全部作者
Bo Ai1, Ali Amin1, Raichelle Aniceto1, Ashwin Balakrishna1, Greg Balke1, Kevin Black1, George Elledge Bokinsky1, Shihao Cao1, Thomas Charbonnier1, Vedant Choudhary1, Foster Collins1, Ken Conley1, Grace B Connors1, James Darpinian1, Karan Dhabalia1, Maitrayee Dhaka1, Jared Di Carlo1, Danny Driess1, Michael Robert Equi1, Adnan Esmail1, Yunhao Fang1, Chelsea Finn1, Catherine Glossop1, Thomas Godden1, Ivan Goryachev1, Lachy Groom1, Haroun Habeeb1, Hunter Hancock1, Karol Hausman1, Gashon Hussein1, Victor Hwang1, brian ichter1, Connor Jacobsen1, Szymon Jakubczak1, Rowan Jen1, Tim Jones1, Gregg Kammerer1, Benjamin Katz1, Liyiming Ke1, Mairbek Khadikov1, Chandra Kuchi1, Marinda Lamb1, Devin LeBlanc1, Sergey Levine1, Xinyu Li1, Adrian Li-Bell1, Vladislav Lialin1, Zhonglin Liang1, Wallace Lim1, Yao Lu1, Enyu Luo1, Vishnu Mano1, Nandan Marwaha1, Aikys Mongush1, Liam Murphy1, Suraj Nair1, Karl Pertsch1, Allen Z. Ren1, Gavin Schelske1, Charvi Sharma1, Baifeng Shi1, Lucy Xiaoyang Shi1, Laura Smith1, Jost Tobias Springenberg1, Kyle Stachowicz1, Will Stoeckle1, Jiaming Tang1, James Tanner1, Shalom Tekeste1, Marcel Torne1, Kyle Vedder1, Quan Vuong1, Anna Walling1, Haohuan Wang1, Jason Wang1, XuDong Wang1, Christopher Whalen1, Sam Whitmore1, Blake Williams1, Charles Xu1, Sukwon Yoo1, Lili Yu1, Wuming Zhang1, Zhuoyang Zhang1, Ury Zhilinsky1
1Physical Intelligence
Physical Intelligence 介绍了 π0.7,这是一个用丰富的多模态提示训练的通用机器人基础模型:提示包括子目标图像、详细的子任务指令,以及描述任务执行方式的元数据(例如质量、失误、速度)。开箱即用时,单个模型在叠衣服、制作意式浓缩咖啡和折纸箱任务上可达到、有时甚至超过经过特定任务 RL 后训练的 π*0.6,并且无需微调就能执行 PI 早期版本中的任务。元数据使模型能够从质量参差不齐的遥操作和评估数据中学习;作者重点介绍了涌现出的组合泛化能力,例如在没有任何 UR5 叠衣数据的情况下,在 UR5 双臂平台上叠衬衫。在问答环节中,他们解释说质量标签主要来自事后的人工标注,而速度是与质量相关性较好的信号之一。