We now have full end-to-end manipulation autonomy for more complicated tasks.
Here's a demonstration of Apollo leveraging AI to make a fresh cup of juice in an end-to-end manner for
#GTC2024
.
A thread 🧵
E-Atlas is finally out. And I love the design so much. Better RoM, torque density, and form factor than any robot I've seen in a while. I'll be adding thoughts in this thread.
-> Humanoid company posts video
-> every other humanoid company posts a similar video because why not
-> manipulation community emphasizes they did this a while back and question why their paper(s) didn't get more attention
If you're gonna do layoffs. Just do it honestly and really think about the impact of the people being let go.
Robotics is a small and passionate community.
Adding Mercedes to our list of pilot partners!
Humanoids continue to show close collaboration with the automotive industry.
Apptronik : Mercedes
Figure : BMW
Agility : Ford
Boston Dynamics : Hyundai
Tesla : Tesla
While I question the validity of the source... I would like to expand the sample size for this.
Apptronik QDH: 6 months
Apptronik Apollo: 8 months
Unitree H1: 6 months
Agility Cassie: ~8 months
Agility Digit: ~6-8 months
... if others would like to add their data. Please do.
this is interesting: someone compiled the timeline for getting to bipedal walking
oftentimes ppl think 2-legged walking is the hardest part, this was true a decade ago but not anymore
I don't know what the future holds for humanoids. But I hope at the very least we'll get a bunch of cracked robotics engineers out of it like the DRC era.
There's a point of "little details" here. Concept is not hard. But the stepping and motions are noticeably less controlled. E.g look at the jitter in the robot as it stands. Here swing heights are smaller, timing shorter, and the CoM moves weird while transitioning.
Never heard of Booster Robotics before. The video description says something along the lines of "we wanted to try the electric Atlas getup motion because it didn't seem so difficult".
Congrats to
@UnitreeRobotics
for releasing today.
Looks very familiar to what we did in 6 months last year. Excited to see where they will be in 1 year.
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
Awesome to see more progress on quad->humanoid transfer with similar/the same RL stacks. This definitely builds behavioral confidence!
I think the footstep placement/balance control could still use some tuning. The robot often gets stutter steps or deviates from nominal walking.
Introducing 🤖🏃Humanoid Parkour Learning
Using vision and proprioception, our humanoid can jump over hurdles, and platforms, leap over gaps, walk up/down stairs, and much more.
🖥️Check our website at
📺Stay tuned for more videos.
Looks great. Even if it becomes 2x it's a good deal showing what's possible. I am extremely doubtful about the pricing.
That's less than any of their larger quadrupeds and they did something similar with H1 where they claimed sub 100k and it ended up being like 160k.
Unitree Introducing | Unitree G1 Humanoid Agent | AI Avatar
Price from $16K 🤩
Unlock unlimited sports potential(Extra large joint movement angle, 23~34 joints)
Force control of dexterous hands, manipulation of all things
Imitation & reinforcement learning driven
#Unitree
#AI
Respect to all the cool new things diffusion policies and similar works have done for manipulation. But too many companies are starting to post videos of teleop demos with open promises of Autonomy. The fact that policy and teleop videos keep getting mixed in demos is confusing
I keep seeing companies making humanoids with scrawny arms. But the manipulation is more important than locomotion when it comes to making a platform that can do anything helpful.
This is my understanding of the dofs. I'm leaning toward it being roatry+pullbar for the linear motions based off some contours and rod motions in video. But could still be linear for ankle/torso. Dotted lines represent offaxis motion.
Gonna start with the elephant in the room. The 360 joint design. Namely for torso, neck, and thighs. While it's still in question if this it's research or product, the ability to spin its body will be extremely helpful for fast turning. Which is a major slowdown for humanoids.
I'm so happy to see people floored by the price point of
@UnitreeRobotics
G1. It's the first major humanoid to show the price most humanoid companies have been seeing for the last few years now. The next few years will be wild as every company iterates.
At long last. Time to see what we can run onboard out of the box and what drive code needs to be written from scratch for the iron ox robots with
@GoddenThomas
!
The drive modules themselves come off with a bolted shaft cap, similar to the brake. Looks like there's a separate bolt-on plate to retain the drive shaft bearing.
Core memory of mine from years ago was watching Atlas spew red oil all over the place before slowly dying an extremely anime death.
Excited to see what comes next 👀
We now have full end-to-end manipulation autonomy for more complicated tasks.
Here's a demonstration of Apollo leveraging AI to make a fresh cup of juice in an end-to-end manner for
#GTC2024
.
A thread 🧵
So I hear that behavior cloning is all the rage now. What if we could do better, but with the same data? :) In CCIL, we show that imitation via BC is improved by synthesizing corrective labels to account for compounding error, without interactive oracles. Lets you do 👇! 🧵(1/9)
@facontidavide
Robotics is not a get rich quick business. I don't think a lot of startups recognize this when they try to pace it like a software startup. Product scoping is super important when taking on investor promises.
It seems like the problem is becoming less "what architecture are you using" and more "do you have a system with good data."
I think we need better ways to transfer data between different robot types when it comes to manipulation policies.
"as long as there are enough parameters, and things are reasonably well-conditioned (i.e. a decent number of nonlinearities and and connections between the pieces) then it really doesn't matter how you arrange them, i.e. any sufficiently good architecture works just fine"
Always fun to see new humanoids. Notes
- Still pretty expensive (granted prototype) because actuation still hasn’t fundamentally changed.
- No demonstration with realistically heavy objects
- Main stability failure scenarios not shown (eg. crossed legs/blind height changes)
China's progress in humanoid robots deserves more attention.
The video below has <300 views on YouTube, while the robot appears to be
- more agile than
@Tesla
's Optimus
- more dexterous than
@agilityrobotics
's Digit
- (likely) a lot cheaper than both
Contemporary discussion (hype?) about LLMs and “pausing AGI development” seems oblivious of Moravec’s paradox.
We’ve hypothesized since the 80s — that the hardest problems in AI involve sensorimotor control, not abstract thought or reasoning.
It
𝗥𝗲𝗮𝘇𝗼𝗻𝗖𝗼𝗯𝗼𝘁 is a low-cost and robust household collaborative robot.
Check out our first teleoperated dual-armed prototype: chest-mounted control, arm elevation, mobile chassis, and lemons 🍋
(full video → )
Hands. Hands. Hands. 2 fingers is more than enough. 3 if you're fancy. Looks like they're keeping to the Barrett-like hands they were using for atlas, which will to just about any task you want while staying robust. Human hands are too prone to breaking.
✨🤖 Robust Standing and Walking: a highly robust humanoid walking controller.
We introduce benchmarks for quantitatively measuring the capabilities of walking controllers.
The results are not trivial, emphasizing why objective measurements are important for our field. ⬇️
Somebody please explain to me how less than 1 day of filming to make a planned announcement that you actually have a functional product is a significant detriment to "pushing the main robot quality."
@KyleMorgenstein
Convex Optimization has remained one of the most helpful classes in my career. So many problems in Robotics can be formulated as Convex problems. If the professor is good, you'll reuse code from it for a long time.
We now have full end-to-end manipulation autonomy for more complicated tasks.
Here's a demonstration of Apollo leveraging AI to make a fresh cup of juice in an end-to-end manner for
#GTC2024
.
A thread 🧵
The motions are lifelike. But I'm honestly just as mesmerized by the exquisite material and paint jobs used for the costumes and skin. Curious how robust they are to wear & tear.
NEW: Disney showcased the new Audio-Animatronics figures coming to Tiana’s Bayou Adventure at Walt Disney World and Disneyland Resort in the first episode of “We Call It Imagineering,” an all-new series from Walt Disney Imagineering that premiered today on YouTube.
@roboticseabass
Pinocchio is great! One of the best dynamics libraries out there for robots. Fast and Accurate. I also highly recommend the robotics winter school that MEMMO put out that involves it:
Breaking down what you see in the video.
Behaviors are learned from a teleoperated human demonstrator.
We process the camera, hand pose, and grasp state to train a library of closed-loop dexterous manipulation behaviors. A task executor then decides which behavior to run.
Astute viewers may notice that we don't need to return to a "home" position with the arms to start a new action. It's something you may notice on other robots. Instead, our lower level planning/control handles l blending between behaviors to minimize human demonstrations per task
I don't think they would commit to such a reversible design if they didn't have a REALLY good custom slip ring design that actually scales. They'd have to be works of art and manufacturable. Either that or it's very careful multi-turn coils of wiring with good software checks.
@roboticseabass
Drake, Crocoddyl, and OCS2 all have some form of support for it via FCL. FCL is still one of the more performant and well supported options to calculate signed distance.
📢 Releasing TRI's open-source Mamba-7B trained on 1.2T tokens of RefinedWeb!
Mamba-7B is the largest fully recurrent Mamba model trained and is a state-of-the-art recurrent LLM. 🚀🚀🚀
Open X-Embodiment wins the Best Paper Award at
#ICRA2024
🎉🤖!
An unprecedented Best Paper 170+ author list (most didn’t fit on the slide) may be a record for ICRA! So amazing to see what a collaborative community effort can accomplish in pushing robotics + AI forward 🚀
All behaviors are driven by NN visuomotor policies. Lower-level motion planners and controllers are then used to track the policy, handle transitions, and maintain posture.
This strong behavioral baseline is a foundational need for more general purpose logic and task execution.
Transmissions seem to be heavily ambidex/WAM cable drive inspired. Probably mostly custom motors. Curious to see what their manufacturing process is because cable drives are notoriously hard to make and manufacture at scale. IIRC 1 ambidex took days to tension with cables.
Design wise... I'm seeing some quadruped inspiration for sure. From body shape to some of the linkages. If they wanted to make this more manufacturable and reduce cost. They have a clear path. From what I can see, they may be able to support backpack-like add-ons.
Phil Lord and Chris Miller’s ‘PROJECT HAIL MARY’, starring Ryan Gosling, will release in 2026 in theaters.
The film follows a man who wakes up from a coma afflicted with amnesia & soon remembers he was sent 12 light-years away from Earth to save humanity.
I like the Pixar lamp head. And the ring lights should also give consistent/good exposure and contrast for object detection because inverse square law.
Following on that backpack train of thought. Having a 360 yaw body means that if you have battery in front and backpack on back and you can easily rotate to access the battery.
As we already saw with
@Figure_robot
.
You can do some incredible tasks when you have language models to interpret action execution.
While we don't demo voice/text interactions, we are closer to it than some may give credit thanks to the recent progress in multimodal LLM'S.