Skild AI Shows S1 Model Learning 10-Minute Tasks From One Video
Skild AI's S1 model performs long-horizon tasks never seen in pre-training from single video prompts.
Skild AI founder Deepak Pathak posted that in-context learning for robotics has arrived with the S1 model. It handles tasks over 10 minutes long that were never seen during pre-training, prompted by one video example and without fine-tuning. Investor Alfred Lin at Sequoia called the single-prompt execution of long-horizon tasks a game changer. Skild researcher Roozbeh Mottaghi noted the absence of test-task leakage and the model's ability to manage extended independent execution. Other observers including Chris Paxton and Sequoia Capital highlighted the demonstration as progress toward general-purpose robot foundation models.
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:
Reactions
12 postsDP 
Deepak Pathak@deepakpathakIn-context learning for robotics is here. - Long-horizon tasks over 10 minutes long - Never seen during pre-training - Prompted with one video, no fine-tuning We are building intelligence from the foundations up, not from the top down.
- likes: 343
- replies: 27
- bookmarks: 73
- reposts: 35
JM Jitendra MALIK@JitendraMalikCV@deepakpathak Nice work, Deepak. Congrats! Now you should do it with multi-finger hands, not these silly parallel jaw grippers :-)
- likes: 1
- replies: 0
- bookmarks: 1
- reposts: 0
PG AL Alfred Lin@Alfred_LinWhile robot olympics may get more media headlines, these types of advancements are really what change the game for robotics. Very impressive work on turning single prompts into long-horizon tasks that robots can then execute independently.
- likes: 61
- replies: 2
- bookmarks: 14
- reposts: 4
SA Skild AI@SkildAIIntroducing 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:
- likes: 3.6K
- replies: 296
- bookmarks: 2.2K
- reposts: 666
RS Robert Scoble@ScobleizerThey wanted to pay me to write this post, and I turned them down. Why? Because I wanted to be able to say, without any prejudice: this is a step forward for robotics. Hopefully, I can get them to sponsor something else in the future.
- likes: 6
- replies: 0
- bookmarks: 0
- reposts: 1
