DeepMind robotics evals: Astra overtakes Gemini by ~8%, still 10x error gap to internal models

m_wulfmeier · x · 2026-09-10

A DeepMind robotics engineer updated the team's robot deployment benchmarks with fresh numbers on frontier models:

• Astra has finally overtaken Gemini, jumping nearly 8% since the Sol checkpoint — though its rival is only the Gemini flash model, priced at $10/50 per million tokens vs Gemini's $0.75/3.75.
• Gemini has now surpassed the dedicated Gemini Robotics ER 2 model.
• A large gap remains to the team's internal models, which show roughly 10x lower error rates.
• Credit for keeping the perception evals running goes to Michał Grzejdziak-Zdziarski and team.

Original post →

More from Embodied

• Unitree deep dive: Wang Xingxing approves every expense over 100 yuan — ChinaTalk · 2026-09-10
• Wristband camera could make robot data collection far cheaper than exoskeletons — CyberRobooo · 2026-09-10
• Training a Unitree G1 to juggle a soccer ball in MuJoCo — 5 keepy-uppies so far — hbouammar · 2026-09-10
• Underwater embodied-AI startup Gongzhi Ocean raises near-$14M angel round for bionic robot — 创业邦 · 2026-09-10
• Guardian Columnist: Driverless Cars Are Taking Us on a Road to Nowhere — nordicinst · 2026-09-10
• First look at Apple's foldable iPhone Duo leaks — pdamodaran · 2026-09-10

Embodied channel →