π Excited to present TOPPO at #NeurIPS2026 in Sydney!
PPO isn't weak at multi-task RL. Its critic is.
Balancing the critic lifts vanilla PPO on Meta-World+ MT50 from 78.5% to 90.9% mean success, and from 6.6% to 56.5% on the 10 worst tasks.
π https://t.co/PYf8qoVARZ