Newer VLA Models Excel in Sim, but π0.5 Still Leads Real-World Eval

DominiqueCAPaul · x · 2026-08-11

Newer VLA models are achieving up to 2x better performance on the RoboDojo simulation benchmark. However, π0.5, a model released 16 months ago by Physical Intelligence, still maintains its lead in real-world evaluations.

This discrepancy highlights a persistent challenge in embodied AI. The author calls on the field to distinguish how much of this gap is driven by model quality versus data quantity and quality.

Original post →

More from Embodied

Embodied channel →