With FastTD3's recipe, QF3 is also fast: it learns G1 motion tracking in ~30 min on a single GPU, ~10x faster than the on-policy FPO++! (and with ~10x fewer env steps)
0:09 1 0 12
0:09
Chung Min Kim
@ChungMinKim
Replying to @ChungMinKim
The same update also makes a pretrained VLA in ABC's simulation (abc.bot) faster and more reliable -- and fine-tunes full flow policies on Robomimic too!
0:35 1 24