机器人视频基准：Gemini 4 Argon 标注得分 49.6 落后 nanoannot-1
新发布的 MicroAnnotationBench 上，自研 nanoannot-1 得 54.8 领先所有前沿模型，Gemini 4 Argon 以 49.6 居次但成本高 10 倍；MicroPlanBench 则由 GPT-6 Astra 领先。
Today we’re releasing two benchmarks for robot learning from egocentric video. MicroPlanBench: can a model predict the next action? GPT-6 Astra leads but still misses 1 in 3. GPT-6.1 Sol is essentially tied at about 1/5 the cost. MicroAnnotationBench: can a model label long clips of real work, with exact timing? Our nanoannot-1 scores 54.8, ahead of every frontier model we tested. The best, Gemini 4 Argon, scores 49.6 at 10x the cost. One takeaway: knowing what happens isn’t knowing when. The GPT models top next-action prediction, but on minute-long clips they cut 50%+ more segments than human annotators and miss most boundaries. Timing is its own skill. Links below.