Can't wait to have my own cricket playing robot soon! 🏏
We trained a robot to play football.
How? Self-play for 140 years in a virtual World Cup.
Meet the #Messinator:
1
18
1,568
Saketh Rambhatla retweeted
In-context learning!!
We’ve spent years in robotics assuming a new task means new data. Collect demos —> finetune —> eval —> repeat.
S1 is much more elegant and intuitive: show it one video of an unseen and the model just works. No fine-tuning or gradient updates needed.
It can do so many complex tasks, even ones that run for 10 minutes or longer. It’s especially impressive to see completely new behaviors emerge at test time, like flipping a pancake, even though that behavior never appeared in the training data.
It’s amazing to see S1 pick up a new task from a single example - something I really thought was still years away.
Saketh Rambhatla retweeted
Language Models are Few-Shot Learners showed that GPT-3 could tackle almost any language task with zero, one, or just a few examples.
We’ve heard “the ChatGPT moment for robotics” many times. IMO, the definition is similar: a robot that can take on arbitrary new tasks from zero, one, or a few demonstrations without further tuning, even when those demos come from humans, other embodiments, or are suboptimal.
Today, I’m excited that we’re one big step closer : )
Enjoy S1 🚀
skild.ai/blogs/s1
Saketh Rambhatla retweeted
This is the first genuine ChatGPT moment in robotics. I head they are already in production?!
Saketh Rambhatla retweeted
Strong scaling laws for one-shot learning in long-horizon manipulation tasks.
It's only a matter of time until these models are ready for widespread deployment with very little task-specific data.
Latest results from our cracked team at @SkildAI . Putting aside the usual "ChatGPT moment of Robotics, robotics is solved, blah blah blah" tweets that show up after every release, this result in particular is very exciting to me. Long horizon tasks are where you can properly evaluate in-context learning (ICL) capabilities. Long horizon tasks (5-10 atomic tasks stacked together) have a much higher chance of revealing a model's true capabilities than simpler tasks like stacking cups/bowls. A long horizon task made of 5 atomic actions at 90% success each only nets you ~60% end to end!
A few more thoughts/observations in 🧵.
Introducing S1, our new foundation model that learns from one example.
It can be taught 10-minute long tasks that it has never seen before, from one video prompt without any fine-tuning.
Watch S1 operate in real-time via in-context learning:
2
57
2,660
My personal favorite: one example is fine, but what does it take for supervised finetuning to "catch up" to the ICL model? Turns out it takes ~380 teleoperated demos for S1. Curious if folks at other labs/companies working on ICL have seen a similar trend, let me know if you've seen similar analyses elsewhere. One thing I think is missing from this blog: how do we push that 66 number up? Does naively adding more examples keep increasing success rate, or is there a catch? Would love to see where this goes.
1
7
143
Saketh Rambhatla retweeted
Video editing is now available on the Meta AI app!
A team of ~10 people built this from scratch in under a year. It was a fun and amazing team effort, with some of the nicest people to work with.
Saketh Rambhatla retweeted
Presenting video editing (up to 10s videos) in Meta AI! Generate high-fidelity edits fast⚡(under 30 seconds) with no usage limits, powered by our model trained in-house last year! Took a while to get this out but excited to share with the world 🧵below with all the capabilities
Saketh Rambhatla retweeted
Warehouses are a critical part of the modern economy, but are still bottlenecked by human labor.
We've acquired Fetch Robotics to tackle previously hard-to-automate parts of warehouse logistics with our omni-bodied AI!
(Our factorio players are delighted)
We have acquired Zebra Technologies’ robotics arm (formerly Fetch Robotics).
This is what happens when orchestration meets intelligence -- a major step toward fully autonomous warehouses.
More robots. More environments. One unified brain.
Saketh Rambhatla retweeted
We hosted Prof. Alyosha Efros (UC Berkeley) at @SkildAI! He didn't believe that robots could actually cook eggs reliably. :)
Tested back-to-back 5times without fail! One batch of scrambled eggs every ~2.5mins nonstop. The same model assembles a GPU on a server rack too.
Skild brain at work at GTC!!! 🚀🚀🤖🤖
Nearly every system today, from energy to chips to food, is bottlenecked by scarce human capital.
We are changing that by building AI-powered industries of the future.
Check out Skild Brain robustly assembling GPU racks, a highly precise task, live at #NvidiaGTC.
5
188
Saketh Rambhatla retweeted
At @SkildAI, we’ve raised $1.4B, bringing our valuation to over $14B.
We’re on a generational mission, and I’m grateful to be working alongside an exceptional team. Thanks to our investors for the long-term conviction towards omni-bodied intelligence 🚀
bloomberg.com/news/articles/…
Announcing Series C
We’ve raised $1.4B, valuing the company at over $14B
With this capital, we will accelerate our mission to build omni-bodied intelligence 🚀
skild.ai/blogs/series-c
Announcing Series C
We’ve raised $1.4B, valuing the company at over $14B
With this capital, we will accelerate our mission to build omni-bodied intelligence 🚀
skild.ai/blogs/series-c













