走出外包式合作 AI开始进入创新药企的骨架里
走出外包式合作 AI开始进入创新药企的骨架里机器人控制测试发现智能体会误报任务成功
RoboARCade把长序列操作任务放进同一套MuJoCo仿真环境，涉及三种机器人本体。团队扩大了原任务的初始条件范围，让智能体必须在执行过程中利用记忆、判断和视觉信息。综合测试中，GPT-6 Astra得分49.8%，Claude Fable 5.1为43.6%。更丰富的视觉界面带来1.5至6.8个百分点的改善，使用成本同时增至原来的1.7至2.3倍。 Satvik Sharma等人的这项研究收录在11月6日更新的计算机科学预印本列表中。
错误分析显示，主要困难在于选择操作策略，或把正确策略执行得足够精确。智能体在尝试过的子任务中，有24%至39%出现错误宣告成功的情况。在另一个五任务子集上，提示中加入一段成功示范后，Astra从53.5%提高到73.0%。后者属于子集测试，不能当作整套基准上的得分。团队公开了代码、视频与执行轨迹。
https://lightroom.adobe.com/shares/6dd8a780ffaebc45922405581e243fce
https://lightroom.adobe.com/shares/a15b806e31a5eb49b87c9fd7f9deb3e8
https://lightroom.adobe.com/shares/82011339c28b0a469666e9a9f951a788
https://lightroom.adobe.com/shares/17822a0fbf85564f842fde65559b4ecb
https://lightroom.adobe.com/shares/4d09ca2031443542ae66aab4351e25c9
https://lightroom.adobe.com/shares/b96064e4dd28ae4785eb1e6588890e22
https://lightroom.adobe.com/shares/e3b7e26851ae6e4f95393b299d5fdc57
https://lightroom.adobe.com/shares/0aaaad2f81b4704c8f4e04b727bcb804
https://lightroom.adobe.com/shares/fd49bfc42084ee4885734b4d602cf465
https://lightroom.adobe.com/shares/cd3920394204eb46a2aa6c9c6c86abb5
https://lightroom.adobe.com/shares/87ed15dcf4593242966ecac448e878bf
https://lightroom.adobe.com/shares/fbf8715b1f3996438b43488c153d4c83
https://lightroom.adobe.com/shares/4a4d95b8008a13489aaf2b755ea117bc
https://lightroom.adobe.com/shares/866b69aabf892d4ba2bd6bdaada952eb
https://lightroom.adobe.com/shares/33d4ffdacedb244aa3bf0d4a91ed6a1e
https://lightroom.adobe.com/shares/bfc3acf337190a4da034878f7e4c0859
https://lightroom.adobe.com/shares/f030354141914a41b2c48de2a89a1c29
https://lightroom.adobe.com/shares/16af9b6d84781f468ea622efca33c2df
https://lightroom.adobe.com/shares/f0c42b2ecfb9a449909a05b3a4bc703a
https://lightroom.adobe.com/shares/9ff371203cfa88488c150b5f05ed9792
https://lightroom.adobe.com/shares/015d5b63e4ee454e88670f03cc93e092
https://lightroom.adobe.com/shares/d84c6cdeb145264dbb115fbe2abbd4bd
https://lightroom.adobe.com/shares/541850484d0dae4db38c26d314c85d4d
https://lightroom.adobe.com/shares/7407774e846f564cb3554a6e201c3c28
https://lightroom.adobe.com/shares/18af3745e7c31c47aa0eec2d2783819c
https://lightroom.adobe.com/shares/717c14983df3764a8446136d921cfede
https://lightroom.adobe.com/shares/3f55e1f30aed374f8d5fdbce30c102aa
https://lightroom.adobe.com/shares/355c1acc2930df4586195b59a4894109
https://lightroom.adobe.com/shares/cc96e768097f9d44b92e2c881dfe2f2c
https://lightroom.adobe.com/shares/b9b175bbf966d8459974f152465e4548
https://lightroom.adobe.com/shares/9836f1104caee740b681a7961b4a7902
https://lightroom.adobe.com/shares/23485b62eb0a6a4881859b1cef5a2ea5
https://lightroom.adobe.com/shares/80cb28d529df7845a3edbce217fe2e9c
https://lightroom.adobe.com/shares/f78f27327e608a4ea9e54178b4df62b8
https://lightroom.adobe.com/shares/b3276692d6234c41a344943e601d28cd
https://lightroom.adobe.com/shares/9d8bd3e8b56bf24482a8f7d828f75c93
https://lightroom.adobe.com/shares/39267e40b423ad4a84538dfb602762c3
https://lightroom.adobe.com/shares/71456bb9ee07ba4590579553f7a1c76d
https://lightroom.adobe.com/shares/2f8a5f68ec0edf4ba17ea8d22b1ad67a
https://lightroom.adobe.com/shares/9455b47cabde034b8aca90f195c5182e
https://lightroom.adobe.com/shares/5c5434dad8174f4ba2cee7f6c63d2564
https://lightroom.adobe.com/shares/56872198b10b9d488ff4d702916d6721
https://lightroom.adobe.com/shares/0115d30f483f3548a0ffcebf190bd83e
https://lightroom.adobe.com/shares/c21980f30a457741b5f783c9b12815ce
https://lightroom.adobe.com/shares/1161c8fb7de3bd4c9f613e006141e648
https://lightroom.adobe.com/shares/6a7f45870c6b904cb9e3f9a46ad7e20f
https://lightroom.adobe.com/shares/ef85484e618eb24fa4260c416894b702
https://lightroom.adobe.com/shares/7f8725524fe32f4cb7fecd68224491cf
https://lightroom.adobe.com/shares/11d8094e51334c49af0e59735ec13ae8
https://lightroom.adobe.com/shares/c382e3706ae9f9488ed8aaba0a8461f3
帮你顶下哈！！
页:
[1]