11,398 real-world robot evals across 124 sites confirm GPT-6 Astra is seriously strong
DJiafei · x · 2026-11-07
RoboColosseum published a 34-page technical report: 11,398 real-world evaluations across 124 diverse sites over 41 days, comparing 5 VLA policies and 3 frontier AI agents. Clear takeaway: OpenAI's GPT-6 Astra is seriously strong, and the demos on X weren't cherry-picked. The team also published every failed run, arguing that sharing failures is what makes a benchmark useful rather than just impressive.
More from Embodied
- Home robot picks up any clothing with 97% success on unseen garments — chris_j_paxton · 2026-11-07
- Zurich team shows tactile + mocap glove is all you need for robot retargeting — arnie_hacker · 2026-11-07
- Musk: SpaceX to attempt first-ever Starship upper stage tower catch with Ship 42 in about a week — XFreeze · 2026-11-07
- Wheeled factory humanoid AIR unveiled in Munich, just 9 months from research start — xmercury_one · 2026-11-07
- Musk: 20-30M humanoid robots could match entire US output; Tesla may rent Optimus instead of selling — XFreeze · 2026-11-07
- Open-source robot Microduck production line first runs at Seeed's Shenzhen factory — huggingface · 2026-11-07