Introducing GEN-0, our latest 10B+ foundation model for robots
@GeneralistAI2.7万 位关注者
501,601 次观看
41 条评论
评论来自 X (Twitter) 上的原帖

@GeneralistAI
Surpassing the intelligence threshold 🧠 in the high-data regime for robotics, we observe a phase transition ⚡️ at 7B params where smaller models exhibit ossification under data overload, while larger ones continue to improve and can quickly adapt to new tasks.
162 个赞

@GeneralistAI
Robotics at Generalist is no longer limited by data. GEN-0 is pretrained on our growing dataset of over 270,000 hrs of real, diverse manipulation data (10,000 hrs/week & accelerating). 1,000s of data collection devices/robots in homes, warehouses, and workplaces worldwide🏠🏭💼
86 个赞

@GeneralistAI
We’ve developed a new approach to training models, Harmonic Reasoning, which creates a "harmonic" interplay between asynchronous, continuous-time streams of sensing and acting tokens. ⚙️🎵 Watch GEN-0 pack a camera. 🤖📸
162 个赞

@GeneralistAI
GEN-0 models exhibit strong scaling laws, in which more pretraining data and compute consistently (and predictably) improve downstream post-training performance of the model across many tasks. 🔬📈🧠
104 个赞

@GeneralistAI
📱Packing the phone into the box: here GEN-0 is tasked with gently packing a phone (including its accessory box) into a snug-fit retail box. Slotting in the phone, inserting the accessory box, flipping, and closing the lid, with all tight fits.
74 个赞

@GeneralistAI
📦🧠🤖🔄 Packing unseen objects - this evaluates GEN-0's generalization and “physical commonsense": reasoning over the highly varied geometry and physics of new objects, nudging and reorienting, packing in differently depending on shape, pushing objects around to make space.
49 个赞

@GeneralistAI
Cross-Embodiment – GEN-0 architecture works on different robots by design. We have tested our models on various 6DoF, 7DoF, and 16+DoF semi-humanoid robots.
60 个赞

@GeneralistAI
To scale GEN-0 capabilities, we are constructing the largest and most diverse real-world manipulation dataset ever built 🌍 including every manipulation task humans can think of – from peeling potatoes, to threading bolts – spanning homes, bakeries, warehouses, factories, & more
113 个赞

@GeneralistAI
Read more in our blog post, including early notes from large-scale ablations on our pretraining data. Blog:
75 个赞

@xiao_ted
Amazing results and very promising scaling laws!
10 个赞

@ATTlKA
These demos are, as always from Generalist, incredible. Would love to see more detail on Harmonic Reasoning since that seems like your alternative to PI's real time action chunking. That seems significant enough of an achievement that it warrants a writeup of its own!
6 个赞

@avizurlo
Amazing work! Where'd you source 270k hours of dexterous data?
5 个赞

@grantg07
Are y’all planning on incorporating hands for more complex assembly or just focusing on grippers?
4 个赞

@helper2424
wow, 270k hours is a lot, what is ratio sim/real?
4 个赞

@MarioChan2002
Extremely large data scale.
2 个赞

@JingjingGong_
Impressive work, I am curious about how Harmonic Reasoning works? Anybody any clue?
2 个赞

@SVenkatachalam
Amazing advances for robotics- congrats to the team!
2 个赞

@carlosdponx
jesus guys
2 个赞

@JonathanScholz2
There's a *lot* to be impressed by in this post, but the biggest thing for me is their early commitment to a scalable data collection strategy (UMI-style in their case) This is how robotics 2.0 companies should be built
2 个赞

@_advaitpatel
awesome results!
1 个赞




















