@tonyzzhao So exciting to see what Memo can do now with ACT-2. Fine-tuning with a single example and achieving 99.1% accuracy?? That’s truly special!
@tonyzzhao one fine-tuning example and it generalizes to held-out setups... yeah that'd blow my mind too
@tonyzzhao still in awe!
Users are excited about a robotics model learning new behaviors like t-shirt folding from a single fine-tuning example because of its impressive generalization, high accuracy, and practical benefits.
Based on 5 visible X reactions from 7 accounts; directional sample.
Ask a question below.
Published answers will appear here.
@tonyzzhao So exciting to see what Memo can do now with ACT-2. Fine-tuning with a single example and achieving 99.1% accuracy?? That’s truly special!
@tonyzzhao one fine-tuning example and it generalizes to held-out setups... yeah that'd blow my mind too
@tonyzzhao still in awe!
more evidence that pretraining data scaling will (eventually) work for robotics, given the right domain and enough data quantity and quality
the robot era is upon us
The robot successfully generalized four folding strategies on unseen setups.
more evidence that pretraining data scaling will (eventually) work for robotics, given the right domain and enough data quantity and quality
the robot era is upon us
Here is the result that blew my mind. https://x.com/tonyzzhao/status/2077806012929851883
@saranormous Well done here.
Here is the result that blew my mind. https://x.com/tonyzzhao/status/2077806012929851883
Just saw the craziest emergent behavior I’ve ever seen in robotics. The team is absolutely losing it. We haven’t posted in a while… should we write it up?
@saranormous Well done here.
the robot era is upon us
We found a general recipe for Solves: scale pretraining, then hill-climb with minimal in-house data. For the first time, one fine-tuning example can teach a new behavior that generalizes. Below: 4 folding strategies, learned from one example each and tested on held-out setups.
more evidence that pretraining data scaling will (eventually) work for robotics, given the right domain and enough data quantity and quality
We found a general recipe for Solves: scale pretraining, then hill-climb with minimal in-house data. For the first time, one fine-tuning example can teach a new behavior that generalizes. Below: 4 folding strategies, learned from one example each and tested on held-out setups.
Users are excited about a robotics model learning new behaviors like t-shirt folding from a single fine-tuning example because of its impressive generalization, high accuracy, and practical benefits.
Based on 5 visible X reactions from 7 accounts; directional sample.
Ask a question below.
Published answers will appear here.
Here is the result that blew my mind. https://x.com/tonyzzhao/status/2077806012929851883
@saranormous Well done here.
