ICRA 2022 Recap / First Blog Post

Last week, I attended the 2022 International Conference on Robotics and Automation (ICRA). This was my first in-person conference, and it was a wonderful opportunity to meet new people in the robotics world and catch up with Laboratory for Progress alumni. Of course, the main reason I attended ICRA was to presented the accepted paper Topologically-Informed Atlas Learning. I had the opportunity to give a 4 minute technical talk in the Representation Learning session and discuss this work further at the subsequent poster session.

While at the conference I attended a variety of interesting technical talks. From our lab, Anthony Opipari gave a workshop talk on his work Differentiable Nonparametric Belief Propagation, and Emily Sheetz gave a technical talk on her publication Composable Causality in Semantic Robot Programming (to appear in the conference proceedings). My advisor Chad Jenkins also gave an invited workshop talk on how best to combine probabilistic reasoning with data-driven neural networks. Outside of our lab, I got very excited about Sampling Over Riemannian Manifolds Using Kernel Herding (Sandesh Adhikary, Byron Boots), and am feeling a growing need to further my understanding of kernel spaces. I spent an entire poster session asking questions about Geometric Fabrics: Generalizing Classical Mechanics to Capture the Physics of Behavior (Karl Van Wyk et. al.) -- although I can't say I fully understand the paper yet, I definitely have a much better handle on it than before. Finally, I enjoyed learning about an approach to learning kinematic models for articulated objects from Kineverse: A Symbolic Articulation Model Framework for Model-Agnostic Mobile Manipulation (Adrian Röfer et. al.)

The conference workshops were another immensely valuable experience. The conferences at ICRA were all day Monday and Friday (while Tuesday through Thursday were reserved for all other parts of the conference). On Monday, I spent practically the whole day at the 2nd Workshop on Representing and Manipulating Deformable Objects. My current research interests lie in the perception, representation, and manipulation of deformable objects, so I treated this workshop as my own tailor-made crash course in the field. The various talks and poster sessions were all excellent, but a few that really stood out to me were Learning Keypoints from Synthetic Data for Robotic Cloth Folding (Thomas Lips et. al.) and Real2Sim2Real: A Model for Deep Learning to Manipulate Deformable Objects (Ken Goldberg). For the Friday workshop, I mainly attended the workshop on Motion Planning with Implicit Neural Representations of Geometry. I'm really interested in the use of neural radiance fields (NeRFs) and similar neural models for representing the world. I was particularly interested in the workshop paper  Learning Prior Mean Function for Gaussian Process Implicit Surfaces (Tarik Kelestemur et. al.), and how the Gaussian process is able to explicitly model uncertainty for different regions of the object model. I think there's a lot of potential for this sort of uncertainty measure to be used for the active learning of object models, to enable robots to intelligently refine their understanding of the world.

I definitely hope to do a deeper dive into several of the papers presented at the conference, and maybe try my hand at re-implementing some of their work to improve my understanding. But that's all for now!


Comments

Post a Comment

Popular posts from this blog

Reimplementing GCS (Shortest Paths in Graphs of Convex Sets)

Image
In a previous blog post , I discussed my experiences reimplementing Iterative Regional Inflation by Semidefinite programming (IRIS) , which is an algorithm for finding large convex regions of collision free space (C-Free). In this blog post, I'll be discussing my experiences reimplementing shortest paths in Graphs of Convex Sets (GCS) , an algorithm for motion planning when the configuration space is described as the union of convex sets. Disclaimer: the intent is not to build a codebase for future use, since the Drake toolbox has mature implementations of both IRIS and GCS . Disclaimer: if you have not read the previous blog post , you should do so before reading this one. In this blogpost, I'll give a brief overview of the GCS paper, and then discuss my experiences reimplementing it. Before reading this blog post, you should read my previous blog post about reimplementing IRIS, as it provides some background and context which I will not repeat in this post. I have also ma...
Read more

Differing Formulations of GCS for Motion Planning

Image
In case you haven't heard of it before, GCS stands for "Graph of Convex Sets", and it is a framework for certain mixed integer convex optimization problems, primarily used for robot motion planning. I previously wrote a blog post on my experience implementing a basic version of it, which you can find here . I'd recommend reading that blog post before looking at this one. If you've read that previous blog post, you may recall that I described two different ways of transcribing the motion planning problem into a graph of convex sets. I implemented one version, which seems more faithful to the original GCS paper . However, the other transcription, presented in the planning-specific followup paper , has several benefits over the previous approach. In this blog post, I'll briefly present both methodologies, and explain their costs and benefits. Formulation 1: Conjugate Graph Fig. 1 of Shortest Paths in Graphs of Convex Sets The classic picture of GCS is the one sho...
Read more

IRIS on a Torus: Growing Convex Sets on Manifolds

Image
My current research focus at MIT is motion planning along manifolds, which we're primarily exploring through the lens of the graph of convex sets framework, abbreviated as GCS. (Refer to a previous blog post for more details, or check out the papers of Marcucci et. al. here and here .) Convexity is a property that is examined through the lens of Euclidean geometry, but it is possible to consider it in the broader context of smooth manifolds. This is reliant on the idea of a geodesic: a shortest path on a manifold. Replacing Euclidean spaces with manifolds and lines with geodesics allows us to consider a graph of geodesically convex sets (GGCS). We've begun laying the groundwork for planning over manifolds with such a tool, and you can read about it in our paper Non-Euclidean Motion Planning with Graphs of Geodesically-Convex Sets , to appear at RSS 2023. (You can read the preprint today on arXiv!) While GGCS assumes the geodesically convex sets are given, computing such set...
Read more