Using Simulations to Aid Robot Learning

Simulations are the primary way for a robot to predict actions and safely learn. This project investigates how we can train robot task models using simulations, and how those models can be helpful for human teachers and operators.

TuneNet: One-Shot Residual Tuning for System Identification and Sim-to-Real Robot Task Transfer
A. Allevato, E. S. Short, M. Pryor, A. L. Thomaz
Conference paper, CoRL 2019
PDF | bibtex | video | code

SAIL: Simulation-Informed Active In-the-Wild Learning
E. S. Short, A. Allevato, A. L. Thomaz
Conference paper, HRI 2019
PDF | bibtex

Affordance Discovery using Simulated Exploration
A. Allevato, E. S. Short, M. Pryor, A. L. Thomaz
Extended abstract, AAMAS 2018
PDF | bibtex

3D Computer Vision for Robotics

Using Point Cloud Library and OpenCV, I developed a 3D pose estimation library based on Robot Operating System (ROS). I showed that the library was effective for use in robotic sorting, remote inspection, and small part picking. We've since found it useful for quickly adding visual object detection capability to various robot experiments.

An Object Recognition and Pose Estimation Library for Intelligent Industrial Automation
A. Allevato
Masters’ thesis, Spring 2016
PDF | bibtex | code

Glovebox Workcell Manufacturing and Integration

In this three-year project, our team at the Nuclear and Applied Robotics Group (NRG) deployed robots in various configurations inside nuclear gloveboxes. Developed in conjunction with Los Alamos National Laboratory, we investigated visual recognition, mixed-waste sorting, and manufacturing tasks such as automated hole drilling with a highly customized robotic system

This project resulted in a number of publicly-available ROS packages, many of which are available on the NRG lab’s GitHub pages, which I helped run.

Automated Glovebox Workcell Design
E. Paredes, C. Petlowany, M. Horn, A. Allevato, M. Pryor
Conference paper, WMS 2018

Demonstrating Autonomous and Robust Sorting in a Glovebox Environment
A. Allevato, M. Horn, M. Pryor
Conference paper, ANS D&RS 2016
PDF | bibtex

Characterizing Glovebox Automation Tasks using Partially Observable Markov Decision Processes
A. Allevato, M. Pryor
Conference paper, ANS D&RS 2016
PDF | bibtex

Using a Depth Camera for Object Classification in Nuclear Gloveboxes
A. Allevato, T. Lu, M. Pryor
Podium Presentation/Technical Session, 2015 ANS Student Conference
PDF | bibtex

Choosing Observation Points for a Camera Robot

We developed an algorithm to select a viewpoint from which to observe a task. This can be useful in a two-robot scenario where one robot is the “cameraman” and the other is the “worker” or “actor.” The camera viewpoint is chosen so as to stay out of the actor’s way, while still providing a good view of the task region.

Getting the Shot: Socially-Aware Viewpoints for Autonomously Observing Tasks
A. Allevato, A. Sharp, M. Pryor
Conference paper, ARSO 2017
site | PDFbibtex | slides | code