I’ll be joining GaTech
@gtcomputing
@ICatGT
as an Assistant Professor in Fall 2022!
Looking forward to continuing my work in Robot Learning as a faculty and collaborating with researchers & students at GTCS
@GTrobotics
@mlatgt
. Reach out for collaborations / joining the lab!
Very proud of my student
@danfei_xu
(co-advised with
@silviocinguetta
) for his wonderful PhD thesis defense today! Danfei’s work in computer vision and robotic learning pushes the field forward towards enabling robots to do long horizon tasks of the real world. 1/2
I'm recruiting! If you are excited about teaching robots to perceive, reason about, manipulate, and move around everyday environments, apply the CS Ph.D program at GT (Interactive Computing) and mention my name. Apps from underrepresented groups in AI&Robo are especially welcome!
A bit more formally: I'm hiring Ph.D. students in Robot Learning this year!
If you are excited about the future of data-driven approaches to robotics, apply through the School of Interactive Computing at
@gtcomputing
by Dec 15th.
I’ll be joining GaTech
@gtcomputing
@ICatGT
as an Assistant Professor in Fall 2022!
Looking forward to continuing my work in Robot Learning as a faculty and collaborating with researchers & students at GTCS
@GTrobotics
@mlatgt
. Reach out for collaborations / joining the lab!
Since we are entering the "BC is all you need" phase of Robot Learning😜 --- Robomimic () allows you to play with SOTA algorithms (BC-Transformer, DiffusionPolicy, etc.) on challenging tasks. Also easy to integration with physical robots!
I often get this question: Is LLM all you need for robot planning?
I'd go: "obviously not, because you need to consider physical constraints, dynamics, ... ", which then turn into a non-stop rant.
Now I'll just point them to this paper 😎
This is clearly going to benefit the privileged. Even the info that this conference/track existed probably will only circulate in a small group with direct tie to academia/tech (parents etc). How about we flip this into a track for creating accessible tutorials, lectures,…
This year, we invite high school students to submit research papers on the topic of machine learning for social impact! See our call for high school research project submissions below.
Excited to share Generalization Through Imitation (GTI)! GTI learns visuomotor control from human demos and generalizes to new long-horizon tasks by leveraging latent compositional structures.
Joint w/
@AjayMandlekar
@RobobertoMM
@silviocinguetta
@drfeifei
Super neat system! It seems that Chinese robotics startups have everything they need to quickly iterate on capable & low-cost hardware. Will US startups be able to compete? Chaining together dynamixals/off-the-shelf motors likely won’t cut it…
Can't believe that I just came across this insanely cool paper. 3D gaussian seems to be such an intuitive representation to model large & dynamic scenes (Lagrangian vs. Eularian). Expect it to drive a whole new wave of dense/obj-centric representation w/ self-supervision.
Dynamic 3D Gaussians: Tracking by Persistent Dynamic View Synthesis
We model the world as a set of 3D Gaussians that move & rotate over time. This extends Gaussian Splatting to dynamic scenes, with accurate novel-view synthesis and dense 3D trajectories.
Putting the promise of AI directly in students’ hands: We’re powering up the Georgia Tech AI Makerspace - a student-focused AI supercomputer hub. Proud to work with
@nvidia
and
@WeAre_Penguin
to make this a reality on campus for our students.
New preprint!
Affordance is a versatile repr. to reason about interactions in a complex world. But it is also *myopic*, because it only means that an action is feasible instead of leading to a long-term goal. How can we use affordances to plan for long-horizon tasks? 1/
Excited to share Generalization Through Imitation (GTI)! GTI learns visuomotor control from human demos and generalizes to new long-horizon tasks by leveraging latent compositional structures.
Joint w/
@AjayMandlekar
@RobobertoMM
@silviocinguetta
@drfeifei
We present Regression Planning Network (RPN), a type of recursive network architecture that learns to perform high-level task planning from video demonstrations.
#NeurIPS2019
(1/3)
Applying imitation learning to real world problems takes more than new algorithms. We are organizing a workshop "Overlooked Aspects of Imitation Learning: Systems, Data, Tasks, and Beyond” at RSS22! Exciting speakers & more to come. Submit by May 7th!
🤖 Inspiring the Next Generation of Roboticists! 🎓
Our lab had an incredible opportunity to demo our robot learning systems to local K-12 students for the National Robotics Week program
@GTrobotics
. A big shout-out to
@saxenavaibhav11
@simar_kareer
@pranay_mathur17
for hosting…
It's a strange time to share this but I'll be co-instructing the Stanford CS231n course next quarter! Now that all courses are pass/fail, we might experiment w/ something new😃
Suggestions / tips on online lecturing are appreciated!
One of the most impressive CV works I've seen recently. Also huge kudos to Meta AI for sticking to open sourcing despite the trend increasingly going towards the opposite direction.
Today we're releasing the Segment Anything Model (SAM) — a step toward the first foundation model for image segmentation.
SAM is capable of one-click segmentation of any object from any photo or video + zero-shot transfer to other segmentation tasks ➡️
We also made a similar transition to ROS-free. The non-obvious thing is that modern NN models (BC policies, VLMs, LLMs) breaks the abstraction of ROS modules. Raw sensory stream instead of state estimation, actions instead of plans, etc. Need new ROS for the next-gen modules.
Interesting (and sad) result here; I really had hoped more people would be able to just run with ROS2. But it seems like it's not quite there, if this is in any way worth doing for a small company/fast-moving startup that should be the target audience.
Presenting two papers at
#NeurIPS2019
! Come say hi if you are around.
1. Regression Planning Networks
We combine classic symbolic planning and recursive neural network to plan for long-horizon tasks end-to-end from image input.
Paper & Code:
1/
Our group headed by
@MarcoPavoneSU
at NVIDIA Research is hiring fulltime RS and interns! Tons of cool problems in planning, control, imitation, and RL. Job posting 👇
Intern:
Full-time:
In 2010 I was a high school senior in Shanghai. I cold-called a company making educational robots and started my first internship in robotics. Almost a decade later, I’m doing a Ph.D. at Stanford, still in robotics, still happy. Let’s see where the next decade leads me.
Annnnd that's a wrap! First semester teaching at GT and it's been an absolute blast. Really happy to see the progression of the student projects and the final poster session joined by ~170 students. Couldn't have made it without my awesome TAs. Thanks
@mlatgt
for the sponsorship!
Detail: 10hz image -> 200hz EEF control. I'm guessing keep the same image token for 20 steps while updating proprio state? Also given how smooth the motion looks --- high-quality OSC implementation?
Finally, let's talk about the learned low-level bimanual manipulation.
All behaviors are driven by neural network visuomotor transformer policies, mapping pixels directly to actions. These networks take in onboard images at 10hz, and generate 24-DOF actions (wrist poses and…
First work coming out of my lab at GT! LEAGUE is a "virtuous cycle" system that combines the merit of Task and Motion Planning and RL. The result is continually-learning and generalizable agents that can carry their knowledge to new task and even environments.
Introducing LEAGUE - Learning and Abstraction with Guidance! LEAGUE is a new framework that uses symbolic skill operators to guide skill learning and state abstraction, allowing it to solve long-horizon tasks and generalize to new tasks and domains. Joint work with
@danfei_xu
1/6
Super excited about this new
#CoRL2023
work on compositional planning! We introduce a new generative planner (GSC) to compose skill-level diffusion models to solve long-horizon manipulation problem, without ever training on long-horizon tasks.
@ICatGT
@GTrobotics
@mlatgt
How to enable robots to plan and compositionally generalize over long-horizon tasks?
At
#CoRL2023
, we introduce Generative Skill Chaining (GSC), a diffusion-based, generalizable and scalable approach to compose skill-level transition models into a task-level plan generator.(1/7)
Introducing Sora, our text-to-video model.
Sora can create videos of up to 60 seconds featuring highly detailed scenes, complex camera motion, and multiple characters with vibrant emotions.
Prompt: “Beautiful, snowy…
Among so many thoughtful & nuanced discussions on regulating AI, the EU chooses to "mitigate the risk of extinction from AI"... This is some sort of joke, right?
Mitigating the risk of extinction from AI should be a global priority.
And Europe should lead the way, building a new global AI framework built on three pillars: guardrails, governance and guiding innovation ↓
We are organizing the Deep Representation and Estimation of State tutorial at the virtual IROS2020!
Fantastic speaker line-up:
@leto__jean
, Yunfei Bai, and
@ChrisChoy208
. Co-organized with
@KuanFang
and
@deanh_tw
.
A short thread about each session👇
Do you really need legs? We don't think so. As much we love anthropomorphic humanoids (our co-founder built one in 9th grade), we believe virtually all menial tasks can be done with two robot arms, mounted on wheels. In our view,
@1x_tech
's Eve robot is the optimal form factor…
T minus 2 hours until we begin our next
#DARPAForward
event
@GeorgiaTech
.
@DoDCTO
Heidi Shyu will kick off a packed agenda featuring experts on pandemic preparedness, cybersecurity, and more. Visit our page for more on how you can join future events:
Interesting trend in AI: the best results are increasingly obtained by compound systems, not monolithic models.
AlphaCode, ChatGPT+, Gemini are examples.
In this post, we discuss why this is and emerging research on designing & optimizing such systems.
Robotics dataset is expanding at an unprecedented pace. How do we control the quality of the collected data? Our
#CoRL2023
work presents an offline imitation learning method that learns to discern (L2D) data from expert in a mixed-quality demonstration dataset. Code coming soon!
Introducing our
#CoRL2023
work Learning to Discern (L2D)! As robotics datasets grow, quality control becomes ever more important. L2D is our solution for handling mixed-quality demo data for offline imitation learning. (1/6)
Interested in playing around with RL? We’re happy to announce the release of Acme, a light-weight framework for building and running novel RL algorithms.
We also include a range of pre-built, state-of-the-art agents to get you started. Enjoy!
We are excited to announce the release of Traffic Behavior Simulation (TBSIM) developed by the Nvidia Autonomous Vehicle research group (), which is our software infrastructure for closed-loop simulation with data-driven traffic agents. (1/7)
We are organizing a workshop on imitation learning at RSS2020 ()! The workshop will bring together well-known researchers in field. CfP includes short-length, full-length, and position papers. Tentative submission deadline Apr 9th. RT and spread the word!
Object representation is a fundamental problem for robotic manipulation. Our
#CoRL2023
work found that *density field* can efficiently represent state and dynamics of non-rigid objects such as granular material. To be presented as a spotlight&poster on Thursday!
How to represent granular materials for robot manipulation?
Introducing our
#CoRL2023
project: Neural Field Dynamics Model for Granular Object Piles Manipulation, a field-based dynamics model for granular object piles manipulation.
🌐
👇 Thread
Join us on Sunday at 9:00-1:30pm PT for the Advances & Challenges in Imitation Learning for Robotics
#RSS2020
Workshop: with an exciting list of speakers! Live streaming at
A thread by my awesome co-instructor
@RanjayKrishna
recapping
@cs231n
for the past quarter. It happens to be the *largest* class on campus for the quarter!
Thanks all the teaching staff, especially our head TA
@kevin_zakka
for making this course possible!
Academic quarter recap: here's a staff photo after the last lecture of
@cs231n
. It's crazy that we were the largest course at Stanford this quarter. This year, we added new lectures and assignments (open sourced) on attention, transformers, and self-supervised learning.
Gearing up for the conference next week, check this interactive feature as you prep for your time at the conference. Discover cool papers and insights.
Did you know that we have 199 contributed papers from 873 authors originating in 25 countries! 🤯
We present 6-PACK, an RGB-D category-level 6D pose tracker that generalizes between instances of classes based on a set of anchors and keypoints. No 3D models required! Code+Paper: w/ Chen Wang
@danfei_xu
Jun Lv
@cewu_lu
@silviocinguetta
@drfeifei
@yukez
It’s that time of the year - first lecture of
@cs231n
!! It’s the 9th year since
@karpathy
and I started this journey in 2015, what an incredible decade of AI and computer vision! Am so excited to this new crop of students in CS231n! (Co-instructing with
@eadeli
this year 😍🤩)
Bay Area Robotics Symposium (BARS) will be happening in person this Friday on October 29!
The registration will close on October 27th, 5 p.m.
Register here:
Program:
Check out our new work on imitation learning from human demos! We released a set of sim&real tasks, demo datasets, and a modular codebase & clean APIs to help you develop new algorithms!
Robot learning from human demos is powerful yet difficult due to a lack of standardized, high-quality datasets.
We present the robomimic framework: a suite of tasks, large human datasets, and policy learning algorithms.
Website:
1/
We are releasing our
#ICCV2019
work on goal-directed visual navigation. We introduced a method that harnesses different perception skills based on situational awareness. It makes a robot reach its goals more robustly and efficiently in new environments.
Blog post by
@deanh_tw
and I summarizing our line of work on generalizable imitation of long-horizon tasks: Neural Task Programming, Neural Task Graphs, and Continuous Relaxation of Symbolic Planner. Enjoy!
What if we can teach robots to do new task just by showing them one demonstration?
In our newest blog post,
@deanh_tw
and
@danfei_xu
show us three approaches that leverage compositionality to solve long-horizon one-shot imitation learning problems.
To carry out long-horizon tasks, robots must plan far and wide into the future. What state space should the robot plan with, and how can they plan for objects & scenes that they have never seen before? See 👇for our new work on Generalizable Task Planning (GenTP).
1/ Can we improve the generalization capability of a vision-based task planner with representation pretraining?
Check out our RAL paper on learning to plan with pre-trained object-level representation.
Website:
Excited to share our milestone in building generalizable long-horizon task solvers at
#CoRL2023
! As part of our long-term vision for a never-ending data engine for everyday tasks, HITL-TAMP combines the best of structured reasoning (TAMP) and end-to-end imitation learning.
How can humans help robots improve? Introducing Human-In-The-Loop Task and Motion Planning (HITL-TAMP), a perpetually-evolving TAMP system that learns visuomotor skills from human demos for contact-rich, long-horizon tasks.
#CoRL2023
Website:
1/
Learning for high-precision manipulation is critical to bridge *intelligence* to repeatable *automation*. C3DM is a diffusion model that learns to remove noise from the input by "fixating" on the target object. To be presented at the Deployable Robot workshop at
#CoRL2023
today!
Introducing C3DM 🤖 - a Constrained-Context Conditional Diffusion Model that solves robotic manipulation tasks with:
✅ high precision and
✅ robustness to distractions!
👇 Thread
If you're a hardware biz or R&D lab in Silicon Valley, you should definitely be keeping your eye on the liquidation auctions, which are on fire right now
This one is auctioning off more than 100 new and used Kuka robot arms:
Fantastic research led by Chen! Continuing our work on hierarchical imitation starting for real-world long-horizon manipulation. It turns out that we can train high-level planner directly from *human video*. This greatly reduces need for on-robot data and improves robustness 1/2
How to teach robots to perform long-horizon tasks efficiently and robustly🦾?
Introducing MimicPlay - an imitation learning algorithm that uses "cheap human play data". Our approach unlocks both real-time planning through raw perception and strong robustness to disturbances!🧵👇
Data fuels the progress in robotics, whether it's sim, real teleoperated, or auto-generated. Our workshop at
#RSS2024
will bring together researchers from academia, industry, and startups around the world to share insights🧐 and hot takes 🔥.
Data is the key driving force behind success in robot learning. Our upcoming RSS 2024 workshop "Data Generation for Robotics” will feature exciting speakers, timely debates, and more! Submit by May 20th.
Very nice post!
Slightly different take: Scaling up should be the **question**, not the answer. Yes we need to scale up to more task, envs, robots, but there should be many possible answers to this question. Training on lots of data may be an answer but should not the only one.
There was a lot of good and interesting debate on "is scaling all we need to solve robotics?" at
#CoRL23
. I spent some time writing up a blog post about all the points I heard on both sides:
Our paper on learning generalizable neural programs for complex robot tasks will appear in
#icra2018
! See you soon. Arxiv: Two minutes paper: , Video:
Trying to better understand contrastive learning: Intuitively, contrastive learning relies on dense pos/neg sample coverage. SimCLR & others increase coverage using image augmentation. But how dense does the space have to be & what about spaces that cannot be augmented easily?
As the Deep Learning course at GT draws to a close this semester, I'd like to extend a heartfelt thanks to
@WilliamBarrHeld
. His exceptional lecture and programming assignment on Transformers and LLMs were truly enlightening. Don't miss out on these incredible resources!
For
@danfei_xu
's Deep Learning course this semester, I made a homework for Transformers and gave a lecture on LLMs.
I'm sharing resources I made for both in hopes they are useful for others!
Lecture Slides:
HW Colab:
Excited about hierarchy, abstraction, model learning, skill learning, planning with LLMs, and benchmarking long-horizon manipulation tasks? Submit a paper to our learning for Task and Motion Planning (L4TAMP) workshop at RSS'23!
We are organizing the RSS’23 Workshop on Learning for Task and Motion Planning
Contributions of short papers or Blue Sky papers are due May 19th, 2023.
Delighted to present our recent work on hierarchical Scene Graphs for neuro-symbolic manipulation planning. We use 3D Scene Graphs as an object-centric abstraction to reason about long-horizon tasks. w/
@yifengzhu_ut
, Jonathan Tremblay, Stan Birchfield
We present 6-PACK, an RGB-D category-level 6D pose tracker that generalizes between instances of classes based on a set of anchors and keypoints. No 3D models required! Code+Paper: w/ Chen Wang
@danfei_xu
Jun Lv
@cewu_lu
@silviocinguetta
@drfeifei
@yukez
The OPT program is crucial for retaining talented international students in the US. I relied on the OPT myself for summer internships during college and for full-time work after graduation.
This is a great effort to collect large robot dataset on standardized hardware setup! Also happy to see that Robomimic is adopted as the core policy learning infrastructure.
After two years, it is my pleasure to introduce “DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset”
DROID is the most diverse robotic interaction dataset ever released, including 385 hours of data collected across 564 diverse scenes in real-world households and offices
Why is learning object-centric representation important for RL/robot learning? If it is merely a form of state dim reduction, and the only useful info it provides is 3D pose / 2d bbox & object appearance, then shouldn't ppl focus on better pose estimator / detector?
The original Perceptual Symbol Systems article and its commentary & author responses are truly goldmines of refs on neural symbolic research & its cogsci background. Also intrigued to find hints of many modern DL ideas in the article such as ...
Our
#ECCV2020
paper is now on arXiv. We show that 3D object tracking emerges automatically when you train for multi-view correspondence. No object labels necessary!
Video: results from KITTI. Bottom right shows a bird's eye view of the learned 3D features.
Can we train visuomotor policies for real-world long-horizon tasks AND generalize across tasks? Join us Wed 8-10am virtually at
#RSS2020
(Paper
#61
) for a live discussion with
@AjayMandlekar
,
@RobobertoMM
, and myself. See 👇 for a short thread about our paper.
Human demonstrations are often used to teach robots new tasks, but how can we achieve generalization when learning by imitation?
Check out our latest blog post about Generalization Through Imitation (GTI) courtesy of
@danfei_xu
and
@AjayMandlekar
:
We have just released our new work on 6D pose estimation from RGB-D data -- real-time inference with end-to-end deep models for real-world robot grasping and manipulation! Paper: Code: w/
@danfei_xu
@drfeifei
@silviocinguetta
Humans heavily rely on visual attention to guide hand movements to perform everyday tasks like reaching and grasping. Can we teach robots similar hand-eye coordination abilities without direct supervision? 👇Our
#IROS2021
work on generalizable imitation via hand-eye coordination.
Can robots learn hand-eye coordination simply from teleoperated human demonstrations? Our new
#IROS2021
paper presents a novel action space to enable this!
Website:
1/9
Can we train visuomotor policies for real-world long-horizon tasks AND generalize across tasks? Join us Wed 8-10am virtually at
#RSS2020
(Paper
#61
) for a live discussion with
@AjayMandlekar
,
@RobobertoMM
, and myself. See 👇 for a short thread about our paper.
Excited to share Generalization Through Imitation (GTI)! GTI learns visuomotor control from human demos and generalizes to new long-horizon tasks by leveraging latent compositional structures.
Joint w/
@AjayMandlekar
@RobobertoMM
@silviocinguetta
@drfeifei
@hausman_k
Agreed that representation is the gap! But I don’t think anyone throught we could solve robotics w/o proper perception, systems modeling, and optimization, and none of which is simply “scripting” (at least not the kind that LLM can solve).
Introducing Menteebot: Groundbreaking Humanoid Robot.
We're proud to unveil Menteebot, the culmination of a two-year journey by our brilliant team!
Menteebot is a groundbreaking humanoid robot designed for versatility.
Visit our website for more demos.