Leaked GPT-6 Astra Scores 46% vs 12% for MolmoAct2 on Bimanual Robot Tasks
damianplayer · x · 2026-09-11
Blogger chooijeq published (unofficial) benchmark results: GPT-6 Astra reportedly scored 46% vs 12% for MolmoAct2 — a state-of-the-art robotics VLA — across 200 trials on five bimanual manipulation tasks, a 3.9x gap.
If accurate, this suggests a general-purpose foundation model dramatically outperforming specialized robotics VLA models, widening the gap as OpenAI moves into embodied manipulation. A thread with test details is linked.
More from Embodied
- Phillip Isola: Astra works because it leverages human-engineered methods, not despite them — phillip_isola · 2026-09-12
- Figure crosses 86,000 weekly active users uploading data for robotics dataset — adcock_brett · 2026-09-11
- OpenArm: Tokyo startup enactic open-sources a humanoid robot with full CAD, firmware and sim tooling — lukas_m_ziegler · 2026-09-11
- NTU's Facet-0 robot model lifts precision assembly success from 16% to 82% — jiqizhixin · 2026-09-11
- UMR: open-source unified motion retargeting for humanoids via learned point-cloud correspondence — ChongZzZhang · 2026-09-11
- Autonomous Lamp open-source companion robot adds dust, air, temperature and humidity sensing — dee_hw · 2026-09-11