When training ACT-1, we treated data from diverse, long-horizon tasks in the wild as a first-class citizen. This makes generalization the default, not an exception. The capability envelope expands. More to come.

is a visionary btw

Moose’s name is Stan and he usually stays in my room. He joined our office for CowBoy day and has stayed ever since

it's a moose!!!

Congrats Alper! Really happy to see you shining!

Thank you Ruoshi 🙏♥️ You've played a big role in this trajectory :)

if I domain randomize my house enough, can we pls get a memo??? @sundayrobotics

@sundayrobotics what animals do you got

…the moose!

colleagues have been insisting that it's a reindeer

@saranormous it IS a reindeer!

@alpercanbe @saranormous How do we know it’s not a baby moose tho??

@alpercanbe @saranormous Because the antlers are not palmate and flattened

@flooburrito @saranormous hallucinated

finally!! 🙌 robotics + moose 🫎

Nice! Did you end up choosing torchcodec for the cpu/gpu decoding of images ?

yessir

Büyük bir adım attın kardeşim. Daha küçük yaşta buralara kadar geleceğini ispatlamıştın. Yürekten kutluyorum seni. 🥰 Başarıların daim olsun. Seninleyiz.🤜🤛

Teşekkürler Hüseyin Abi ❤️🙏

@chris_j_paxton I’m looking forward to training and testing one and hoping that I can do it myself without a lot of outside help.

Çok güzel.. emeğinize sağlık. Market raflarına ürün dizerken veya seralarda ürün toplama senaryoları da yaptırabilir misiniz? Ellerde üç parmak kullanmak çok verimli. İnsanlar günlük işlerinin %99’unda sadece üç parmağını kullanır. Ellere beş parmak ekleme saçmalığına devam etmemeniz ayrı güzel.

Give it legs and make training boots.

this is rad

Exciting approach to training data! Generalization as default is key. Can't wait for more!

Please don't show at 4/5x accelerated pace. As bullish I'm about personal home robots it's misleading.

Great work with no fluff.

The greatest human invention after pre-chopped onions.

While you should keep trade secret aspects of how your RL works; imho it would attract attention if your company can share “along the way; these approaches failed/ turned out to be dead ends” stories.

finger precision looks tight 🙌 need to get my hands on one for testing. got a few office task automations in mind that could work

i can't help but get lego movie vibes & i love it

Exciting approach to training! Looking forward to seeing the results.

congrats with the launch Alper!

thanks Kate 🙏

Congratulations. This is so cool!

Thank you Changhao 🫡

It's exciting to see such a focus on diverse data!

most humanoid robots have human-like legs to imitate bipedal locomotion, but Sunday's robot has a wheel base for movement. was this a deliberate choice because you found fully imitating human biomechanics to be ineffective for the tasks you're targeting? curious if the trade-off was stability/energy efficiency vs terrain adaptability, or if the manipulation tasks mattered more than the locomotion form factor?

@EMostaque Interesting approach. Did you see any patterns across the different tasks that surprised you?

I hate my socks folded like that though. How does he do with folding shirts? Folding shirts is the worst chore.

Making generalization the default is huge, can't wait to see how far that capability envelope goes

Exciting approach to training! Looking forward to seeing the results.

but can it fold socks in a submarine?

Did you face issues in dataset training, synthetic vs real implemtations etc? Can we talk about it? Just curious in robotics intelligence

Very cool robot! Looks like an evolved form of the Aloha

What is going on with that PCB? I cant read the descriptor but this is going to come bite you later during manufacturing.

Hello Alper, I wanted to contact you about the robots podcast, but your DM is closed. Please follow me if you're interested!

Models that learn across long tasks unlock everything next