Testing visual off-policy RL paper: poor results on real tasks

eigenron · x · 2026-08-25

A researcher questions the practicality of a visual off-policy RL paper. Despite showing fast convergence in sim, their replication on a YAM arm with ManiSkill3 yielded poor results, suggesting the background handling acts as a black box for real tasks.

Original post →

More from Embodied

Embodied channel →