11,398 real robot trials put Astra out front
RoboColosseum ran 11,398 real-world evaluations across 124 sites and says OpenAI's GPT-6 Astra is seriously strong, lining up with the demos people have been arguing about on X.
Shown this week
RoboColosseum's results land: Astra strongest across 11,398 real trials
A day after launching, RoboColosseum published numbers: 41 days, 124 diverse sites, 11,398 real-world evaluations comparing 5 VLA policies (models that turn images and text into robot actions) and 3 frontier AI agents. Jiafei Duan says OpenAI's GPT-6 Astra is seriously strong and that the X demos line up with what they see at scale.
Scalable evaluation is the bottleneck to robot learning
NVIDIA's VFlash decodes robot actions 12x faster, and more accurately
NVIDIA researchers released VFlash, which outputs a whole trajectory or bounding box in one shot instead of one token at a time, because the output length is known in advance.
every coordinate is written in ink and every later point loses the chance to backspace
Younghyo Park asks what heavy robot arms cost you in learning
Younghyo Park posted findings on whether arm weight hurts robot learning, pointing out that metal links and highly geared motors make arms extremely heavy next to the objects they manipulate. The question he puts to the field is whether learning algorithms are strong enough to absorb that, or whether there is a hidden cost nobody is counting.