Welcome to the website of the Geometric and Physical Computing group at Northeastern University’s College of Computer and Information Science. We develop algorithms, software, and hardware to address open problems in robotics, perception, manufacturing, and human-machine interfaces.
PI: Marsette Vona
Location: 214 West Village H, 440 Huntington Avenue, Boston MA (directions, map)
Mouseover each image for a description, click for details.
January 7, 2013: Our paper Sparse Surface Modeling with Curved Patches was accepted for publication at ICRA 2013. This builds on our IROS 2011 paper Curved Surface Contact Patches with Quantified Uncertainty.
August 21, 2012: Our paper Teaching Robotics Software with the Open Hardware Mobile Manipulator was accepted for publication in the IEEE Transactions on Education special issue on Robotics Education. This paper presents our new open-hardware/open-source OHMM platform for teaching robotics software and algorithms, along with pedagogical results from using it in our course CS4610.
August 21, 2012: Controlling a Robot from Another Planet, an interview with Prof. Vona about the NASA Curiosity rover, was featured on the NEU website today.
July 26, 2012: Ph.D. student Dimitrios Kanoulas gave talks on our patch-based perception algorithms to the e-Motion and PERCEPTION teams at INRIA and at LAAS-CNRS Toulouse.
July 6, 2012: Our new paper Moving Volume KinectFusion has been accepted at the British Machine Vision Conference (BMVC) 2012. Newcombe and Izadi et al’s KinectFusion is an impressive new system for real-time tracking and dense 3D mapping using the Kinect. The original algorithm is limited to a relatively small volume fixed in the world, limiting applications for mobile robot perception. Our approach adds remapping algorithms that allow the camera to roam freely. Shortly after we submitted our work for peer review in May 2012 we learned that two other groups are also developing alternative approaches to scale up the KinectFusion map. A key distinction of our method is the ability to rotate the volume. We have implemented our system based on the kinfu implementation in the Point Cloud Library, and we are preparing our code for an open-source release. Videos and more info here.
May 14, 2012: The Driverless Car, an interview with Prof. Vona about the robotic technology behind driverless cars, was featured on the NEU website today.
May 1, 2012: Ph.D. students Henry Roth and Dimitrios Kanoulas will travel this summer for research internships at NASA/JPL in Pasadena, CA and INRIA in Grenoble, France. Henry will be working with Jeff Norris’s OPS Lab on a new augmented reality system using KinectFusion; Dimitrios will be working with the e-Motion group on perception algorithms for road and traffic environments.
April 27, 2012: M.S. student Jessica Lowell presented her thesis BlueSANE: Integrating Functional Blueprints with Neuroevolution today. Prof. Vona was her advisor and Joe Ayers was also on the committee. Jessica will enter the Ph.D. program at Tufts University in the fall.
February 28, 2012: Building a Better Robot, a story about our research and Prof. Vona’s recently awarded NSF CAREER grant on 3D perception and compliant contact, was featured on the NEU website today.
January 29, 2012: The new website for OHMM is now live with details on the robot, photos, hardware designs, software sources, and curriculum materials.
We are currently focusing on several problems in perception and control for bipedal and humanoid locomotion on very uneven 3D terrain. These were motivated by some of our prior work in compliant climbing and stair-stepping, which led us to consider how compliant and proprioceptive motion strategies can be combined with on-line perception of uncertain contact surfaces.
We are developing a set of curved and flat patch models to represent both nearby environment surfaces and contacting surfaces on a robot including foot soles and fingertips. Unlike prior approaches (e.g. algebraic surfaces) we use geometrically meaningful minimal parameterizations, we quantify uncertainty in patch shape and pose, and we include the patch boundary. Fast perceptual algorithms can detect instances of these patches around a moving robot, and as patches are observed and re-observed a spatial map can be built to support motion planning and control.
We recently released a new software package called the Surface Patch Library (SPL) which includes models of 10 types of curved surface patches and an algorithm to fit them to potentially noisy range sensor data. Uncertainty is quantified throughout using covariance matrices. Some of the mathematical foundations of the system are described in this paper.
We are also working with KinectFusion to develop algorithms that will enable its use on a mobile robot in rough terrain. This could provide both localization and a low-level dense model of nearby terrain upon which we can fit higher-level surface patches. In our first paper in this area we present experiments with a remapping approach to allow the KinectFusion modeling volume to move with the sensor through the world. More info and videos here.
A primary advantage of bipedal locomotion (vs wheels or tracks) is the potential to negotiate highly faceted 3D trails as humans do. An understanding of how to address this challenge will help enable intelligent prosthetics, human locomotion aids, and robots that can safely follow and assist humans in rugged terrain.
Sensing and mapping upcoming footholds in real time seems inescapable for this task. Though some recent robots have achieved impressive advances in mechanics and control for walking on rough terrain, they largely operate without significant forward-looking perception. Work to implement these perceptual functions, or even to study human perception in this task, is still in its infancy. We are thus developing a body-worn sensor system to acquire quantitative datasets observing human locomotion on rocks and correlated perceptual behavior.
We are studying humanoid locomotion on terrain composed of sloped planar facets as an in-lab approximation for hiking on a rocky trail. We are considering a four-phase strategy: (1) acquisition of a next-facet pose estimate by range sensors, including a quantified estimate of uncertainty; (2) passive compliant foot placement based on this estimate, with leg stiffness modulated according to the degree of uncertainty; (3) post-contact proprioception (joint angle sensing) to reduce uncertainty; (4) follow-through of torso and trailing leg with up-modulated leg stiffness for control.
Though most existing humanoids have high-impedance joint actuators, recently a few low-cost mini-humanoids have become available which enable modulation of actuator stiffness by a combination of passive backdriveability and on-line adjustment of servo position loop proportional gain. Our experimental apparatus is based on this approach to control joint compliance.
Group Founder and Principal Investigator
|BlueSANE: Integrating Functional Blueprints with Neuroevolution|
|Accepted into the Ph.D. program at Brandeis University.|
Past and present research funding:
NSF CAREER: Reliable Contact Under Uncertainty: Integrating 3D Perception and Compliance (PI Marsette Vona)
NSF MRI-R²: Development of a Second-Generation Applications-Driven Wireless Sensor Networking Instrument (PI Guevara Noubir)
The best way to contact us is to send email to Marsette Vona.
If you will be visiting us, our lab is located in 214 West Village H, 440 Huntington Avenue, Boston MA (map). Take the MBTA Green Line E train to the Northeastern stop, then walk one block west on Huntington Ave. WVH is the glass-facade building at the corner of Parker and Huntington, diagonally southeast across Huntington from the Museum of Fine Arts.
Our mailing address is
360 Huntington Ave 202 WVH, attn Marsette Vona Boston, MA 02115
Some of our colleagues and collaborators include