Eel news is back baby
Tony Fortunato@tfortunato
The most important part of the morning dog walk is catching up on the latest Eel News
English
Tony Fortunato
1.9K posts
Tony Fortunato
@tfortunato
Builder of fine robots and robot accesories. Chewy Physical AI and Robotics. Prev Boston Dynamics, Amazon Robotics, Ginkgo Bioworks
Rockport, MA Katılım Mart 2009
1.4K Takip Edilen364 Takipçiler
Mode Collapse Goblin is a strong contender for best character of 2026
English
@kevin_zakka @pabbeel Congrats Doctor! You've already had an impact on my work (just shared my latest mjlab run with team this morning!), and I'm sure you'll continue kicking ass in whatever robotics problems you point yourself at!
English
Gave my PhD dissertation talk on Friday!
It's been an incredible journey made possible by the best advisor who believed in me and gave me the freedom and support to explore. Thank you @pabbeel!
And thank you to everyone who came to support and share this milestone with me 🙏
English
@KyleHessling1 Thank you, this is awesome! I've been curious how to get into some of this kind of model hacking, and folks like you sharing are super appreciated!
English
Hello again everyone, the official Qwopus-GLM5.1 frankenmerged 18B is LIVE!
12-16GB GPU bros, today is your day!
The healing fine tune worked beautifully to fix the seam between the two merged models!
It was unusable before the healing run, now it’s surprisingly excellent! Getting a nice mix of Qwopus and GLM 5.1 intelligence in a tiny model!
Of course it would be nothing without my friend Jackrongs two excellent Qwen 9b fine tunes! Qwopus3.5-9B-v3.5 and Qwen3.5-9B-GLM5.1-Distill-v1
Please use it and let me know what’s awesome and what needs work! Send us examples of what it can do!
It one shotted a really nice snake game, as well as some other great front end tests, all raw outputs in the repo! And I’ll link the video of them below as well!
Have fun everyone! This is just the start! I have so many potentially great theories to further this method!
I have included a full documentation of the merge and finetune-heal method in the repo, feel free to experiment with it! Will link that below in this thread as well!
Everything done locally in my RTX 5090!
huggingface.co/KyleHessling1/…
English
Tony Fortunato retweetledi
@i2cjak Big pyrotechnics doesn't want you to know that but any cheap component can be a pyrotechnic initiator if you suck enough at electronics
English
Tony Fortunato retweetledi
Introducing Operator, our newest industrial AI robot built to work, not demo.
Operator handles your warehouse's most repetitive tasks: packing, sorting, and kitting. Up to 24 hours a day, with flexibility and consistency that allows businesses to scale quickly.
This is what we've been building. ↓ [1/4]
English
@kevin_zakka Awesome! I was literally looking for something like this earlier today when I was prepping to port a new bot into mjlab
English
Really excited to release mjviser, a web-based MuJoCo viewer, powered by Viser. It has almost all the features of the native MuJoCo viewer, but runs in your browser. Load and simulate any MuJoCo model with a single uv command 👇
uvx mjviser
English
@lauriewired @FeepingCreature Not entirely the same, but reminds me of an issue we had at a former robot job. The controls required an intense amount of real-time compute. If CPU got too hot and throttled, the controller ran too slow and robot stumbled, causing a shutdown. Basically computes itself to death.
English
yeah, I almost mentioned boost clocks + power envelopes, but that’s a little different because you’re sliding along a continuum. (you’re always running, but just faster or slower)
with the cray 2, it was truly pulsed in the sense that you are turning the whole module on and off based on thermal constraints.
I think the interesting thought is if we’ll ever re-enter this era of a truly pulsed modulation, where the minimum envelope/threshold is still high enough to not be able to function continuously even with downclocking. maybe something like dense space compute with a really strict power budget
English
I’ve always wondered about the possibility of a “pulsed” computer.
That is, a computer so dense, either due to thermal or electrical limits, it can’t be run continuously.
Kind of like a camera flash, intense burst, then a recovery period.
Turns out, it sorta exists.
English
Tony Fortunato retweetledi
Robotics lacks infrastructure, not intelligence.
Everyone wants to build bigger robot models, but most Physical AI papers complain about the same things:
data collection is slow, sim-to-real is fragile, teleop is painful, evaluation is messy, long-horizon control still breaks.
English
@RealJoshuaYang @KyleVedder Like yes, the hardware is going to limit you on a bunch of interesting tasks. Its practicality is as a learning tool, not as an industrial arms you'd take to production. From that POV I think its great I can cheaply give everyone in the office a pair of arms to play with
English
@RealJoshuaYang @KyleVedder I was pretty skeptical at first as well after working on "real" robots, but it really won me over. Its been a great teaching and learning tool for folks in my lab, especially those new to robotics. Nice way to get feet wet with hardware, collecting data, training policies, etc.
English
im often asked “how do i break into robot learning without a phd?”
buy an SO-101 ($300) and do something interesting
understand the stack, train models, implement a paper, do serious on-robot evals, document your work
the pool of talent with real experience is v small
English
Tony Fortunato retweetledi
It took me a while to come to grips with Unitree's game being a bit like PC vs. Mainframe in past. You might not win on the tech merits of the HW at first, but something to be said about quickly getting low cost HW into as many hands as possible.
Chris Paxton@chris_j_paxton
Owning the platform upon which everyone is building worked for Apple. Its going to work for unitree.
English
Tony Fortunato retweetledi
Shout out again to @ZeYanjie for sharing the TWIST2 design. Just finished assembling our first pan-tilt rig, excited to try it out on our bots!
Yanjie Ze@ZeYanjie
Excited to introduce TWIST2, our next-generation humanoid data collection system. TWIST2 is portable (use anywhere, no MoCap), scalable (100+ demos in 15 mins), and holistic (unlock major whole-body human skills). Fully open-sourced: yanjieze.com/TWIST2
English
- Train on picking small rigid items from 63 of 64 spaces on a chess board
- Successfully pick an item from the 64th space in test 40% of the time.
"Dexterous Manipulation generalizing to out of domain data solved"
Chris Paxton@chris_j_paxton
My research shows extremely strong generalization throughout the training set thank you very much
English
@aholtzma @ptrschmdtnlsn Can a connector rated for only 10 cycles stand up to ultrasonic cleaning?
English
@ptrschmdtnlsn Get an ultrasonic cleaner, so worth it. Squeaky clean boards.
English
Can you spot the bit of gunk that might have made me lose days of work? One *huge* downside of connectors only rated for 10 mating cycles is that you try everything you possibly can before you risk reseating the connector.
English
@Sentdex (Also, he's absolutely correct that Robotics community is best community)
English
Love @Sentdex sharing stuff like this. Not just for the technical content, which is great, but for the attitude of "just try things and learn".
Been a huge inspiration for folks at the lab, including myself, an EE turned Roboticist now diving deeper into RL/ML!
Harrison Kinsley@Sentdex
New video is out, teaching a Unitree G1 humanoid to walk using reinforcement learning (PPO). First time I've got sim2real to work with robotics, sharing what I've learned and testing out how good the policy actually is by walking around outside on some semi challenging terrain.
English