@MTSlive I don’t think this counts as distillation, because Claude itself was also trained on knowledge created by other models/people without permission as welll. This is how AI is trained today. Rejecting distillation outright is standing in the way of technological progress.
This new paper is a fascinating read! 🤖📄
Unlike RISE and WMPO, this approach achieves robot policy self-improvement without relying on a world model. It drives the learning only through an ingenious reward design.
This is huge news for post-deployment robot manipulation! 🚀
@danfei_xu@Dantong_Niu I think the egocentric video is the one of the most promising ways to achieve the gpt3 moment of the robotics, and this work just give us some very good guidance and insights.
@JitendraMalikCV@bhawna_paliwal_@HarithejaE@willjhliang@pabbeel@notmahi This is crucial, especially for dexterous hands, where data is extremely scarce. Once human-hand videos can be converted into robot data, we can move beyond RL-only training in simulation, and we can unlock nearly unlimited internet-scale data😂😂