's intended high-dimensional action from low-dimensional inputs remains an open problem. Our key idea is to adapt the assistive map at training time to additionally estimate high-dimensional action quantiles, and then calibrate these quantiles via rigorous uncertainty quantification methods.

In this work, we investigate how to generate multi-agent strategy explanations for human-robot collaboration. We formulate the problem using a generic multi-agent planner, show how to generate visual explanations through strategy-conditioned landmark states and generate textual explanations by giving the landmarks to an LLM.

We propose a method for representing human and robot contribution constraints in collaborative human-robot tasks. Additionally, we present an approach for learning a human partner's contribution constraint online during a collaborative interaction. We evaluate our approach using a variety of simulated human partners in a collaborative decluttering task.

This paper presents a framework for defining artificial intelligence (AI) that adapts to individuals within a group, and discusses the technical challenges for collaborative AI systems that must work with different human partners. Collaborative AI is not one-size-fits-all, and thus AI systems must tune their output based on each human partner's needs and abilities.

This work autonomously recognizes available task-completion strategies by observing human-human teams performing a collaborative task. By transforming team actions into low dimensional representations using hidden Markov models, we can identify strategies without prior knowledge. Robot policies are learned on each of the identified strategies to construct a Mixture-of-Experts model that adapts to the task strategies of unseen human partners.

We show by learning the set of hidden states representing a team’s observed collaborative process behaviors over time, we both learn information about the team’s collective intelligence (CI), predict how CI will evolve in the future, and suggest when an agent might intervene to improve team performance.

Workshop Papers

While today’s robot learning algorithms increasingly enable people to teach robots via diverse forms of feedback, they place the burden of responsibility on the human to perfectly understand what the robot doesn’t know and provide the “right” data. My research contends that robots should be proactive participants— they should bear some of the burden of knowing when they don’t know and should ask for targeted help. I tackle this problem by extending foundational uncertainty quantification techniques to the HRI setting, enabling robots to rigorously “know when they don’t know” even when relying on black-box policies and to ask for strategic help.

Best Poster Presentation Award

We consider the problem of teaching a human partner a joint reward function, which captures how both human and robot should contribute to the task. This reward, which is known only to the robot, is joint over human and robot actions, and encompasses constraints over how the human and robot should contribute to a task. By adapting existing machine teaching frameworks for our collaborative domain, we seek to provide a minimal number of demonstrations such that a human can learn the rewards.

We propose a method for generating a customizable quantity of synthetic data that reflects the variability in task execution styles seen in the real-world task, and enables us to train a baseline sequential model that predicts the next action a participant will take within a cooking activity.

We present a closed-loop interaction framework that adapts the level of information complexity based on the human partner’s observable cognitive understanding. This work investigates how knowledge and preparation impact the suitability of di erent complexity levels, motivating dynamic interaction.

Other

Design by Jon Barron.