Cannot be more excited about my PhD student
@ShuchengK
's new work after spending his first year diving into moment and sums of squares relaxations.
Fast and Certifiable Trajectory Optimization
We have seen so many pendulum swing-ups, but I bet you haven't seen this one:
- The
Grad school applicants: my Computational Robotics group at Harvard SEAS is hiring PhD students in Fall 2023!
If you are interested in robotics, computer vision, machine learning, applied math, and their application in safe autonomy, you are the person I am looking for!
I'll join
@Harvard
as an Assistant Professor of Electrical Engineering in
@hseas
School of Engineering and Applied Sciences in Fall'23. I'll build a
#ComputationalRobotics
lab studying the algorithmic foundations of robot perception, action and learning. Contact me if interested!
Grad school applicants: this year I am particularly looking for a strong candidate in **statistics** and **machine learning** who is also interested in robotics applications. Please apply to Harvard and reach out if you’re interested!
Our Harvard Computational Robotics group has
- 2 papers accepted to
#L4DC2024
- 1 paper accepted to
#RAL
- 1 paper accepted to
#ICML2024
- 1 paper accepted to
#RSS2024
80% of these paper were authored by undergrad visitors last summer, who will go to the best PhD programs this
Super excited to share STRIDE, the first algorithm that can solve high-order tight semidefinite relaxations to high accuracy, which are extremely powerful global optimization tools for general polynomial optimizations but notoriously difficult to solve.
What is the true optimal value function of the infinite-horizon inverted pendulum problem?
Despite being a simple nonlinear optimal control problem, its optimal value function remains a mystery until
@Hyhan0118
uncovered the answer.
a thread (1/n)
It’s such an honor to win the best paper award in robot vision at
@icra2020
! Many thanks go to my co-authors
@AntonanteP
@VassilisTzoumas
and
@lucacarlone1
for their amazing feedback in both writing a great paper and doing a great presentation!
Video:
So excited our paper "Certifiably Optimal Outlier-Robust Geometric Perception" is accepted to IEEE TPAMI!
It unifies my PhD work in designing tractable geometric estimation algorithms with optimality guarantees, in outlier-contaminated realistic setups.
We are organizing a Frontiers of Optimization for Robotics workshop in the upcoming Robotics: Science and Systems conference at Netherlands.
A great lineup of speakers covering four diverse topics: Geometry & Global Optimality, Speed & Scale, Optimization through Contact, and
Super excited to share new work “TEASER: Fast and Certifiable Point Cloud Registration” with Jingnan Shi and
@lucacarlone1
Paper:
Code:
TEASER is the first algorithm of its kind in many practical and theoretical aspects:
Our group () led by
@drmapavone
is looking for PhD research interns for next year. If you’re excited about 3D deep learning, motion planning, and control for robotics/autonomous driving, please consider applying to our group!
Training deep
#ReinforcementLearning
agents can be unstable. This instability gets magnified in a Lifelong Learning setup, leading to the intriguing phenomenon called Loss of Plasticity (Negative Transfer), where the agent gradually loses the ability to adapt to new information
⭐New Paper Alert ⭐
How can your
#RL
agent quickly adapt to new distribution shifts ? And without ANY tuning?🤔
We suggest you get on the Fast TRAC🏎️💨, our new Parameter-free Optimizer that surprisingly works. Why?
Website:
1/🧵
Super honored to appear on today's MIT Spotlight! Thank you
@MIT
for sharing our work on
#CertifiablePerception
and our vision about a future with safe robots and trustworthy autonomy.
I am deeply grateful to
@lucacarlone1
for the tremendous support!
Proud of
@YuXihang
(undergrad from UM)’s intern project at Harvard!
SIM-Sync, Certifiably Optimal camera trajectory estimation directly from 2D images with learned depth.
2D kpts + learned depth —> scaled point clouds —> sync over 3D similarity group
#RSS2021
check out
- "Optimal Pose and Shape Estimation for Category-level 3D Object Perception" (w/
@jingnanshi
,
@lucacarlone1
) in poster session I (ET 11:15AM tmr!) and IV
- "Certifiable Outlier-Robust Machine Perception" in the RSS Pioneers poster session 0 (ET 10AM tmr!)
"Fast and Certifiable Trajectory Optimization" just got accepted to the International Workshop on the Algorithmic Foundations of Robotics (WAFR)
@wafr_conf
, great job
@ShuchengK
!
Cannot be more excited about my PhD student
@ShuchengK
's new work after spending his first year diving into moment and sums of squares relaxations.
Fast and Certifiable Trajectory Optimization
We have seen so many pendulum swing-ups, but I bet you haven't seen this one:
- The
Computer vision friends: is there a large-scale structure from motion dataset (e.g., over 2000 frames) whose images are labelled with keypoint correspondences?
Was looking at the building Rome in a day dataset but it seems no keypoints are provided.
I have got many questions about how to use TEASER++ for 3D registration in practice. Finally, I wrote a step-by-step tutorial about using FPFH and TEASER++ on the 3DMatch dataset, in our favorite language Python:
Our ICRA 2020 paper on graduated non-convexity for robust spatial perception has been implemented by Matlab navigation toolbox and featured by
@MathWorks
news! If you haven’t read the paper, here it is:
Self-supervised Geometric Perception, with W. Dong,
@lucacarlone1
, V. Koltun, is accepted as
#CVPR2021
Oral. Appreciated the discussion w/
@ducha_aiki
on difference b/w SGP and reconstruction-based supervised learning. Check out future research on Page 8:
Presenting "Fast and Certifiable Approximation of Pose Uncertainty Sets" tomorrow at the INFORMS Optimization Society Conference, a series of two works from my group.
Given a nonconvex pose uncertainty set (PURSE), how to approximate its minimum enclosing (geodesic) ball with a
In an era of rapid robotics demos, maybe we should get together in a workshop, revisit some of the “old problems” we used to care about, and ask how do our new methods work on the old problems?
🧲3/3
The answer: sort of... but not really.
Control problems require continuous adjustments for stability, which the diffusion policy, being a less-than-optimal control approximation, could not maintain.
Could
@ChaoyiPan
's Model-based Diffusion work?
Join me 10am ET 11/21!
Super excited to talk about new work on computing the minimum enclosing ellipsoid of set-membership estimation in control and perception, via convex optimization (of course).
Examples from system identification and object pose estimation will be presented.
Talk 2:
Speaker: Prof.
@hankyang94
,
@Harvard
Title: Revisiting the Minimum Enclosing Ellipsoid of Set-Membership Estimation in Control and Perception
When? November 21 2023, 16:00 CET
Where?
#autonomytalks
In spotting objects amid clutter, how can a machine tell the difference between a success and a failure, and correct its failure, if at all possible? Certifiable Outlier-Robust Geometric Perception, enabling robots to see the world with confidence.
Paper:
Congrats to my incoming PhD student Shucheng Kang on getting his paper accepted to IEEE CDC 2023!
I believe this multilevel polynomial optimization formulation has promise beyond verification and synthesis of robust control barrier functions!
The sparse Moment-SOS hierarchy just got even more powerful -- handle unbounded sets!
With this, we solve optimal control of the unstable Van der Pol oscillator to certifiable global optimality!
Great work led by Lei,
@ShuchengK
, and Jie!
Was a busy week but had a lot of fun hosting Jean-Bernard Lasserre at Harvard. Such a nice, energetic, and fun person! I hope the wonton with chili sauce was good :)
Glad to be a finalist of MIT's first Research SLAM featuring 3-Minute Thesis! What a broad range of cool science!
#CertifiablePerception
Come join the public showcase at 5PM EDT Monday 29th!
NVIDIA Autonomous Vehicle Research group resumed hiring! Check out this opportunity if you’re interested in autonomous driving, robotics, foundation models etc.
If you are attending
#CVPR2020
, check out our paper: In Perfect Shape: Certifiably Optimal 3D Shape Reconstruction From 2D Landmarks. I will be holding two live Q&A sessions in EDT: 1-3PM June 16th, and 1-3AM June 17th (3AM is crazy!) Join me in discussing
#CertifiablePerception
!
Our paper "Graduated Non-Convexity for Robust Spatial Perception: From Non-Minimal Solvers to Global Outlier Rejection" has been nominated as Best Paper Award Finalist in Robot Vision at
#ICRA2020
. GNC is the new RANSAC, check out the paper here:
At
#RSS2024
?
Check out three works from the
#HarvardComputationalRobotics
lab presented at the main conference and workshops:
- CLOSURE: GPU-accelerated fast end-to-end uncertainty quantification for 6D pose estimation, main conference "Perception" session
- STROM: Fast and
Excited to share a new preprint:
Verification and synthesis of ROBUST control barrier functions for control-affine polynomial systems with bounded state-dependent additive uncertainty and convex polynomial control constraints.
Three key techniques👇
@gabrielpeyre
This is interesting, I think you will like this similar idea called graduated non-convexity that solves non-convex robust estimation using a sequence of surrogate functions.
With
@imZhiyuZ
's expertise in adaptive online learning, we are initiating an exciting research thread in our group "lifelong learning amid distribution shifts" and bridging the theory there to robot learning applications.
In our ICML paper, Zhiyu revisited the notion of a
A short thread about our ICML paper, joint work with David Bombara and Heng Yang (
@hankyang94
).
tldr: Follow the Regularized Leader (FTRL), as a well-known gradient descent replacement in online learning, is particularly useful for lifelong learning and conformal prediction.
I’d like to share an interesting dynamical perspective on point cloud registration, where Newton-Euler dynamics and Lyapunov theory are adopted to analyze nontrivial convergence under virtual springs and damping. Glad to see F=ma still shines.
Excited to share the paper with
@yukai_tang02
(an amazing undergrad!) and Jean Lasserre on revisiting the computational challenges of minimum enclosing ellipsoids:
The presentation I gave at Autonomy Talks:
Join me 10am ET 11/21!
Super excited to talk about new work on computing the minimum enclosing ellipsoid of set-membership estimation in control and perception, via convex optimization (of course).
Examples from system identification and object pose estimation will be presented.
TEASER is now an accepted paper in Transactions on Robotics! Gratitude goes to coauthors
@jingnanshi
,
@lucacarlone1
, reviewers, and many others who read, tried and gave feedback. Stay tuned for more
#CertifiablePerception
.
Paper:
Code:
Graduated Non-Convexity (GNC) is the counterpart for RANSAC: while RANSAC robustifies minimal solvers, GNC robustifies non-minimal solvers. The intriguing duality for robust estimation: {Consensus Maximization, Minimal Solver, RANSAC} and {M-estimation, Non minimal Solver, GNC}😀
GNC is the new RANSAC: our paper on robust perception via Graduated Non-Convexity and non-minimal solvers has been accepted on RA-L! congrats to Yang, Pasquale, and Vasileios
#mitSparkLab
Excited to present recent work with
@drmapavone
that applies conformal prediction to obtain probabilistically correct object pose estimation with uncertainty quantification at ECCV Workshop on 3D Perception for Autonomous Driving. See attached image for a sneak peak!
Before that, I will be a Research Scientist in the Autonomous Vehicles Research Group
@NVIDIAAI
led by
@MarcoPavoneSU
starting July 2022. Deep thanks to all the people who supported me, most importantly, my amazing advisor
@lucacarlone1
!
Join us for the [HYBRID] Fall 2021 GRASP Seminar: Heng Yang, Massachusetts Institute of Technology , “Certifiable Outlier-Robust Geometric Perception: Robots that See through the Clutter with Confidence” - This Wednesday (12/15) @ 2:00pm in Levine 307 and via Zoom!
Our GNC paper got a RAL best paper honorable mention in addition to the ICRA best paper award in robot vision last year! I still remember that day, the four of us looking at the same monitor, polishing the paper WORD by WORD. Good paper really pays back!
It was wonderful to have
@tatjunchin
talk about “Quantum Robust Fitting” to my group and visit Harvard SEAS. One of the most fun and educative talks I have heard in a while. I didn’t know Boscovich discovered Earth is oblate using robust fitting!
#RSS2021
check out
- "Optimal Pose and Shape Estimation for Category-level 3D Object Perception" (w/
@jingnanshi
,
@lucacarlone1
) in poster session I (ET 11:15AM tmr!) and IV
- "Certifiable Outlier-Robust Machine Perception" in the RSS Pioneers poster session 0 (ET 10AM tmr!)
We are organizing an
#ACC2023
Workshop on Safe & Robust Learning for **Perception-based Planning and Control**, consider submitting your best work to our workshop👇 deadline May 5th, 2023
Dynamics and Perception? With C. Doran and J.-J. Slotine, we propose DynAMical Pose estimation (DAMP), solving five pose estimation problems by simulating rigid body dynamics from virtual springs and damping:
Paper:
Video:
Our RSS'21 work shows (i) Certifiable algorithms can be fast; (ii) Robust estimation is general; (iii) Once you get the math correct, everything should work like a charm even on real nasty data! Check out category-level object pose and shape estimation
Traditional approaches for category-level object pose and shape estimation are sensitive to outliers and get stuck in local minima. We propose the first approach that avoids local minima and is robust to 70-90% outliers:
#computervision
#mitsparklab
Exciting to be presenting our recent work on PAC-Bayes generalization guarantees for conformal prediction at
#NeurIPS23
! Come by our poster (
#1818
) on Wednesday at 5pm to chat if you're around. Read on to learn more: 1/6
Sharing the work that I am most proud of: STRIDE for solving rank-one semidefinite relaxation of polynomial optimization. We handle inexactness, prove global convergence, design a modified L-BFGS, and solve important optimization in math and engineering:
#CVPR2021
Come check out "Self-supervised Geometric Perception" during paper session 11 at 10PM-12:30AM June 24 EDT (tomorrow night), or just have some casual chat!
Self-supervised Geometric Perception, with W. Dong,
@lucacarlone1
, V. Koltun, is accepted as
#CVPR2021
Oral. Appreciated the discussion w/
@ducha_aiki
on difference b/w SGP and reconstruction-based supervised learning. Check out future research on Page 8:
With
@lucacarlone1
, we present Shape*, the first certifiably optimal non-minimal solver for 3D shape reconstruction from 2D landmarks in a single image, and Shape#, a robust shape reconstruction algorithm that tolerates 70% outliers.
Life winds up in ways we can never expect, the best thing to do is to keep integrating.
10 yrs ago I started college in Tsinghua. In no way I expected my Lorentz cone would lead me to MIT.
Hope we all find thrills walking inside (maybe outside) the cone.
Happy New Year🎆
Attending
#ICRA2022
next week. Can’t believe this will be my FIRST in-person ICRA! Look forward to hearing cool projects and meeting friends, old and new!
SEAS welcomes 10 new faculty. Their areas of expertise include robotics, bioengineering, artificial intelligence, machine learning, and quantum engineering.
Proud of this new preprint that proposes a general framework to design
#certifiably
robust geometric perception algorithms that are able to decide whether or not outliers have been correctly rejected. I believe that algorithms with guarantees will safeguard robot perception.
I had a lot of fun talking about
#CertifiablyRobustPerception
with
#outliers
in both SPARK Lab and Marine Robotics Group at MIT. Today I also had the chance to present to a broader audience at our
#RSS2020
tutorial. Check out the video here:
At
@NVIDIAGTC
, I presented my group's strategy on leveraging foundation models (FMs) to develop next-gen autonomous vehicles. Slides: Recording: Three pillars: standing up the FMs, using them within an AV program, and AI safety.
Honda, BD and Agility for sure make cooler humanoids. But Tesla, a public company with a spirit of bravery, promises a $20K humanoid definitely brings unprecedented hope and hype. We need hope and with correct feedback from roboticists perhaps we can do it differently this time.
Thanks to Wei Dong, a clean and general implementation of "Self-supervised Geometric Perception" [CVPR2021] is now released, with detailed instructions.
Try it on 3D registration, camera relative pose estimation, and other problems of interest.
Self-supervised Geometric Perception, with W. Dong,
@lucacarlone1
, V. Koltun, is accepted as
#CVPR2021
Oral. Appreciated the discussion w/
@ducha_aiki
on difference b/w SGP and reconstruction-based supervised learning. Check out future research on Page 8:
I spent 2+ hours googling how to embed fonts into a pdf on Mac to pass IEEE PDF eXpress check, and none worked ... Feeling desperate, I opened Preview and exported the pdf with "Create PDF/A" and it worked ... just in case you also suffered from this ...
@ducha_aiki
@wdong397
@lucacarlone1
@Jimantha
Thanks Dmytro for the swift review, hours after we uploaded the paper! The difference between our method SGP and 3D reconstruction-based supervised training has been discussed many times among ourselves, and the difference can be both subtle and deep. Here is my rebuttal.
Very happy to be a coauthor of
@jingnanshi
's nice work about ROBIN, which applies invariance and graph theory to prune outliers in robot perception (like a charm)! Looking forward to exploring more applications of ROBIN.
do you want to boost the robustness of your RANSAC/robust estimator? Check this preprint showing how to solve problems with >90% outliers using max clique and invariance: we prune outliers in milliseconds!
#mitsparklab
#robustPerception
#computerVision
Grateful for the enlightening discussion on causal RL with
@hankyang94
and the amazing group of peers at Harvard SEAS! 🙌 Thanks for the wonderful accommodation and the lunch at HBS 🙈
Flying to Oxford tomorrow for
#L4DC
to present the "sister paper" of CLOSURE, where we use SOS relaxations to find outer approximations of uncertainty sets!
How to apply:
(1) Choose an area (e.g., EE, CS, AM) for your PhD application (e.g., )
(2) Submit the official application (e.g., )
(3) List me as a potential faculty to work with in the application
Thanks and best of luck!
If you have polynomial optimizations in your applications and you care about certifiable global optimality. Let us know.
Deeply grateful to collaborators from NUS who are world-renowned SDP experts!
our outlier-robust algorithms for SLAM and geometric perception are now available in GTSAM and MATLAB: pointers to code after the abstract here: . see also:
#mitSparkLab
#robustPerception
My cat Cola was born in NYC. So
#dalle
created a portrait for him! “A black and white British shorthair drinking Coca-Cola in New York City, digital art” Fun fact is if I change NYC to Boston it does not work. Should collect some Boston training data :)
New SPARK pre-print shows how to simultaneously do robust estimation and learn the inlier noise statistics (among many other cool results) - check out our minimally-tuned algorithms: .
#mitsparklab
#robustEstimation
#robotperception
Feel that reviewers are so good at picking common weaknesses (novelty and experiments). Do they even read the paper carefully enough to understand the novelty and significance that authors want to say through the experiments? A good researcher should learn to appreciate, too.
That time of year where I wonder why we all stake our careers on a glorified random number generator whose output we regard as a respected mark of quality. \end{rant}
Introducing EmerNeRF, our answer to the challenging dynamic NeRF in-the-wild problem. EmerNeRF is the best-ever project I've got involved in, led by our only
@JiaweiYang118
(stay tuned for his even more impressive works), in collaboration with
@NVIDIAAI
colleagues
@iamborisi
Many people have tried TEASER and I am glad to see TEASER works great for their applications! New demos and examples will make TEASER even easier to use 😃
Our paper on robust spatial perception is a finalist for the Best Paper Award in Robot Vision at ICRA 2020! The paper provides a general framework to extend newly developed global solvers for 3D vision to problems with outliers.
#mitSparkLab
Excited to attend ICCOPT and present STRIDE for solving rank-one semidefinite relaxations of polynomial optimization problems, with applications in robot perception!
Session: Recent Developments in Solving Structured Semidefinite Programs
When: Wednesday July 27th 2:20pm
@antoine_leeman
@Modern_Gangster
For trajectory optimization problems that can be written as polynomial optimization problems (e.g., using Lie group variational integrators), one can apply a sparse moment-SOS relaxation that is empirically tight at the second order (as shown by a paper last RSS). We make this
We are very excited to host
#RSSPioneers21
@RoboticsSciSys
Scroll into our Pioneeers' research statements👉
Get to know our rising stars and interact with them at the main
#RSS2021
interactive poster session on Monday 👉
In this preprint,
@jingnanshi
and
@hankyang94
investigate category-level object pose and shape estimation from 2D and 3D keypoints and propose an elegant and general framework for graph-theoretic outlier rejection:
#mitSparkLab
(i) A MULTILEVEL polynomial optimization formulation
(ii) Reduction to single-level and min-max polynomial optimization via KKT conditions
(iii) Single-stage and two-stage semidefinite relaxations with asymptotic global convergence
w/ Shucheng Kang,
@Yuxiao_Chen_
,
@drmapavone
The punchline:
With clever loss functions, we can distill the optimal value function into a simple neural network.
The neural value function induces a controller that also globally swings up and stabilizes the pendulum!
So proud of Haoyu!
(n/n)