to me, showing scaling laws in robot performance with human-hand data is one of the most important problems to solve in robotics. beyond the common narrative that human-hand data is the most abundant data source compared with teleop and umi-style data, there are two more reasons people often miss: - human-hand data is one of the few types of data that is forward-compatible. the promise of little to no data-to-deployment gap with teleop and umi will disappear once your end effectors change. - human-hand data collection is also the least intrusive way to gather physical data without hurting labor productivity. adding any device to human hands will only slow down dexterity and reduce precision. this cannot be emphasized enough if the goal is to reach human-level performance. DYNA2 takes a stab at this challenging but rewarding problem. huge props to Dyna team for pulling this off! if you look closely, you will also find previews of other exciting work happening at Dyna!