MolmoSpaces benchmark launched: GPT-Astra beats all open-source VLA baselines zero-shot

notmahi · x · 2026-09-15

The Molmo team released MolmoSpaces-v1, a benchmark for robotic spatial understanding, asking whether robotics is having its "GPT moment."

Key result: GPT-Astra outperformed all open-source VLA/world-model baselines on a zero-shot subset of the benchmark. The authors published execution traces and argue general-purpose models are showing a leap in spatial reasoning for robotics.

Original post →

More from Embodied

Embodied channel →