@CVL_ETH
A bit off the topic:
never devalue yourself, no matter what others say or do on you, even when it is your advisor, the one you are emotionally dependent on, and the one you have intimate relationship with.
finally made it public!
We trained end2end model-free policies for risky terrains that were reserved for model-based methods.
Super agile, no motion references. Paves our way to the ultimate solution of quadruped locomotion.
For anyone interested
We have significantly updated the ABS manuscript (now it's 18 pages and almost everything is explained):
Also, code here
Also, it's now an RSS accepted paper with full score reviews
Agile But Safe
Learning Collision-Free High-Speed Legged Locomotion
Legged robots navigating cluttered environments must be jointly agile for efficient task execution and safe to avoid collisions with obstacles or humans. Existing studies either
It's also inspired by
@BostonDynamics
's awesome works.
The key question is -- to which extent we need model-based methods for motion synthesis, when
#RL
may support robust and versatile control w/ efficient computation?
Another work for this Question:
๐จ Without Any Motion Priors, how to make humanoids do versatile parkour jumping๐ฆ, clapping dance๐คธ, cliff traversal๐ง, and box pick-and-move๐ฆ with a unified RL framework?
Introduce WoCoCo:
๐ง Whole-body humanoid Control with sequential Contacts
๐ฏUnified designs for minimal
Things are getting hotter and I feel hard to catch every paper..
# of papers after 2023 is almost equal to the number before 2023, and many are on humanoids!
More and more challenging problems conquered, IMO, starting by
@zhaomingxie
's early work in 2018
#RL
for
#humanoid
control is good -- but HOW-TO enable humanoid to fulfill contact sequences like PARKOUR jumping by END-TO-END RL, w/o any motion prior, and apply adaptive contact forces?
See our work WoCoCo!
I also made a brief explanation video here:
๐จ Without Any Motion Priors, how to make humanoids do versatile parkour jumping๐ฆ, clapping dance๐คธ, cliff traversal๐ง, and box pick-and-move๐ฆ with a unified RL framework?
Introduce WoCoCo:
๐ง Whole-body humanoid Control with sequential Contacts
๐ฏUnified designs for minimal
Some lessons on Unitree H1 (and perhaps G1๐ฅ)
Good:
1 cheap
2 100% rigid body, com centered, friendly for RL walking
3 ~direct motor, accurate torque when small
4 large battery
5 large torque and velo limit
--
Bad:
1. Huge impact - not very durable, imu-unfriendly
(Cont'd)
Make Pixels Dance: High-Dynamic Video Generation
paper page:
Creating high-dynamic videos such as motion-rich actions and sophisticated visual effects poses a significant challenge in the field of artificial intelligence. Unfortunately, current
@zdeborova
If you can't find experts for each paper, students are quite good choices.
1. Many ug (mostly 4y) have top conf pub.
2. As students, we are serious with reviewing. We know we are 'noobs', we respect papers.
3. We also do reviewing to build our taste in our first academic stages.
People should try to close the conceptual gap between frontier researchers and general audience. This was something published 2y ago with techniques (mostly optimization without AI) matured at least 3y ago put together as a journal paper. But today's audience still find it new.
This swarm of 10 bright blue drones lifts off in a bamboo forest in China, then swerves its way between cluttered branches, bushes and over uneven ground as it autonomously navigates the best flight path through the woods.
[๐น Zhejiang University]
Trivial today.
Triple backflip:
Learning from reference (more likely) or have phase observation and crafted stage-wise rewards (I once did w/ SAC)
Others:
Heavy orientation rewards together with velocity tracking. Sufficient randomization. Multi-expert likely but not necessary.
Daily Training of Robots Driven by RL
Segments of daily training for robots driven by reinforcement learning.
Multiple tests done in advance for friendly service humans.๐
The training includes some extreme tests, please do not imitate.
#AI
#Unitree
#AGI
#EmbodiedIntelligence
Seems that there are no commercial legged robots with both powerful and agile upper body?
Talos seems quite decent with 6kg payload at ~1m arm spread, and has torque transparency. But the motions are very slow.
@Toukairinn_FUZZ
Hi, For your interest, we report the peak torques and joint velocities in the paper (and power = torque * joint_vel).
Compared to the commonly used trotting gait (the sota is named "rapid" here), ours have higher velocity but much lower energy consumption!
Weird fact: I've somehow become the only person who helped all of existing projects of
[human->humanoid]+[whole-body control]+[teleoperation->autonomy], though none of these are led by me:
1. H2O, as co-author
2. Omni H2O, as co-author
3. HumanPlus, minor help on sim2real
#RSS2024
Denoising World Model Learning (DWL) is accepted by RSS 2024 with all 4.0 scores!
DWL is my first RSS paper ever, thanks to
@JianyuChen_THU
, Xinyang Gu from
@roboterax
, and Xiang Zhu, Chengming Shi,
@GYanjiang
, Yichen Liu from
@Tsinghua_IIIS
.
We will publish this paper
I've always thought this too. Although I'm mostly still an RL skeptic, I think the place this CLEARLY isn't true is walking robots; even Boston Dynamics has been getting in on the RL action these days
solid papers from academia and open source RL infra put humanoids around the world on the same level. Some hyped companies seem to fall behind tho.
But the fast progress will stop here - winner of next stage must have good state estimator and perception solution.
The Beijing Humanoid Robot Innovation Center has unveiled "Tiangong," a full-size humanoid robot capable of running on electric power.
Developed as a general-purpose humanoid robot platform, Tiangong stands 1.63 meters tall, weighs 43 kilograms, and can maintain a steady
Welcome to our
#ICRA2024
presentation at 10.30 am on 14 May. Room 301, award session for best paper on cognitive robotics (or robot+learning).
#robotics
#robotlearning
My coauthor Jin will present our work.
Project page:
Got 2/2 co-first-authored papers accepted by
#RSS2024
.
Thanks to all reviewers and ACs -- rly constructive and insightful comments.
The rebuttal results are confusing tho. Seems that AC manually updated the reviews based on the reviewer feedback. ๐
For
@CMU_Robotics
people:
if you attend
@corl_conf
in Germany this year, you can:
1. travel in Schengen freely
2. go to Iceland, which is also in Schengen area
3. enjoy aurora, etc.
4. fly DIRECT from Iceland to Pittsburgh with 400 dollar.
@Toukairinn_FUZZ
and here is another plot comparing the energy consumption for different gaits and controllers.
I think, there is indeed another trade-off in agility vs energy๐
For autonomous legged robots in the wild, local navigation with online SLAM is an essential function. However, the perception is not always ideal-- how to save in such cases?
Our exploration: use e2e RL to react to perception failures!
video:
Today's people/investors are far more tolerant of the new era grandma bots
and seem to have a weird belief that AI and data can make humanoids break the fundamental limits of hardware systems.
ASIMO is essential on any list of all-time real-life robots.
Honda unveiled ASIMO in 2000 after being developed in secret since the 1980s. It could walk, dance, handle objects, and ran shockingly fast โ it held the fastest humanoid robot record (9 kph) for over a decade.
For autonomous legged robots in the wild, local navigation with online SLAM is an essential function. However, the perception is not always ideal-- how to save in such cases?
Our exploration: use e2e RL to react to perception failures!
video:
Our works (Human to Humanoid and Agile But Safe) are featured by IEEE Spectrum Video Friday!
@TairanHe99
@_wenlixiao
Twice within a month as the cover videos -- thanks for sharing!
The RL policy: win the DARPA subterranean challenge, hike on Alps without a fall, published on Science
#Robotics
๐ฅฐ
Also the same policy: falls dozens of times during hundreds of hours test in the wild ๐จ
And now we find it can fail under small noises and reasonable commands ๐ค
RL controllers for quadruped locomotion are great but black-box. Are they safe enough? Our work in
#RSS2024
reveals the risk of small noises and low-frequency velocity commands falling SOTA controller, which helped win DARPA SubT Challenge, even on the simplest flat terrain.
Humanoid is an interesting topic and works on it can bring innovative techniques for nonlinear multi-contact system control -- you don't need to trust those commercial boosters for making it work in warehouse
ADA regulations in the US ensure virtually all public spaces where robots might work *must* be accessible by wheelchairs. Through extensive industry research and customer interviews, we've identified only two environments where anthropomorphic humanoid robots have a strict
LLMs won't enable this -- need more advanced perception & low-level control & fast reaction
Language intelligence is not all you need. At least we still need motor intelligence.
Finally someone tried to roughly reimplement
@ki_ki_ki1
's SR work on
@UnitreeRobotics
quadruped lol!
Not thoroughly tho (didn't guide the encoders, DR not enough), but all in good ways (elev map recon and state estim. --there's no official api AFAIK)
For autonomous legged robots in the wild, local navigation with online SLAM is an essential function. However, the perception is not always ideal-- how to save in such cases?
Our exploration: use e2e RL to react to perception failures!
video:
stop hyping unitree and check what Jemin did
Actuation Transparency is so FKING IMPORTANT for such [agile but precise] motions. It may be unnecessary for a commercial application, but necessary for top-tier control research.
I feel the same during my WoCoCo and earlier works.
Gonna travel around US next week and visit my friends.
Currently in my plan:
UT Austin
Caltech
Stanford
anyone wants to catch me -- ping me and I will update my list
Then what's in the future?
More powerful generative models, more modalities for weakness identification, and more effective co-design evolution.
And someone needs to put them together.
The examples here are with legged robots, but the philosophy is above.
We derive a bound on this sample error and find that back-propagating through contact and long trajectories drastically reduces gradient accuracy.
๐งต(3/4)
With a short trend,
AI people mind: it's exp(t)
Control people mind: it's 1-exp(-t)
I still feel very strong limitations for biped platforms (also for humans)
Moreover, hard problems still exist and are not solved by (even scaling of) cheap systems
@keenanisalive
Simply take the derivative of the gravitational potential w.r.t. each block's orientation (see the whole system as an articulation where the joints are the "corners").
#bostondynamics
has a very fast-paced dev team. Even when they were only doing MPC - if you didn't feel I'd say you don't understand control.
Now it's too easy for them to add some learning things. Everyone doing research on
#humanoid
must be clever about choosing the topic now.
IMO, the most important point of such works is that, we can do whole-body behavior cloning for complex tasks in the future.
#humanoid
is important because they are anthropomorphic -- more than being able to locomote and manipulate
๐ค Introducing H2O (Human2HumanOid):
- ๐ง An RL-based human-to-humanoid real-time whole-body teleoperation framework
- ๐ Scalable retargeting and training using large human motion dataset
- ๐ฅ With just an RGB camera, everyone can teleoperate a full-sized humanoid to perform
โฆ
WoCoCo explained.
One RL framework for
#humanoid
sequential contact tasks, e.g., parkour jumping, clap-tap dancing ... with task-agnostic sim2real.
(Previously shared once, but this seems to create a YouTube preview friendly to twitter people)
@Stone_Tao
@RoboDepot
cool and close to being deployable. You may try to penalize (1) ground reaction force (2) foot velocity when contact (slippage) (3) foot acceleration (4) joint acceleration. Then the motion will be natural
Water bird
* Locomotion:
Biped walk โ
Swim โ
Fly โ
* Manipulation:
With beak โ
With wing โ
With feet โ
* Supported API
Language โ
Vision โ
Tactile โ
Smell โ
Sound โ
humanoid < water bird
Wthhhh
Such straight running while dragging the gantry.
It says 3.5m/s but I estimate the peak velo to be around 4m/s.
Also the gait is very stable-can be further accelerated and sustain long running.
Disclaimer: these are some personal feelings for some people curious about experience with unitree H1. It's only about the very specified one robot. Nothing about unitree support & others.
@EugeneVinitsky
In control we often say f(x,u) = Ax + Bu + w
w is the noise -- so it looks quite reasonable to me
and in robust control, we need to optimize u under w*=sup_{w} cost
so in robust rl: optimize for largest noise can sometimes be even more effective (tho hard to explore)
>>> and these AGI people successfully confused general people about robot and AI
(Talk to random people that I work on robotics โ "oh you study AI! Will they destroy the world?")
2. Huge backlash (at least 2 deg on each joint of H1)
3. Limited versatility
4. No state estimator provided, no perception solution provided except two cheap sensors
5. Very inaccurate actuation when torque is high
6. Very small ankle torque compared to Cassie
I once thought for good research you need to be serious.
No.
I began my robotics journey for fun, by my own, without any background or advisor. Building robots and making them move -- from building toys to building our life companions.
Who else is excited for Humanoid Robots to take over? I found 34 companies building humanoid robots and made the most complete video featuring every one in existence on the internet! (for today at least)
I spent over a month trying to find them all,
Two recent papers from Nature and Science Robotics show that it's *good* (not just possible) to use event cameras/sensors for mobile robots:
We need advanced perception for animal-level advanced sensorimotor skills.
@Stone_Tao
@EugeneVinitsky
GPUs are better at rendering (ray tracing). Indeed the accuracy of dynamics is still bad on GPU. the contact forces can be arbitrarily huge on PhysX tho the motion looks "normal". And I don't feel mujoco-gpu does better. CPU+RLlib was a good solution (case: Cassie)
I used to save my paper reading notes locally but I never open them again.
So I'm considering putting them on twitter ->
@RoboReading
Reading notes of new papers will be shared and discussions are welcomed.
Excited to share our latest work, Vision-Language Frontier Maps โ a SOTA approach for semantic navigation in robotics. VLFM enables robots to navigate and find objects in novel environments using vision-language foundation models, zero-shot! Accepted to
#ICRA2024
!
๐งต
@CSProfKGD
I would vote for yes if there are practical ways that can
1) clearly separate each author's contributions in big projects,
2) consider those failed trials and discussions into contributions
3) make it reader-friendly without overwhelmingly massive statements
โ ๏ธ How our humanoid robot got more and more tapes on it...
#WoCoCo
And maybe we need a savior controller!
๐ Let's share more funny ๐ค videos like this by retweet!
๐ง We really need safety guarantee for humanoid โผ๏ธ Otherwise agile humanoid becomes fragile robot๐ญ
We are really fascinated by recent humanoid demos in academia that trying to pushing the limit of
- autonomous: HumanPlus, OmniH2O;
- agility: Humanoid Parkour Learning, WoCoCo.
@KyleMorgenstein
I would care more about sim2real pipelines. There are no mature MJX solutions for sim2real (though Google released one at some initial stage, results are not convincing at all) -- then even it's 2x faster I wouldn't use it.
@the_real_btaba
@KyleMorgenstein
Technically I want
1) procedure generation of random terrains
2) perceptive policy training with raw sensor obs
3) NN-based system ID from real to sim
4) neural state estimation from raw obs
Many of these simply need a few lines of easy codes though.
Imitation learning worksโข โ but you need good data ๐ฅน How to get high-quality visuotactile demos from a bimanual robot with multifingered hands, and learn smooth policies?
Check our new work โLearning Visuotactile Skills with Two Multifingered Handsโ! ๐
Each of these works has pros and cons -- so we need to build upon them towards more perfect solutions.
btw I do think Expressive Whole-Body Control by
@xuxin_cheng
is also a good work and close to the topic!
also, PHC from
@zhengyiluo
is great!
@_akhaliq
Rly inspiring. But the tracking "flows" are on the task irrelevant background, which can make it challenging to generalize in the wild. There needs to be some "attention"
@kevin_zakka
No worries. Others cannot cite your work if they cannot achieve something based on yours. LM things are far easier to implement than robotic things, but very few really โworkโ (e.g., in production). Hence, you can always expect you are cited by only high-quality impactful papers.
Figure CTO - Jerry Pratt.
20 years. Worked for "Institute for Human and Machine Cognition" and accomplished much? No.
11 years. Co-founded "Boardwalk Robotics". Did the bot get commercialized? No.
5 years. Interestingly, worked for a non-profit "Pensacola Mess Hall"? Chill
This is a bit far away from what I expect to be Parkour for humanoid (compared to atlas ofc), but definitely a good job and nice follow up of robotic parkour
Introducing ๐ค๐Humanoid Parkour Learning
Using vision and proprioception, our humanoid can jump over hurdles, and platforms, leap over gaps, walk up/down stairs, and much more.
๐ฅ๏ธCheck our website at
๐บStay tuned for more videos.
The May issue of Science
#Robotics
is out!
Taking inspiration from a robber flyโs eye, researchers developed a lens-free pinhole compound eye with a perovskite nanowire photodetector array. It exhibited a wide field of view and dynamic motion tracking.
Other cases: people doubt why limx dynamics build a wheeled-legged robot with biped mode (or think wheeled legged is a new design) while SwissMile already proves w/ demo for 2 years.
Also, people don't understand why parallel simulation but it brought sota RL on legged in 2019.
Funny as it is
We can buy a $10000 camera for a Mars
#robot
, but only $100 for a street one. Same for special robots vs commercial ones. So it's vital to estimate the upper bound of reliability with a fixed budget before we build the robot.
How? My discussion below๐งต(1/n)
Guess we did make a breakthrough months ago. How fast should we reach as for a safety boudary?
Consider transferring this to wheeled-legged robot + 3D locomotion and see what's gonna happen
but len(todos) seems >10. ๐