[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"aihot-art-88081":3},{"itemId":4,"vertical":5,"category":6,"source":7,"score":8,"title":9,"summary":10,"analysis":11,"url":12,"coverUrl":13,"direction":13,"marketSignal":13,"publishedAt":14},"88081","ai","论文研究","新智元",66,"银河通用WAM-TTT：首个具身智能大模型测试时后训练框架","银河通用发布WAM-TTT框架，机器人通过观看第一视角视频即可适应新场景，泛化能力超越硅谷大模型。","银河通用发布 WAM-TTT，全球首个具身智能大模型测试时后训练框架，用低成本人类视频替代昂贵机器人数据。\n· WAM-TTT 在跨环境实测中泛化保持率达 76%，而 Physical Intelligence 的 π0.5 成功率砍半。\n· 部署时只需人类用运动相机录制第一视角视频，无需遥操设备或动作标注，成本极低。\n· 人类视频可 1:1 等效替代昂贵的机器人轨迹数据，在相同预算下性能持平。\n· 模型采用“大模型主干冻结 + 专用参数快学习”架构，避免灾难性遗忘。\n影响\u002F看点：该框架大幅降低具身智能部署成本，可能加速机器人进入家庭和工厂的商业化进程。","https:\u002F\u002Fmp.weixin.qq.com\u002Fs\u002FD6S9Ta7O2y8M0wCN9Ey1Fw",null,"2026-07-16 16:18:11"]