Have you always wondered whether your extrinsic sensor transformation globally minimizes its calibration cost function? With our latest work on certifiable monocular hand-eye calibration, wonder no more! Check out the IEEE MFI 2020 paper – we prove that trajectories satisfying observability requirements lead to convex relaxations that are inherently stable to measurement error. The open source implementation of our method is fast, requires no calibration targets, and works for a wide variety of sensors, including monocular cameras!
Congratulations to lab members Valentin Peretroukhin and Matthew Giamou and to our collaborators, David M. Rosen, W. Nicholas Green, and Nicholas Roy at MIT for winning this year’s RSS Best Student Paper Award! Full details and code for the paper, “A Smooth Representation of SO(3) for Deep Rotation Learning with Uncertainty,” are available here. Great work all!
Just a couple of days until the RSS 2020 Workshop on Power-On-and-Go Robots: ‘Out of the Box’ Systems for Real-World Applications! We are extremely excited about the event, which will be streamed live via Zoom. Full details are available at https://www.power-on-and-go.net/.
The workshop will bring together researchers from diverse backgrounds to address topics related to power-on- and-go robots: robotic systems that are able to successfully deal with new situations fluidly and to adapt immediately to new environments or to changes in their own operating parameters. We have a fantastic lineup of speakers and panelists, including Hadas Kress-Gazit (Cornell), Stefan Leutenegger (Imperial College), Nathan Michael (CMU), Arne Sieverling (Realtime Robotics), Luca Carlone (MIT), Ali Agha (JPL), Dorsa Sadigh (Stanford), and Gaurav Sukhatme (USC)!
As a follow-on to the workshop, our Call for Papers for a special issue of the journal Autonomous Robots is out now as well, with more details and deadlines available here. We hope you will be able to join us for an insightful virtual event!
We are thrilled to announce that our paper on a smooth representation of belief over SO(3) for deep rotation learning will be presented at Robotics: Science and Systems 2020. This work was a collaboration between STARS lab members Valentin (now alumni!) and Matthew, together with Prof. Nicholas Roy, W. Nicholas Greene and Dr. David M. Rosen at MIT. Watch the 5 minute video summary below!
Several STARS lab members helped organize the Debates on the Future of Robotics Research workshop that was held (virtually) on Friday, June 5th as part of ICRA (2020). The workshop live stream received over 1100 unique viewers and was enthusiastically received by both participants and viewers! Congratulations to STARS students (and alumni) Matthew Giamou, Valentin Peretroukhin and Lee Clement, as well as Prof. Kelly, who helped organize the event!
We’re excited to have three lab papers that will be presented at this year’s (virtual) ICRA 2020 conference! Highlights and video links are below.
Check out our new extension to DPC-Net (from ICRA 2018): we show that DPC networks can be trained in a fully self-supervised manner, which improves accuracy and allows for retraining online in new environments!
Got features? Our recent RA-L and ICRA 2020 work demonstrates how to learn maximally-matchable image mappings to dramatically reduce the data needed for experience-based navigation.
Check out our work on a QCQP approach to inverse kinematics for redundant manipulators. We show that this difficult, nonconvex problem often admits a provably tight convex relaxation that can be efficiently solved! Coming soon to MoveIt!
We’re delighted to be co-organizing the RSS 2020 Workshop on Power-On-and-Go Robots: ‘Out of the Box’ Systems for Real-World Applications! More details at https://www.power-on-and-go.net/
Substantial advances have been made over the past two decades in the area of mobile robot autonomy, in part due to the development of sophisticated methods to fuse data from multiple information sources. However, these gains come with the caveat that proper system initialization and calibration are essential. Starting with or quickly discovering the “right” initial conditions for the selected estimation, planning, and control algorithms is a crucial but largely overlooked problem that has not yet been fully tackled by the community—instead it is often regarded as a post-hoc ‘engineering’ issue rather than a key safety concern, for example. In a future where robots actively operate alongside people in human environments, businesses and consumers will demand that the machines work correctly the first time, every time, anywhere, with minimal external (human) intervention.
The workshop will bring together researchers from diverse backgrounds to address topics related to power-on- and-go robots: robotic systems that are able to successfully deal with new situations fluidly and to adapt immediately to new environments or to changes in their own operating parameters.
Please consider submitting an extended abstract for presentation at the virtual workshop! The deadline has been extended and is now June 21st, 2020.
We’re excited to announce the release of our University of Toronto Canadian Planetary Emulation Terrain Energy-Aware Rover Navigation Dataset. The dataset was gathered by a small, four-wheeled rover at a planetary analog test facility in Canada. The rover was equipped with a suite of sensors designed to enable the study of energy-aware navigation and path planning algorithms. Our International Journal of Robotics Research data paper is available here.
The dataset includes more than 14,000 colour omnidirectional stereo panoramas captured from a synchronized 10-camera cluster and 16,000 high-resolution monocular terrain images. IMU, pyranometer (solar irradiance), drive power consumption, wheel encoder, and GPS measurements are also included. All data are presented in human-readable text files and as standard-format images; additional Robot Operating System (ROS) parsing tools and several georeferenced aerial maps of the test environment are also included. The full dataset is accessible from:
Congratulations to Dr. Valentin Peretroukhin for successfully defending his PhD on March 6! His dissertation, “Learned Improvements to the Visual Egomotion Pipeline” described ways to augment classical visual egomotion techniques with learned models. We wish him the best as he moves to Boston to postdoc at MIT!