We dramatically sped up Diffusion policies through consistency distillation.
With the resulting single step policy, we can run fast inference on laptop GPUs and robot on-board compute.
👇
Diffusion Policies are powerful and widely used. We made them much faster.
Consistency Policy bridges consistency distillation techniques to the robotics domain and enables 10-100x faster policy inference with comparable performance.
Accepted at
#RSS2024
Back in Summer 2019, a group of Roboticists and Machine Learning Researchers met at the Robot Learning Summit in Montreal.
We had lively discussions on the Challenges and Opportunities for Embodied Intelligence.
Here are our thoughts:
Recognizing symbols like "dish in dishwasher" or "cup on table" enables 🤖 task planning
But how do we get data to train models for recognizing symbols?
Introducing "Grounding Predicates through Actions" to automatically label human video datasets 🧵
New
@ICEgov
policy regarding F-1 visa international students is horrible & will hurt the US, students, and universities. Pushes universities to offer in-person classes even if unsafe or no pedagogical benefit, or students to leave US amidst pandemic and risk inability to return.
We made the first place at the NuScenes Tracking Challenge (AI Driving Olympics
@NeurIPSConf
)
Spoiler alert: It's a Kalman filter!
We beat the AB3DMOT baseline by a large margin.
arXiv
Don't we all want our robots to do many different tasks?
We learned a single policy for 78 manipulation tasks.
We don't use goal images. We use natural language instructions to index the task.
Chat with us tomorrow 8am PDT
#RSS2020
1/4
Excited to share our new work on vision-only navigation using NeRFs!
We show how to use a pre-trained NeRF to both represent collision geometry and localize a robot with an onboard camera.
Project page:
Paper:
We propose a method for in-hand manipulation that combines low-level manipulation primitives with a learned, mid-level policy that orchestrates these primitives.
Our method keep the object firmly grasped while transporting it to distant goal poses.
Seeing TidyBot come together in my lab has been fun! And yes, our lab has been exceptionally clean 🧹
Many of you commented on the mobile platform which I'm so pleased about seeing it move!
Let me tell you the story behind these platforms!
🧵
When organizing a home, everyone has unique preferences for where things go. How can household robots learn your preferences from just a few examples?
Introducing 𝗧𝗶𝗱𝘆𝗕𝗼𝘁: Personalized Robot Assistance with Large Language Models
Project page:
People often use customized tools for a variety of manipulation tasks: 🥄🍴🥢🪛🪝🪓🛠️✂️🧵📎
We look at the problem of automatically designing customized tools for robots and leverage differentiable simulation and continual learning.
Video & Paper:
🧵
How do you sequence learned skills for a manipulation task?
Use STAP to plan with learned skills and maximize the expected success of each skill in the plan where success is encoded in Q-functions.
🐙 Code for STAP is now on github
I am very grateful to the Sloan Foundation to recognize and support the research done at my lab
@StanfordIPRL
!
This wouldn't have been possible without the hard work of my students, postdocs and the support by my mentors!
We are delighted to announce the winners of this year’s Sloan Research Fellowship! These outstanding researchers are shining examples of innovation and impact—and we are thrilled to support them. Meet the winners here: 🎉
#SloanFellow
#STEM
#ScienceTwitter
How do you perform a manipulation task with a novel object that is heavily occluded?
We propose a method for task-driven in-hand manipulation of unknown objects with tactile sensing.
🧵 1/n
3D multi-object tracking is a challenging problem, because it requires effective data association, track lifecycle management, false-positive elimination, and false-negative propagation.
To address all 4 of these problems, we propose ShaSTA!
🧵 1/5
Thank you so much for all the congratulations to the 2020 RSS Early Career Award!
I'm honored to receive it and astounded and happy to hear that I serve as an inspiration.
I want to mention who inspires me
Our multi-robot task allocation algorithm is out!
It addresses the key computational challenges of sequential decision-making under uncertainty and multi-agent coordination.
RSS'20 Paper, Code, Reviews:
Hanging objects is a common daily task.
We propose OmniHang to learn to hang arbitrary objects onto a diverse set of supporting items using contact point correspondences and neural collision estimation.
Project webpage:
Paper:
Target networks in deep RL can slow down the learning process. We propose a self-regularized and self-guided actor-critic (GRAC) method and achieve state of the art in MuJoCo continuous control tasks.
No target network needed.
Project Page:
How do you generate stable grasps for n-fingered hands?
Check out our new work UniGrasp! One model that generates n contact points for novel objects and novel hands.
Project Page
arXiv
Manipulation of deformable objects is hard.
Check out our approach!
We propose a self-supervising training objective on real images to robustly perceive a high-dimensional object state w/ occlusions, some clutter, appearance variation.
Project + arXiv
Tracking by detection is a powerful paradigm.
Multi-object tracking performance is largely driven by detection accuracy: the better the detector, the better tracking.
So just work on better object detectors?
Not so fast .... 🧵
Details:
At
#CVPR2023
, we present CARTO - a model that reconstructs articulated objects from a single stereo image.
This includes the object's 3D shape, 6D pose, size, joint type, and joint state. All this in a category agnostic fashion.
I was invited to the
#IROS2020
Tutorial on Deep Representation and Estimation of State for Robotics by
@deanh_tw
@danfei_xu
@KuanFang
:
I created a mini-course on Representations and their interplay with decision-making & control:
So many ways to generate a sequence of robot skills to solve a task that requires long-horizon reasoning: LLMs, task planners, high-level policies, …
But how do you ensure that each skill is executed such that the next skill can be successful at all?
So happy to see
@Stanford
joining our peer institutions in an amicus brief to support Harvard and MIT in a lawsuit against the
#StudentBan
The letter of MTL to the acting secretary of the Department of Homeland Security
BREAKING: MIT and Harvard just filed suit against ICE and the DHS seeking to reverse their ban on international students being able to stay in the US while taking classes remotely.
Full complaint:
Are there more Roboticists out there right now, on a deadline grind, wondering: How is this even possible?
Such an incredible achievement. So many possible points of failure. So many years of work.
JUST, WOW.😍
Grab the popcorn because the
@NASAPersevere
rover has sent us a one-of-kind video of her Mars landing. For the first time in history, we can see multiple angles of what it looks like to touch down on the Red Planet.
#CountdownToMars
Learning contact-rich manipulation skills is hard for robots!
But what if the robot could modify its environment to help skill learning?
We let a robot discover how to place fixtures that funnel uncertainty and dramatically speed up skill learning.
We perform precision grasping with multi-fingered hands on objects seen from only a single RGB-D camera.
What is new and different about this work is that we model grasping as a dynamic process where the object is allowed to move while the hand establishes a stable grasp.
New Grasping Data Set Alert 🚨
We trained a model called UniGrasp to generate stable grasps for multi-fingered hands.
Code + Data now available including not only a variety of objects but also hands:
How do you generate stable grasps for n-fingered hands?
Check out our new work UniGrasp! One model that generates n contact points for novel objects and novel hands.
Project Page
arXiv
Understanding dynamic 3D environment is crucial for robotic agents.
We propose MeteorNet for learning representations of dynamic 3D point cloud sequences (Oral
@ICCV19
).
Project Page:
Arxiv:
Learning of manipulation skills often considers point-to-point motion only. Many manipulation tasks are however periodic and repeated indefinitely:
🧵🥣🛠️🪚🧼🪠🧹🪡✂️ 🪛🪥
We learn and represent periodic policies with complex deformable objects or granular material from vision
How can robots learn to perform various periodic tasks in our everyday life, like wiping tables and stirring food? Check out our approach ViPTL that can learn to execute a periodic task from a single human demo!
Paper:
Website:
👇
Ever worked on a real robot in a not so controlled environment?
Then you may have experienced changing lighting conditions, sensor drift or just a broken sensor.
We propose a way to deal with these unexpected measurements through crossmodal compensation
🧵
What happens if your camera gets occluded? Use remaining sensors to compensate for the corrupted image!
Introducing Crossmodal Compensation Model, a representation model that compensates for corrupted sensors.
Paper:
Project:
We want our robots to extrapolate from a few examples of a manipulation task to many variations.
Embedding equivariance in both, our object representation and policy architecture allows our 🤖 to do just that.
From a few examples of solving a task, humans can:
🚀 easily generalize to unseen appearance, scale, pose
🎈 handle rigid, articulated, soft objects
0️⃣ all that with zero-shot transfer.
Introducing EquivAct to help robots gain these capabilities.
🔗
🧵↓
Phase 2 of the challenge starts this Monday, October 12! Amazing effort by all the teams in Phase 1, thank you for participating! 🙂
Congratulations to the selected teams!! Come Monday, training on the real TriFinger platforms will be as simple as running jobs on a cluster 😉
Mark your calendars for this new public Robotics Seminar Series starting Friday May 15th 1PM EST with
@AjdDavison
!
Upcoming amazing speakers:
@AjdDavison
, Leslie Kaelbling, Allison Okamura,
@ancadianadragan
Hope to see you there and participate in the Q&As!
We are launching a new
#virtual
#robotics
seminar “Robotics Today--A Series of Technical Talks”. Andrew Davison
@AjdDavison
will give the first seminar “From SLAM to Spatial AI” on Friday May 15th 1PM EST.
Watch and learn more about the seminars here:
We embedded equivariance in Diffusion policies to let them generalize to scenarios that would otherwise be out of distribution.
👇 Check Jingyun's thread for all the details!
Want a robot that learns household tasks by watching you?
EquiBot is a ✨ generalizable and 🚰 data-efficient method for visuomotor policy learning, robust to changes in object shapes, lighting, and scene makeup, even from just 5 mins of human videos.
🧵↓
Thank you Scott Kuindersma
@BostonDynamics
, for this very exciting and entertaining talk! On top of all the controller details, I loved hearing about how you integrate perception.
Re-watch talk and Q&A here!
1/3 There are all these rich human activity datasets out there! Wouldn’t it be great if a robot could learn task-oriented grasping from them?
Unfortunately, these datasets often lack 6D hand-object pose annotations.
Here are some ideas on how to automatically annotate them.
Figure 01 has learned to make coffee ☕️
End-to-end AI system, trained in 10 hours, just by watching humans make coffee
Our neural networks are taking video in, trajectories out
Join us to ship a fleet of AI robots:
Working with robots? Sometimes having trouble with coordinate transforms and rotations? Here are STL files for printing these extremely useful 3D coordinate systems with magnets, so you can attach them to your whiteboard. Kudos to Felix from
@MPI_IS
Looking forward to present our line of work on enabling robots to do long-horizon task planning by leveraging large-scale human activity datasets!
#CVPR2021
Thanks for the invite!
Sun 20/6 10:00 EDT (full day)
EPIC Workshop with 5 challenges announcement, and live talks with Q&A by winners. Keynotes by L Torresani, D Krandall
@leto__jean
@davsca1
and K Grauman. Program:
2/6
We have previously shown how multimodal representation learning enables generalization in contact-rich manipulation tasks.
I'm excited about this new
#CVPR2022
work that is object-centric and includes vision, touch and audio.
Excited to share our
#CVPR2022
paper ObjectFolder 2.0, a multisensory object dataset with visual, acoustic, and tactile data for Sim2Real transfer.
Paper:
Project page:
Dataset/Code:
Details with narration👇
Solving long-horizon tasks requires task and motion planning!
Our robot optimizes hand poses relative to target objects. The plan remains valid even with moving objects.
Check out for the Tower of Hanoi and robot demos!
URL:
To do daily chores, robots need to understand articulated objects. Sometimes a single picture of an object is deceiving. We propose a novel method that leverages temporal data to estimate the object articulation mechanism.
🧵 1/9
Fusing multiple modalities for state estimation requires dynamics and measurement models.
Differentiable filters enable learning these models end-to-end while retaining the algorithmic structure of recursive filters.
Project Page:
#IROS2020
Happy to share our new paper on Making Sense of Vision and Touch:
Self-Supervised Learning of Multimodal Representations for Contact-Rich Tasks. Featuring RL on a real robot.
How do you autonomously learn a library of composable, visuomotor skills?
At
#IROS2023
we present an approach that can create novel yet feasible tasks to gradually train the skill policies on harder and harder tasks.
Monday morning Poster Session: MoAIP-19.9
Active Task Randomization (ATR) learns to create novel and feasible tasks for acquiring generalizable visuomotor skills. The learned skills can be composed to solve unseen sequential manipulation tasks in the real world.
Interested in creating a truly immersive multimodal virtual reality experience?
We propose a learned model to generate realistic haptic sensory data for a multitude of virtual materials.
Project:
arXiv:
I'll be presenting EquiBot today (Friday) at
#RSS2024
in the Workshop on Geometric and Algebraic Structure in Robot learning.
Room: Room: ME B - Newton
Poster Session: 3-4pm
Want a robot that learns household tasks by watching you?
EquiBot is a ✨ generalizable and 🚰 data-efficient method for visuomotor policy learning, robust to changes in object shapes, lighting, and scene makeup, even from just 5 mins of human videos.
🧵↓
How do you sequence learned skills for a manipulation task? Use STAP to plan with learned skills and maximize the expected success of each skill in the plan where success is encoded in Q-functions.
Today 3-4pm
#ICRA2023
at Pod 42 WePO2S-21.1
@agiachris
Maybe unpopular opinion:
Unless you control the environment, manipulation robots will always make mistakes ... just like people.
We looked at how a robot can online learn from failure and overcome it:
No, this is not about RL!
This afternoon
#ICRA2024
, we are presenting RoboFuME - a method that implements the pre-train + fine-tuning paradigm for multi-task robot policies.
4:30-6pm at "Big Data in Robotics and Automation"
Room: CC-313
Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that
1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data;
2. runs RL fine-tuning online with the VLM reward model.
🔗
🧵↓
It is this time of the year: We are teaching rotations in CS336.
Check out these awesome 3D printed coordinate frames for visualizing rotations and making your life easier.
Working with robots? Sometimes having trouble with coordinate transforms and rotations? Here are STL files for printing these extremely useful 3D coordinate systems with magnets, so you can attach them to your whiteboard. Kudos to Felix from
@MPI_IS
Really cool, out-of-the-box hand-design allowing for new ways to think about in-hand manipulation!
My student
@linshaonju
also contributed with his wisdom on policy learning - in this case through imitation
Manipulating objects using a robot grasper with steerable rolling fingertips. Our Roller Grasper V2 project is being presented at
#IROS2020
Details at:
Large Language Models promise to replace task planners in robotics.
But how do we verify that these plans are correct - especially for tasks that require long-horizon reasoning?
👇Check out Kevin's 🧵 on text2motion!
Large language models (LLMs) can readily convert language instructions into high-level plans.
However, should we trust robots to execute these plans without verifying that they actually satisfy the instructions and are feasible in the real world?
Differentiable recursive filters allow end-to-end learning of process and observation models + noise profiles.
This is incredibly useful for state estimation of systems that are hard to model.
But HOW exactly do you train your filter?
The nitty gritty:
Check out our paper “Vision-Only Robot Navigation in a Neural Radiance World” at
#ICRA2022
! Come learn how to use NeRFs for trajectory optimization, state estimations, and MPC.
We’ll be presenting at 4:15pm on Thursday in Room 124!
More info:
A lot of us
@StanfordEng
spend a lot of time at Bytes and Coupa. The employees there are hourly, and with Stanford’s effective shutdown, their hours are being cut or people are being let go.
This fundraiser is to help them with financial assistance:
Thanks so much
@siddhss5
for this wonderful talk on motion planning! You can re-watch it below.
Some of the audience questions were on "What should I work on?"
Sidd's advise: Be fearless!
To unwrap this, check out his talk on this topic here:
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.
How do you autonomously learn a library of composable, visuomotor skills?
We propose an approach that can create novel yet feasible tasks to gradually train the skill policies on harder and harder tasks.
Active Task Randomization (ATR) learns to create novel and feasible tasks for acquiring generalizable visuomotor skills. The learned skills can be composed to solve unseen sequential manipulation tasks in the real world.
Tying a knot is a topological task that has infinitely many geometric solutions. How can a robot translate a topological plan into a motion plan?
We learn a library of topological motion primitives for creating various knots.
#iros2020
Project:
Thanks
@StanfordEng
for this profile!
Also thanks to
@ai4allorg
for everything you do and specifically for organizing the fireside chat described in this story!
#IAmAnEngineer
: I'm a first-generation university graduate who grew up in communist East Germany. I'm passionate about mentoring young women in science and engineering. -Jeannette Bohg, Assistant Professor of Computer Science
We want robots that can manipulate any articulated object 🚪🪟📪🗃️🗄️📒📦
Turns out: this is still harder than you think!
AO-Grasp proposes robot grasps on articulated objects and generalizes 0⃣ shot from sim2real.
Talk to us at
#BARS2023
Finding grasps that enable robots to interact with articulated objects is challenging because:
🗄️articulated objects exist in infinite joint configurations
🤖grasps need to be both stable and actionable
We tackle this in our work AO-Grasp
🔗
[1/7] 🧵
Can we grasp everyday objects with a dexterous hand from only a single RGB-D image?
We propose SpringGrasp, an optimization-based grasp planner that can synthesize compliant dexterous grasp for shape-uncertain objects.
Using contact feedback leads to a more robust grasp acquisition policy - especially for objects with a complex shape and when given noisy pose estimates. Kudos to Hamza, Miroslav, Daniel and Ludo!
We're excited to announce our
@ieee_ras_icra
workshop "Motion Planning with Implicit Neural Representations of Geometry!"
We'll discuss the future of INRs -- like DeepSDFs, NeRFs, and more -- in robotics.
Submissions due April 15.
We are happy to announce the organized by
@MPI_IS
,
@MILAMontreal
and
@NYU
. Each team passing the simulation stage will have more than 100 real-robot-hours. Coding in Python, simulator in PyBullet and tutorials are provided. Start 3rd Aug.
Dedicated to giving girls the opportunity to explore robotics. Age doesn't matter 🤖 Featuring Naos, Athena & Apollo
@MPI_IS
,
@FRANKAEMIKA
& Ocean1
@Stanford
So excited to see these news! Matt is the best. Congratulations to CMU and to Matt.
He was a PostDoc at KTH in Stockholm when I was doing my PhD. Working with Matt gave me such a boost of energy! Truly enjoyed it!
Can't wait to see him leading RI
Need a method for your robot to visually track objects and its own arm during manipulation? Look no further: Bayesian Object tracking now also runs on Ubuntu 16.04 LTS
Super excited about co-organizing this virtual Robotics Retrospectives workshop
@rssconf
!
The good thing about making it virtual? We don't have to worry about visas! 😊
Check out the webpage for submission info and our virtual formats
Very excited about the program of our virtual Robotics Retrospectives workshop ! Want to participate? Submit your reflections on your own past work or a subfield of robotics.
with
@leto__jean
, Arunkumar Byravan and Akshara Rai.
We want our robots to extrapolate from a few examples of a manipulation task to many variations.
Embedding equivariance in both, our object representation and policy architecture allows our 🤖 to do just that.
Today Wed
#ICRA2024
Room: CC-418
Time: 1:30-3pm
Poster: 4:30-6pm
From a few examples of solving a task, humans can:
🚀 easily generalize to unseen appearance, scale, pose
🎈 handle rigid, articulated, soft objects
0️⃣ all that with zero-shot transfer.
Introducing EquivAct to help robots gain these capabilities.
🔗
🧵↓
My first time at
#CVPR2019
. I'll be giving a talk at the Vision meets Cognition ws in 103c along with other great speakers! Come by if you like to hear about the challenges and opportunities of bringing together vision and robotics
Imagine you’re in a dark room searching for something when an object drops to the floor. The sound will tell you a lot about the object's material.
Our model enables robots to do the same: infer object properties from their sound (in the future w/o clumsily dropping objects).
Watch live: 1 PM Friday, June 12:
@Berkeley_EECS
’s Anca Dragan
@ancadianadragan
#humanrobot
interaction "Optimizing Intended Reward Functions: Extracting all the right information from all the right places"
@black_in_ai
launches the 2020/21 Graduate Application Program for those of you who self-identify as Black and/or African and intend to apply for grad school this cycle.
More info here:
Paging
@BlackInRobotics
for signal boosting 🤖
Proud that our lab has contributed to this long-term data collection effort called Droid - an in-the-wild robot manipulation dataset of unprecedented diversity.
Amazing leadership by
@SashaKhazatsky
,
@KarlPertsch
and
@chelseabfinn
🧵👇
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
Love this!
A lecture that gives insights into how research directions come about, how much effort and time they take to follow through and about the context of the research community in which they blossom.
Last semester, a guest lecture fell through last minute. So improvised a lecture on the backstories behind some research papers -- i.e. not about the research results, but about how we ended up doing research in that direction.
#metaresearch
Online now:
Love the idea of this retrospective NeurIPS WS!
The overarching goal of retrospectives is to do better science, increase the openness and accessibility of the machine learning field, and to show that it’s okay to make mistakes.
One of those at
#ICRA2020
?
Hi
#ICRA2023
folks, we are presenting text2motion that can verify LLM plans in the 'Pretraining for Robotics' Workshop this morning.
Location: ICC Capital Suite 7
Spotlights 10am, Poster session 10:20am. CU there with
@agiachris
Large language models (LLMs) can readily convert language instructions into high-level plans.
However, should we trust robots to execute these plans without verifying that they actually satisfy the instructions and are feasible in the real world?
How can we adopt the successful pre-training + finetuning paradigm in Robotics?
Presenting RoboFuME 🤖that can learn new manipulation tasks by finetuning a multi-task policy.
And all of that with minimal human supervision!
👇
Announcing RoboFuME🤖💨, a system for autonomous & efficient real-world robot learning that
1. pre-trains a VLM reward model and a multi-task RL policy from diverse off-the-shelf demo data;
2. runs RL fine-tuning online with the VLM reward model.
🔗
🧵↓
Key Insight of TidyBot:
Summarization with LLMs is an effective way to achieve generalization in robotics from just a few example preferences.
Jimmy presents TidyBot on Monday afternoon at
#IROS2023
: MoBIP-16.5
Come by to chat!
When organizing a home, everyone has unique preferences for where things go. How can household robots learn your preferences from just a few examples?
Introducing 𝗧𝗶𝗱𝘆𝗕𝗼𝘁: Personalized Robot Assistance with Large Language Models
Project page:
What is happening in Iran is heartbreaking and difficult to watch.
With fellow roboticists I condemn any form of violence against innocent children, women and the people in Iran.
@linshaonju
@StanfordIPRL
@krishpopdesu
Thanks for all the congrats 😊
Maybe some perspective on the 100% success rate:
4 of those papers were rejected before. Although it can be frustrating to be rejected, we worked hard on improving them based on the reviews.
So I am extra happy for those 4! And never give up 🙌
While humans can seamlessly combine our sensory inputs, can we teach robots to do the same?
@michellearning
writes about how we can use self-supervision to learn a representation that combines vision and touch.
We worked on a new, more convenient way to goal-condition robot policies: user sketches! 👨🎨
Sketches are less ambiguous than language and not too specific (like goal images).
We can tell our robots what we want them to do, but language can be underspecified. Goal images are worth 1,000 words, but can be overspecified.
Hand-drawn sketches are a happy medium for communicating goals to robots!
🤖✏️Introducing RT-Sketch:
🧵1/11