Announcement
SteerAct reportedly improves real-robot mean success from 31.1% to 52.2%
A post describes the Stanford–MIT project as steering one feature while keeping VLA and WAM models frozen.
TLDR
A post about SteerAct says the Stanford–MIT project uses 24 online rollouts per task and steers just one feature while keeping VLA and WAM models frozen. It reports real-robot mean success rising from 31.1% to 52.2%, a 21.1-percentage-point gain, or about 68% relative.
Combined views
68
2 Sources, first seen ago
