Peter Mitrano

I am a PhD candidate at the University of Michigan advised by Professor Dmitry Berenson, defending December 2023. I started in the fall of 2018 after I completed my Bachelors at WPI double majoring in Robotics Engineering and in Computer Science. I have published in the journal Science Robotics, and at top-tier conferences such as RSS and ICRA. I currently have over 200 repositories on my github., including open source contributions to MoveIt, MuJoCo, and Gazebo. In my personal time, I enjoy rock climbing, flow arts, woodworking, and a wide variety of music.

Goals

I would like to continue doing robotics research with a focus on real robots that learn. I hope to contribute to open-source and articles published to academic conferences, and fight over-hyping or misleading the public on the state of robotics. Most importantly, I want to work on a team that is excited to make a positive impact, and considers equity and ethics in all aspects of their work.

Research

My research interests include learning, planning, and simulation for robotics.

During my PhD, I have studied deep learning and motion planning and have a particular interest in deformable object manipulation. My thesis explores a core challenge in deformable object (and most other types) of manipulation. How do we make the best use of unreliable dynamics models and limited data? I have designed and implemented several dynamics learning methods, as well as planning methods which use learned models. For deformable object manipulation, our methods for planning with learned models must be aware of modeling inaccuries, and I have explored a number of ways for doing this. Throughout my PhD, I have read and practice ideas and techniques including instance segmentation, graph neural networks, generative modeling, video prediction, model predictive control, uncertainty qunaitification for planning, and others.I also have experience working with real robot hardware. This includes working with low-level control and sensors, embedded systems, and dealing with real data collection and online learning on real robots. I have extensive experience with Gazebo and MuJoco, have used PyBullet and Isaac Sim, and have a practical knowledge of the underlying numerical methods used for simulation of rigid body and deformable motion and contact.

I have worked with Professor Carlo Pinciroli on a swarm robotics project. We are studing the space of possible behaviors for memoryless computeless robots, specifically showing that they are capable of n-class segregation with a single primitive sensor and no communication. You can learn more by reading the paper we wrote for the project as a part of the Swarm Intelligence course at WPI.

I also worked with Professor Scott Barton on the intersection of machine learning and music. As a musician myself, this is an interesting application of machine learning.

As a freshman at WPI, I worked as a research assistant for Russell Torris, under Professor Sonia Chernova. I also worked with Professor Chernova for a summer at Georgia Tech. My project focused on using demonstration to improve indoor robot navigation.

Papers

Videos for recent papers can be found on the ARMLab YouTube channel.

Grasp Loop Signature: A Topological Representation for Manipulation of Ropes and Cables. Under Review for ICRA, 2024

Focused Adaptation of Dynamics Models for Deformable Object Manipulation. ICRA, 2023

Data Augmentation for Manipulation. RSS, 2022

Learning Where to Trust Unreliable Models in an Unstructured World for Deformable Object Manipulation. Science Robotics, May 2021

Learning When to Trust a Dynamics Model When Planning With Physical Constraints. ICRA 2020 Workshop MLPC

Learning When to Trust a Dynamics Model for Planning in Reduced State Spaces. ICRA 2020

A Minimalistic Approach to Segregation in Robot Swarms. The 2nd IEEE International Symposium On Multi-Robot and Multi-Agent Systems

Using Recurrent Neural Networks to Judge Fitness in Musical Genetic Algorithms. The 5th International Workshop on Musical Metacreation (MUME)

Invited Talks

Learning Where to Trust a Dynamics Model for... - IAM Lab, Carnegie Mellon University, April 2022

Social Media

Twitter
PickNik Blog Roboflow Blog
My (very casual) Blog.

Internships

PickNik Robotics

I integrated instance segmentation into a traditional manipulation planning pipeline, enabling the first autonomous perception-driven behaviors in the MoveIt Studio beta product.

Uber ATG Pittsburgh

I worked on the problem of predicting future paths of actors detected by the autonomous vehicle. My work consisted mostly of software engineering.

Open Robotics (formerly OSRF)

At OSRF, I worked on the Gazebo robot simulator. I worked on getting Gazebo running on windows, and on allowing high school students in the FIRST Robotics Competition to use Gazebo as a simulator for their robots.

Service and DEI

Participated in Rackham Professional Development Diversity, Equity, and Inclusion Certificate

Completed 10 workshops on a variety of DEI topics and completed the Intercultural Development Inventory.

Robotics Graduate Student Council

I have served as President, Social chair and Colloquium chair. I participated in several outreach events for local girl scout troops and middle schools, including hand-on engineering activities and lab tours. I have also organized colloquiums for UM Robotics students to share and discuss their research.

Education

University of Michigan - PhD in Robotics

Worcester Polytechnic Institute - BS in Computer Science, and in Robotics Engineering

3.95 GPA

PDF CV/Resume

You can view a PDF form of my resume here.

Inspiration for website design from here.