Meet LeRobot, my first library at
@huggingface
robotics ๐ค
The next step of AI development is its application to our physical world. Thus, we are building a community-driven effort around AI for robotics, and it's open to everyone!
Take a look at the code:
Big news for me. I left Tesla!
Iโve been very fortunate to be part of the Autopilot and Optimus teams.
Five years ago, I started as an intern with Andrej, witnessing the impressive achievements of a small team. At this time, I improved the neural backbone, a fundamental module
After 3 years
@tesla
and Optimus, I am thrilled to announce that I joined Hugging Face to start an ambitious open robotics project! (open as in open-source, not as in Open AI) Looking for engineers to build real robots in Paris ๐ซ๐ท ๐ค๐ค
Optimus can now sort objects autonomously ๐ค
Its neural network is trained fully end-to-end: video in, controls out.
Come join to help develop Optimus (& improve its yoga routine ๐ง)
โ
Yeah! Teleoperation on low-cost robot arm from
@taurobots
working natively at 200Hz with and a very simple code. We are working a video tutorial to make it easy for anyone to reproduce our setup (in simulation or with a real robot). Stay tuned ๐
Tesla AI is building next-generation autonomy on a single foundation video network that directly drives the car
Join the team and build state-of-the-art end-to-end models using massive fleet data on one of the world's largest training clusters
We made Reachy2 from
@pollenrobotics
autonomously doing household chores and interacting with us. It can move its full body, including its neck. Even the dog was impressed! ๐ถ You can do the same at home on smaller robots with and
Collecting data at
@pollenrobotics
in Bordeaux ๐ซ๐ท!
Impressed by their low latency teloperation through VR.
You can control the full mobile body and the neck of the robot to see through its eyes ๐
This cool robot is called Reachy2 and you can buy it soon.
Did you miss that
@Stanford
open-sourced a humanoid robot dataset? We just made it available on
@huggingface
hub to allow everyone to train their own humanoid robot with LeRobot:
Here is a cool visualization with
@rerundotio
.
More in this thread ๐งต
Introduce HumanPlus - Shadowing part
Humanoids are born for using human data. We build a real-time shadowing system using a single RGB camera and a whole-body policy for cloning human motion. Examples:
- boxing๐ฅ
- playing the piano๐น/ping pong
- tossing
- typing
Open-sourced!
With just 50 episodes and 30mn of training, our policy works pretty well! It generalizes to new robot positions and light conditions ๐
Fun fact,
@pollenrobotics
is one of the biggest humanoid robot company worldwide with 100 robots sold to customers. They work 4 days a week!
Thanks for trusting me to start this adventure! After 2 months,
@huggingface
is now one of the best places to work on open-source robotics ๐ Who wants to do a summer internship with us?
Excited to announce Tau Robotics (
@taurobots
). We are building a general AI for robots. We start by building millions of robot arms that learn in the real world.
In the video, two robot arms are fully autonomous and controlled by a single neural network conditioned on different
I am mind blown by this new technology!
AI is now embodied.
And we are open-sourcing it all.
Listen to
@HaixuanT
casually discussing with its cute robot at the
@linuxfoundation
:
๐ What's your name?
> I am Reachy, a robot from
@pollenrobotics
, I have two arms.
๐ What do you
I was also impressed by the last video from
@Tesla
Optimus about scaling up data collection to dozen of robots and more tasks in their lab!
I can't imagine what the open-source robotic community will be able to achieve with thousands of affordable robots in various places! ๐๐
Meet LeRobot, my first library at
@huggingface
robotics ๐ค
The next step of AI development is its application to our physical world. Thus, we are building a community-driven effort around AI for robotics, and it's open to everyone!
Take a look at the code:
A Rerun Viewer for the DROID Dataset!
DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset is a robot manipulation dataset by
@SashaKhazatsky
et al. with 76k demonstration trajectories or 350h of interaction data, collected across 564 scenes and 86 tasks.
Tesla AI 10k H100 cluster, go live monday.
Due to real-world video training, we may have the largest training datasets in the world, hot tier cache capacity beyond 200PB - orders of magnitudes more than LLMs.
Join us!
Important paper from Chelsea Finn and Sergey Levine lab:
The authors trained a unique diffusion policy on a real-world dataset of various robots and tasks (manipulation arms, wheeled robots, self-driving cars, robot dogs, drones).
Co-training gets 5%-20%
What strikes me with this video from our lab is how easy we can give a friendly personality to an inanimate object. It's basically possible to make anything cute and lovely. By any chance, do you have good resources for this? (e.g. there is a book on how they made the alien look
@huggingface
's LeRobot team is awesome
Last week in Paris, their teleop was already up to 500Hz thanks to
@HaixuanT
's dora-rs. Feels very natural!
Great work
@RemiCadene
and team!
The year of robotics continues to ๐
Folding clothes with $250 robot arms. I've added another motor to improve mobility and extend the reach. The CAD files and the code are public at:
(video at 2x speed)
Thrilled that researchers around the world are converging for
#ICCV2023
in Paris โ truly a gem among cities! ๐ซ๐ท
First-timer? Cruise along the Seine for iconic views โด๏ธ
Places to hang out
๐ณ Jardin du Luxembourg
๐บ Rue Mouffetard
๐ธ Jardin des Plantes
๐ฒ Bois de Vincennes
๐๏ธ
Super excited about what we can achieve with the power of open source!
@pollenrobotics
built a pretty cool and capable robot. And it's just the beginning! Hope I can try it soon ๐
Image generated by Stable Cascade by
@stabilityai
thanks to with the prompt: "hugging cute little smiling robots with like wall-e with dad high definition scifi 5k holding flowers in paris eiffel tower"
Mobile ALOHA's hardware is very capable. We brought it home yesterday and tried more tasks! It can:
- do laundry๐๐
- self-chargeโก๏ธ
- use a vacuum
- water plants๐ณ
- load and unload a dishwasher
- use a coffee machineโ๏ธ
- obtain drinks from the fridge and open a beer๐บ
- open
Additionally, the same models can be trained on real-world datasets. Here is a cool data visualization with
@rerundotio
which is fully integrated with our video format optimized for training. The data originally comes from the Aloha project:
Do you have recommendation on papers for robot navigation in home? Is end-to-end navigation a thing? Is it possible to avoid SLAM or use it only as conditioning/input?
Really interesting work!
A big issue with current AI for robotics approaches is that you dont really know when to early stop the training. It's not like with image classification where you just take the best checkpoint on the validation set. When doing imitation learning on real
Donโt have a real robot/setup but want to evaluate policies trained on real world datasets? Check out SIMPLER, fast, safe, and reliable evaluation of real robot policies in sim via ManiSkill 2. The ManiSkill 3 beta will port SIMPLER over soon so stay tuned!
LeRobot is to robotics what the Transformers library is to NLP. It offers clean implementations of advanced AI models with pre-trained checkpoints. We also reimplemented 31 datasets from academia, and some simulation environments, allowing to get started without a physical robot.
Another visualization with LeRobot, this time on Mobile Aloha data, to learn navigation and manipulation totally end-to-end.
Both datasets have been collected on
@trossenrobotics
robot arms.
Teleoperation though a VR headset will certainly be the cheapest and most efficient approach to teach a robot to achieve complex tasks. I recently came across from
@soumithchintala
robotic group at NYU that does just that! Thx for open sourcing the code!
LeRobot also features the Diffusion Policy, a powerful imitation learning algorithm, and TDMPC, a reinforcement learning method that includes a world model, continuously learning from its interactions with the environment.
Unitreeย H1 Breaking humanoid robot speed world record [full-size humanoid] ย Evolution V3.0 ๐ฅฐ
The humanoid robot driven by the robot AI world model unlocks many new skills!
Strong power is waiting for you to develop!
#Unitree
#AI
#subject3
#BlackTech
Introducing ๐๐จ๐๐ข๐ฅ๐ ๐๐๐๐๐๐ -- Hardware!
A low-cost, open-source, mobile manipulator.
One of the most high-effort projects in my past 5yrs! Not possible without co-lead
@zipengfu
and
@chelseabfinn
.
At the end, what's better than cooking yourself a meal with the ๐ค๐งโ๐ณ
Can we collect robot data without any robots?
Introducing Universal Manipulation Interface (UMI)
An open-source $400 system from
@Stanford
designed to democratize robot data collection
0 teleop -> autonomously wash dishes (precise), toss (dynamic), and fold clothes (bimanual)
Really great contribution from
@kenny__shaw
,
@pathak2206
and colleagues! We need more affordable robotic platforms to teach and study the fundamentals ๐คฉ
Robotic hands are daunting -- costly yet super fragile.
After yrs of frustration, we decided to make one that is better, stronger & anyone can build!
Open sourcing LEAP Hand ๐๐ค
- low cost ($2K)
- 3D printed. Easy to assemble (3hr)
- sim2real code etc.
LeRobot codebase has been validated by replicating state-of-the-art results in simulations. For example, here is the famous ACT policy which has been retrained and made available as a pretrained checkpoint:
It was my pleasure to discuss with
@parmy
about Hugging Face and the importance of open source AI and robotics in our society. You should also check out other insightful articles from Parmy Olson on
@Bloomberg
๐
โ[Open-source AI] is really fundamental because it allows everyone to seize the technology, to diminish the fear of limited understanding or of not being qualified to use AIโ -
@RemiCadene
๐ฅ๐ฅ
Great piece in
@opinion
by
@parmy
on the open-source landscape
Time for the open-source AI robots revolution ๐
Weโve been playing with a low-cost DJI robot controlled by 3 local open-source AI models (Whisper, Idefics2, Parler-TTS - all Apache2) & orchestrated by Dora-cs
In comments a 250 lines code gist to build on top of it => enjoy!!
We are honoured to host
@chichengcc
, author of Diffusion Policy and UMI (Universal Manipulation Interface), for our next LIVE paper discussion. To access:
1. Join our discord:
2. Click on google meet link of this post:
For our next
#LeRobot
Paper Discussion,
@chichengcc
will present his first authored Diffusion Policy!
The event will start at 6 PM GMT on 12 June and will officially run for 1 hr.
Join our Discord and check out the events channel for full details. ๐ค
OpenAI + humanoid robots โ weโre collaborating with
@Figure_robot
to expand our multimodal models to robotic perception, reasoning, and interaction.
@pollenrobotics
Again, we only trained on 50 episodes for 2 hours. We tested all checkpoints and saw our network learning new concepts! At 20k steps, it learned to put the cup on the dish rack reliably. At 40k steps, it learned to rotate its base to give back the apple! At 60k steps, it learned
Really cool :) Maybe some possible future collaborations with Hugging Face for open source robotics: crow-sourced datasets, foundational models, affordable hardware?
University of Bonn is part of the new BMBF-funded Robotics Institute Germany (RIG).
Five groups from Bonn contribute:
* Sven Behnke
* Cyrill Stachniss
* Maren Bennewitz
* Chris McCool
* Heiner Kuhlmann.
Press release:
The ML ecosystem in France is on fire๐ฅ It has amazing talent and resources. Here are 10 facts you might not know:
1. There are great research labs - from
@MistralAI
and
@kyutai_labs
to large ones from
@AIatMeta
and
@GoogleDeepMind
. The Llama 2 and CodeLlama authors are based in
Congrats to torchrl developers and community!
Torchrl is on its way to become THE library for reinforcement learning and robotics in pytorch.
You should check it out ;)
๐ Exciting news! TorchRL paper has been accepted as a spotlight talk at
@iclr_conf
in Vienna this May! Huge thanks to the amazing OSS community, there would be no torchrl if it wasn't for our contributors. See you there!
#ICLR2024
#Vienna
just finished printing first Universal Manipulation Interface (UMI) - now assembling time with
@RemiCadene
and on to labeling in the wild
very excited about what's coming for open-source $500 robots
h/t
@chichengcc
& co for pushing more robot data collection democratization
Come join our Discord channel. We are building a diverse community from various backgrounds, software and hardware, to develop the next generation of smart robots in the real-world!
@pollenrobotics
We also made their pretrained models available (which is useful for the ~100 customers of Pollen-Robotics):
Of course, we plan to do the same for affordable robot arms that everyone can own! Stay tuned ๐
New open Vision Language Model by
@Google
: PaliGemma ๐๐ค
๐ Comes in 3B, pretrained, mix and fine-tuned models in 224, 448 and 896 resolution
๐งฉ Combination of Gemma 2B LLM and SigLIP image encoder
๐ค Supported in
@huggingface
transformers
Model capabilities are below โฌ๏ธ
After spending just 20 minutes with the
@MistralAI
model, I am shocked by how unsafe it is. It is very rare these days to see a new model so readily reply to even the most malicious instructions. I am super excited about open-source LLMs, but this can't be it!
Examples below ๐งต
Our primary goal: get more robotics and AI researchers to go beyond constrained lab environments and start getting into homes!
So, we open-sourced our entire system.
Dataset, code, robot software & hardware:
Github repo: (8/n)
I used
#lerobot
's diffusion policy implementation to figure out if spatial softmax learns semantically sensible keypoints. Short answer: no. Here's my short writeup with animated visualizations:
After 3 years
@tesla
and Optimus, I am thrilled to announce that I joined Hugging Face to start an ambitious open robotics project! (open as in open-source, not as in Open AI) Looking for engineers to build real robots in Paris ๐ซ๐ท ๐ค๐ค
We have just released ๐ท FineWeb: 15 trillion tokens of high quality web data.
We filtered and deduplicated all CommonCrawl between 2013 and 2024.
Models trained on FineWeb outperform RefinedWeb, C4, DolmaV1.6, The Pile and SlimPajama!
Looking forward to at Station F in Paris ๐ซ๐ท
Eric Schmidt (Google founder), Jensen Huang (Nvidia founder), Xavier Niel (Free founder), among other speakers.
Big AI industry event at Station F in Paris, Friday Nov 17.
Keynote Speakers: Eric Schmidt (Schmidt Future), Jensen Huang (CEO Nvidia), Xavier Niel (Iliad/Scaleway).
Speakers include: Thomas Scialom (Meta, of Llama fame), Arthur Mensch (CEO Mistral, or Mistral-7B fame), Jason
How to do good research? Like making good music. Everyone have their style. You eventually find your own. Trying to copy styles that work best can be useful at first.
careful about overfitting to lists like this. there are many ways to do good research -- my fav papers were born out of getting "stuck in rabbit holes" that no-one else went down...
Impressive work! Especially, finetuning with online RL with reward provided by a CLIP model + automated reset :)
Any plan to open-source the dataset like Mobile ALOHA?
@HaoyuXiong1
@mendonca_rl
@pathak2206
Adaptive Mobile Manipulation for Articulated Objects In the Open World
paper page:
Deploying robots in open-ended unstructured environments such as homes has been a long-standing research problem. However, robots are often studied only in closed-off lab
In 2024, explaining what a model learned or why it generated a certain output is still challenging!
LENS is tackleing this by providing all concepts used by a pretrained model for all 1000 classes of ImageNet.
Check-out by yourself, it's super interesting ๐ค
๐๐จโ๐ณ๐ต After a year of cooking up a secret project, I'm thrilled to officially reveal:
The ๐๐๐๐ ๐๐ซ๐จ๐ฃ๐๐๐ญ.
By combining modern tools of Explainable AI, how much can we explain a ResNet50?
๐งถ