Vol.232 产业观察44｜具身智能+AI4S，能诞生什么新物种？

[517.0s -> 518.0s] 确实对中国的, [518.0s -> 519.0s] 机器人的, [519.0s -> 520.0s] 研发, [520.0s -> 521.0s] 和, [521.0s -> 522.0s] 生产制造能力, [522.0s -> 523.0s] 是绝对认可的, [523.0s -> 524.0s] 所以当我们, [524.0s -> 525.0s] 真正做这么一个, [525.0s -> 526.0s] 机器人大会, [526.0s -> 527.0s] 是真正兑现的, [527.0s -> 528.0s] 它叫, [528.0s -> 529.0s] 世界机器人大会, [529.0s -> 530.0s] 这个比例还是很高的, [530.0s -> 531.0s] 比如说, [531.0s -> 532.0s] 到我们这儿, [532.0s -> 533.0s] 展台的就很多这样的, [533.0s -> 534.0s] 就希望跟我们合作, [534.0s -> 535.0s] 能不能帮你们, [535.0s -> 536.0s] 其实推广到, [536.0s -> 537.0s] 这个欧洲啊, [537.0s -> 539.0s] 包括这个中东啊, [539.0s -> 540.0s] 这个日韩也有, [540.0s -> 541.0s] 包括像, [541.0s -> 542.0s] 半岛电视台, [542.0s -> 543.0s] 就是中东, [543.0s -> 544.0s] 这个, [544.0s -> 545.0s] 半岛电视台, [545.0s -> 546.0s] 特地跑到我们展台, [546.0s -> 547.0s] 就一直在我们这儿, [547.0s -> 548.0s] 这个, [548.0s -> 549.0s] 举景, [549.0s -> 550.0s] 跟我们去聊, [550.0s -> 551.0s] 也交流一下, [551.0s -> 552.0s] 我们对整个, [552.0s -> 553.0s] 行业, [553.0s -> 554.0s] 包括整个, [554.0s -> 555.0s] 机器人的发展, [555.0s -> 556.0s] 如何在, [556.0s -> 557.0s] 他们那边, [557.0s -> 558.0s] 去推广起来, [558.0s -> 559.0s] 他们对这方面, [559.0s -> 560.0s] 是极其关注, [560.0s -> 561.0s] 也认可我们现在, [561.0s -> 562.0s] 真正的逐渐, [562.0s -> 563.0s] 成为一个, [563.0s -> 564.0s] 事实上的, [564.0s -> 565.0s] 机器人的一个, [565.0s -> 566.0s] 全球的中心。 [566.0s -> 567.0s] 对,好, [567.0s -> 568.0s] 那你觉得, [568.0s -> 569.0s] 接待的这些, [569.0s -> 570.0s] 外国友人里边, [570.0s -> 571.0s] 有多少是, [571.0s -> 572.0s] 带着, [572.0s -> 573.0s] 不是学习, [573.0s -> 574.0s] 而是, [574.0s -> 575.0s] 商业和, [575.0s -> 576.0s] 需求目的, [576.0s -> 577.0s] 来, [577.0s -> 578.0s] 确定想找, [578.0s -> 579.0s] 解决方案, [579.0s -> 580.0s] 或者想找, [580.0s -> 581.0s] 机器人的, [581.0s -> 582.0s] 有多少是, [582.0s -> 583.0s] 就是, [583.0s -> 584.0s] 看看热闹, [584.0s -> 585.0s] 说, [585.0s -> 586.0s] 大概是个什么样, [586.0s -> 587.0s] 因为毕竟, [587.0s -> 588.0s] 中国现在免签, [588.0s -> 589.0s] 所以说, [589.0s -> 590.0s] 来变得比以前, [590.0s -> 591.0s] 容易了很多, [591.0s -> 592.0s] 就是, [592.0s -> 593.0s] 前者, [593.0s -> 594.0s] 希望真的, [594.0s -> 595.0s] 能够有, [595.0s -> 596.0s] 实质的合作, [596.0s -> 597.0s] 至少是, [597.0s -> 598.0s] 尝试, [598.0s -> 599.0s] 实质的合作, [599.0s -> 600.0s] 他们去当地, [600.0s -> 601.0s] 就是, [601.0s -> 602.0s] 去做这样的事情, [602.0s -> 603.0s] 这样比例, [603.0s -> 604.0s] 是很高的, [604.0s -> 605.0s] 而非, [605.0s -> 606.0s] 就是像后者, [606.0s -> 607.0s] 可能后者, [607.0s -> 608.0s] 我们也没有怎么, [608.0s -> 609.0s] 去交流, [609.0s -> 610.0s] 如果是, [610.0s -> 611.0s] 前者, [611.0s -> 612.0s] 带着, [612.0s -> 613.0s] 需求来的, [613.0s -> 614.0s] 就你能够, [614.0s -> 615.0s] 接触到的, [615.0s -> 616.0s] 这个小样本, [616.0s -> 617.0s] 范围里, [617.0s -> 618.0s] 他们提出的, [618.0s -> 619.0s] 需求有多少是, [619.0s -> 620.0s] 可行的, [620.0s -> 621.0s] 有多少是, [621.0s -> 622.0s] 还暂时, [622.0s -> 623.0s] 机器人能力上, [623.0s -> 624.0s] 范围之外的, [624.0s -> 625.0s] 我觉得, [625.0s -> 626.0s] 其实, [626.0s -> 627.0s] 从能力来讲, [627.0s -> 628.0s] 绝大多数, [628.0s -> 629.0s] 是都可以做到的, [629.0s -> 630.0s] 因为, [630.0s -> 631.0s] 在机场上, [631.0s -> 632.0s] 要上一个机床商量, [632.0s -> 633.0s] 我们要做一个, [633.0s -> 634.0s] 装配, [634.0s -> 635.0s] 甚至有时候, [635.0s -> 636.0s] 我们要装配自行车, [636.0s -> 637.0s] 我不知道他们, [637.0s -> 638.0s] 为什么,

[638.0s -> 639.0s] 当地做自行车, [639.0s -> 640.0s] 应该从中国买, [640.0s -> 641.0s] 当地还是有一些, [641.0s -> 642.0s] 这样的, [642.0s -> 643.0s] 提议劳动, [643.0s -> 644.0s] 包括这个, [644.0s -> 645.0s] 超市服务的, [645.0s -> 646.0s] 包括像, [646.0s -> 647.0s] 找到我们这边, [647.0s -> 648.0s] 有很多做, [648.0s -> 649.0s] Bio-type公司, [649.0s -> 650.0s] 有做治药公司, [650.0s -> 651.0s] 他们是, [651.0s -> 652.0s] 实实在在, [652.0s -> 653.0s] 有这样的诉求的, [653.0s -> 654.0s] 如果这些诉求, [654.0s -> 655.0s] 发生在中国, [655.0s -> 656.0s] 我们同样, [656.0s -> 657.0s] 肯定也是能做的, [657.0s -> 658.0s] 但发生在海外, [658.0s -> 659.0s] 是不是, [659.0s -> 660.0s] 这时候, [660.0s -> 661.0s] 可以去, [661.0s -> 662.0s] 投入大量的人力, [662.0s -> 663.0s] 去, [663.0s -> 664.0s] 双方协同起来, [664.0s -> 665.0s] 因为当地, [665.0s -> 666.0s] 这方面的, [666.0s -> 667.0s] 工程师, [667.0s -> 668.0s] 其实是, [668.0s -> 669.0s] 适缺, [669.0s -> 670.0s] 是对, [670.0s -> 671.0s] 数量, [671.0s -> 672.0s] 是低于国内, [672.0s -> 673.0s] 所以, [673.0s -> 674.0s] 这个可能是一个挑战, [674.0s -> 675.0s] 所以, [675.0s -> 676.0s] 投入肯定是更大一些的, [676.0s -> 677.0s] 另外一方面, [677.0s -> 678.0s] 当然这个, [678.0s -> 679.0s] 地位政治, [679.0s -> 680.0s] 肯定也是一个因素的, [680.0s -> 681.0s] 这个不可控, [681.0s -> 682.0s] 你刚才已经隐含了, [682.0s -> 683.0s] 我想, [683.0s -> 684.0s] 好奇的另外一个小问题, [684.0s -> 685.0s] 就是, [685.0s -> 686.0s] 听起来假定, [686.0s -> 687.0s] 需求绝大部分, [687.0s -> 688.0s] 就看起来, [688.0s -> 689.0s] 当时谈的, [689.0s -> 690.0s] 没有谈到价格, [690.0s -> 691.0s] 就是, [691.0s -> 692.0s] 从他们, [692.0s -> 693.0s] 角度讲, [693.0s -> 694.0s] 初期的, [694.0s -> 695.0s] 我觉得, [695.0s -> 696.0s] 小量, [696.0s -> 697.0s] 肯定都是可以做的, [697.0s -> 698.0s] 然后, [698.0s -> 699.0s] 如果真的上量的话, [699.0s -> 700.0s] 这个要看, [700.0s -> 701.0s] 还是回到, [701.0s -> 702.0s] 之前的本质, [702.0s -> 703.0s] 你是不是能够, [703.0s -> 704.0s] 可靠的去做, [704.0s -> 705.0s] 比如说, [705.0s -> 706.0s] 给人, [706.0s -> 707.0s] 推一台, [707.0s -> 708.0s] 这一台, [708.0s -> 709.0s] 是一个, [709.0s -> 710.0s] 特定了一个东西, [710.0s -> 711.0s] 不管怎么定价, [711.0s -> 712.0s] 上方可以谈, [712.0s -> 713.0s] 但是你, [713.0s -> 714.0s] 总共要形成一个标准, [714.0s -> 715.0s] 然后规模化, [715.0s -> 716.0s] 那你这个标准, [716.0s -> 717.0s] 可靠的, [717.0s -> 718.0s] 为什么叫可靠, [718.0s -> 719.0s] 在某一个地方以后, [719.0s -> 720.0s] 他跑了一天, [720.0s -> 721.0s] 他第二天, [721.0s -> 722.0s] 他是不是还愿意开机, [722.0s -> 723.0s] 如果他自己, [723.0s -> 724.0s] 主动都愿意开机, [724.0s -> 725.0s] 他肯定付费愿是强的, [725.0s -> 727.0s] 那他愿意付出的, [727.0s -> 728.0s] 就可以覆盖这个成本, [728.0s -> 730.0s] 如果他不愿意开机, [730.0s -> 731.0s] 你说你不来, [731.0s -> 732.0s] 你们工程师不来, [732.0s -> 733.0s] 你就不开机, [733.0s -> 734.0s] 那肯定这个成本, [734.0s -> 735.0s] 是高的, [735.0s -> 736.0s] 这个, [736.0s -> 737.0s] 确实还是要, [737.0s -> 738.0s] 筛选一下, [738.0s -> 739.0s] 是一个动态的过程, [739.0s -> 740.0s] 你这个也涉及, [740.0s -> 741.0s] 我们过去, [741.0s -> 742.0s] 行业, [742.0s -> 743.0s] 传递出来的, [743.0s -> 744.0s] 信心, [744.0s -> 745.0s] 或者预期, [745.0s -> 746.0s] 就造成市场的预期, [746.0s -> 748.0s] 有点是偏高的, [748.0s -> 749.0s] 因为如果不偏高, [749.0s -> 750.0s] 肯定是百分之百去, [750.0s -> 751.0s] 我们绝对都可以满足, [751.0s -> 752.0s] 但是, [752.0s -> 753.0s] 就是过去几天停下来, [753.0s -> 754.0s] 有一些预期, [754.0s -> 755.0s] 确实是比较高的, [755.0s -> 757.0s] 经常会举一个很, [757.0s -> 758.0s] 复杂的一个, [758.0s -> 759.0s] 生产流程, [759.0s -> 761.0s] 一个场景流程, [761.0s -> 762.0s] 说, [762.0s -> 763.0s] 这个感觉,

[763.0s -> 764.0s] 你们应该很容易做呀, [764.0s -> 765.0s] 就是他们那儿都能, [765.0s -> 766.0s] 做什么什么东西了, [766.0s -> 767.0s] 是吧, [767.0s -> 768.0s] 这样的这个预期, [768.0s -> 769.0s] 还是过高, [769.0s -> 770.0s] 但是, [770.0s -> 771.0s] 我觉得, [771.0s -> 772.0s] 也是一个好事, [772.0s -> 773.0s] 这也造成了, [773.0s -> 774.0s] 我们有很多这样的, [774.0s -> 775.0s] 需求, [775.0s -> 776.0s] 可以从一个更大的, [776.0s -> 777.0s] 池子里面, [777.0s -> 778.0s] 来去筛学。 [778.0s -> 779.0s] OK, [779.0s -> 780.0s] 对,那看起来, [780.0s -> 781.0s] 再往下这个, [781.0s -> 782.0s] 机器人的国际化, [782.0s -> 783.0s] 确实有可能是, [783.0s -> 785.0s] 中国在新出口当中的, [785.0s -> 786.0s] 一个潜在增相了, [786.0s -> 787.0s] 虽然现在还很小, [787.0s -> 788.0s] 那这个话题, [788.0s -> 789.0s] 我们先放一边, [789.0s -> 790.0s] 还有另外一个, [790.0s -> 791.0s] 今天比较热点, [791.0s -> 792.0s] 和时髦的话题, [792.0s -> 793.0s] 就是, [793.0s -> 794.0s] 在两年以前, [794.0s -> 795.0s] 我们录上一期的时候, [795.0s -> 796.0s] 那个时候你就提了一个, [796.0s -> 798.0s] 当时其实不太受关注, [798.0s -> 800.0s] 当时在创业方向上, [800.0s -> 802.0s] 也没有很多人做的事情, [802.0s -> 804.0s] 那时候你的词汇叫做, [804.0s -> 805.0s] 怎么解决直觉物理的问题, [805.0s -> 807.0s] 就是人有个直觉物理, [807.0s -> 808.0s] 就是说白了, [808.0s -> 810.0s] 你知道怎么拿东西, [810.0s -> 811.0s] 这个东西, [811.0s -> 812.0s] 用什么姿态, [812.0s -> 813.0s] 或者用什么力道, [813.0s -> 814.0s] 或者用什么状况来, [814.0s -> 815.0s] 这个执行, [815.0s -> 816.0s] 或者, [816.0s -> 817.0s] 来对这个物体做动作, [817.0s -> 818.0s] 那, [818.0s -> 819.0s] 当然也包括一些, [819.0s -> 820.0s] 常识性的物理, [820.0s -> 821.0s] 那这个, [821.0s -> 822.0s] 那时候你把它称之为, [822.0s -> 823.0s] 直觉物理, [823.0s -> 824.0s] 说这个是当时, [824.0s -> 825.0s] 你们在国外的时候, [825.0s -> 826.0s] 就一直在考虑, [826.0s -> 827.0s] 怎么解决这个问题, [827.0s -> 828.0s] 那我提这个话题的, [828.0s -> 829.0s] 原因是, [829.0s -> 830.0s] 从圈年里, [830.0s -> 831.0s] 年间初开始, [831.0s -> 832.0s] 那个机器人相关的, [832.0s -> 833.0s] 创业方向, [833.0s -> 834.0s] 教世界模型, [834.0s -> 835.0s] 或物理模型, [835.0s -> 836.0s] 反正这个, [836.0s -> 837.0s] 定义有非常多, [837.0s -> 838.0s] 总而言之, [838.0s -> 839.0s] 就是解决这个交互, [839.0s -> 840.0s] 物理状态改变, [840.0s -> 841.0s] 怎么让机器人, [841.0s -> 842.0s] 理解这种, [842.0s -> 843.0s] 带物理量的世界, [843.0s -> 844.0s] 这些事情, [844.0s -> 845.0s] 我们也投了几件, [845.0s -> 846.0s] 这样的初创公司, [846.0s -> 847.0s] 当然他们确实也是, [847.0s -> 848.0s] 很热, [848.0s -> 849.0s] 因为他们的估值, [849.0s -> 850.0s] 都上涨得非常快, [850.0s -> 851.0s] 在过去的六个月, [851.0s -> 852.0s] 好, [852.0s -> 853.0s] 我的问题是, [853.0s -> 854.0s] 考虑到, [854.0s -> 855.0s] 你两年以前, [855.0s -> 856.0s] 就一直在思考, [856.0s -> 857.0s] 这样的问题, [857.0s -> 858.0s] 你觉得, [858.0s -> 859.0s] 今天, [859.0s -> 860.0s] 世界模型, [860.0s -> 861.0s] 和世界模型, [861.0s -> 862.0s] 因为中国非常热, [862.0s -> 863.0s] 美国也算比较热, [863.0s -> 864.0s] 今天, [864.0s -> 865.0s] 这个事情的出现, [865.0s -> 867.0s] 这个事情的发展阶段, [867.0s -> 868.0s] 和这个事情的解决方案, [868.0s -> 869.0s] 大概是个什么样子, [869.0s -> 870.0s] 过去很多年, [870.0s -> 871.0s] 思考这个问题, [871.0s -> 872.0s] 相比较而言。 [872.0s -> 873.0s] 这是, [873.0s -> 874.0s] 挺有意思的一个问题, [874.0s -> 875.0s] 当时, [875.0s -> 876.0s] 确实提了, [876.0s -> 877.0s] 这个, [877.0s -> 878.0s] 直觉物理, [878.0s -> 879.0s] 然后怎么去, [879.0s -> 880.0s] 理解世界, [880.0s -> 881.0s] 怎么去跟世界交互, [881.0s -> 882.0s] 然后后来, [882.0s -> 883.0s] 发生了, [883.0s -> 884.0s] 像分数总结的, [884.0s -> 885.0s] 各种各样的路线, [885.0s -> 886.0s] 都在致力于解决, [886.0s -> 887.0s] 这样的问题, [887.0s -> 888.0s] 因为机器人, [888.0s -> 889.0s] 需要做的事情, [889.0s -> 890.0s] 就是, [890.0s -> 891.0s] 我要知道,

[891.0s -> 892.0s] 去动到什么地方, [892.0s -> 893.0s] 第二个点, [893.0s -> 894.0s] 就是, [894.0s -> 895.0s] 怎么, [895.0s -> 896.0s] 安全可靠的, [896.0s -> 897.0s] 动到脑, [897.0s -> 898.0s] 其实, [898.0s -> 899.0s] 本质上, [899.0s -> 900.0s] 就是在解决这两个问题, [900.0s -> 901.0s] 抽离出来, [901.0s -> 902.0s] 一个, [902.0s -> 903.0s] 我们可以把, [903.0s -> 904.0s] 运动大体的, [904.0s -> 905.0s] 分成两类, [905.0s -> 906.0s] 一类, [906.0s -> 907.0s] 是这种, [907.0s -> 908.0s] 自由空间, [908.0s -> 909.0s] 我要从, [909.0s -> 910.0s] 任何的, [910.0s -> 911.0s] A点到B点, [911.0s -> 912.0s] 以及呢, [912.0s -> 913.0s] 如果有, [913.0s -> 914.0s] 接触, [914.0s -> 915.0s] 有交互, [915.0s -> 916.0s] 我怎么, [916.0s -> 917.0s] 接触这东西, [917.0s -> 918.0s] A点到B点, [918.0s -> 919.0s] 这个A点到B点, [919.0s -> 920.0s] 不光是空间上, [920.0s -> 921.0s] 也是状态上, [921.0s -> 922.0s] 另外一个问题, [922.0s -> 923.0s] 就是, [923.0s -> 924.0s] 如何去决定, [924.0s -> 925.0s] 我的B点在哪? [925.0s -> 926.0s] 是, [926.0s -> 927.0s] 目标是什么? [927.0s -> 928.0s] 所以, [928.0s -> 929.0s] 之前, [929.0s -> 930.0s] 行业, [930.0s -> 931.0s] 包括, [931.0s -> 932.0s] 学界也好, [932.0s -> 933.0s] 还是, [933.0s -> 934.0s] 工业界也好, [934.0s -> 935.0s] 很多在推的, [935.0s -> 936.0s] 不管是, [936.0s -> 937.0s] 世界圆动的模型, [937.0s -> 938.0s] 还是, [938.0s -> 939.0s] 这个世界模型, [939.0s -> 940.0s] 想去, [940.0s -> 941.0s] 理解这个东西, [941.0s -> 943.0s] 想去理解这个常识, [943.0s -> 944.0s] 现在的方式, [944.0s -> 945.0s] 是, [945.0s -> 946.0s] 通过数据采集, [946.0s -> 947.0s] 更多的是, [947.0s -> 948.0s] 说, [948.0s -> 949.0s] 我采完这个数据, [949.0s -> 950.0s] 以后, [950.0s -> 951.0s] 我要去做一个, [951.0s -> 952.0s] 回归, [952.0s -> 953.0s] 我从, [953.0s -> 954.0s] 我观测到的量, [954.0s -> 955.0s] 直接去决定, [955.0s -> 956.0s] 我被来往哪儿动, [956.0s -> 957.0s] 它, [957.0s -> 958.0s] 其实是一个, [958.0s -> 959.0s] 结尽, [959.0s -> 960.0s] 绕过了理解, [960.0s -> 961.0s] 直接去生成动作, [961.0s -> 962.0s] 但是, [962.0s -> 963.0s] 后来, [963.0s -> 964.0s] 大家可能发现, [964.0s -> 965.0s] A这个好像, [965.0s -> 966.0s] 有点, [966.0s -> 967.0s] 对, [967.0s -> 968.0s] 有点问题, [968.0s -> 969.0s] 这个是过你和的, [969.0s -> 970.0s] 对, [970.0s -> 971.0s] 是点, [971.0s -> 972.0s] 确实, [972.0s -> 973.0s] 但是, [973.0s -> 974.0s] 它的繁华性, [974.0s -> 975.0s] 没有大家想象的好, [975.0s -> 976.0s] 这个东西, [976.0s -> 977.0s] 从一个1厘米, [977.0s -> 978.0s] 往左移, [978.0s -> 979.0s] 0.5厘米, [979.0s -> 980.0s] 还能抓, [980.0s -> 981.0s] 就叫繁华了, [981.0s -> 982.0s] 我就是, [982.0s -> 983.0s] 从来不觉得, [983.0s -> 984.0s] 这个叫繁华, [984.0s -> 985.0s] 就是, [985.0s -> 986.0s] 这是很自然的, [986.0s -> 987.0s] 应该有的能力, [987.0s -> 988.0s] 所以, [988.0s -> 989.0s] 反而, [989.0s -> 990.0s] 大家往这个, [990.0s -> 991.0s] 世界圆动的模型, [991.0s -> 992.0s] 去走的时候, [992.0s -> 993.0s] 反而把这个, [993.0s -> 994.0s] 标准降的, [994.0s -> 995.0s] 低了一点, [995.0s -> 996.0s] 对, [996.0s -> 997.0s] 然后, [997.0s -> 998.0s] 那, [998.0s -> 999.0s] 慢慢的, [999.0s -> 1000.0s] 大家意识到的这个问题, [1000.0s -> 1001.0s] 又开始, [1001.0s -> 1002.0s] 做这个世界模型, [1002.0s -> 1003.0s] 世界模型, [1003.0s -> 1004.0s] 其实是, [1004.0s -> 1005.0s] 模型, [1005.0s -> 1006.0s] 再去, [1006.0s -> 1007.0s] 提特征, [1007.0s -> 1008.0s] 然后用, [1008.0s -> 1009.0s] 用这些特征, [1009.0s -> 1010.0s] 去做决策的时候, [1010.0s -> 1011.0s] 不希望, [1011.0s -> 1012.0s] 去丢掉,

[1012.0s -> 1013.0s] 那些, [1013.0s -> 1014.0s] 理解世界, [1014.0s -> 1015.0s] 那些信息, [1015.0s -> 1016.0s] 对, [1016.0s -> 1017.0s] 不光要生生动作, [1017.0s -> 1018.0s] 我说, [1018.0s -> 1019.0s] 理解世界, [1019.0s -> 1020.0s] 理解, [1020.0s -> 1021.0s] 客观, [1021.0s -> 1022.0s] 物体的, [1022.0s -> 1023.0s] 发展规律, [1023.0s -> 1024.0s] 但是, [1024.0s -> 1025.0s] 过去呢, [1025.0s -> 1026.0s] 这些世界模型, [1026.0s -> 1027.0s] 往往还是, [1027.0s -> 1028.0s] 一个pixel to pixel, [1028.0s -> 1029.0s] 对, [1029.0s -> 1030.0s] 从, [1030.0s -> 1031.0s] 像素领到像素领, [1031.0s -> 1032.0s] 对, [1032.0s -> 1033.0s] 一个像素, [1033.0s -> 1034.0s] 把大量的, [1034.0s -> 1035.0s] 这个, [1035.0s -> 1036.0s] 表达能力, [1036.0s -> 1037.0s] 都用于去重建了, [1037.0s -> 1038.0s] 去, [1038.0s -> 1039.0s] 去恢复这个, [1039.0s -> 1040.0s] 也有一些, [1040.0s -> 1041.0s] 学术界, [1041.0s -> 1042.0s] 会提些方法, [1042.0s -> 1043.0s] 比如说, [1043.0s -> 1044.0s] 我是不是不用关心, [1044.0s -> 1045.0s] 所有的, [1045.0s -> 1046.0s] 那是不是, [1046.0s -> 1047.0s] 能够去, [1047.0s -> 1048.0s] 压缩, [1048.0s -> 1049.0s] 我只去预测一些, [1049.0s -> 1050.0s] 物体的状态, [1050.0s -> 1051.0s] 或者去做一些分类, [1051.0s -> 1052.0s] 分割, [1052.0s -> 1053.0s] 可能会加一些职任物, [1053.0s -> 1054.0s] 然后, [1054.0s -> 1055.0s] 我们其实很早, [1055.0s -> 1056.0s] 但是一直在, [1056.0s -> 1057.0s] 渐行着, [1057.0s -> 1058.0s] 以物体为中心的, [1058.0s -> 1059.0s] 一个, [1059.0s -> 1060.0s] 思路去走, [1060.0s -> 1061.0s] 不用去过度的, [1061.0s -> 1062.0s] 去建模, [1062.0s -> 1063.0s] 视觉世界模型里面, [1063.0s -> 1064.0s] 没有的, [1064.0s -> 1065.0s] 反而, [1065.0s -> 1066.0s] 我们是要关心的, [1066.0s -> 1067.0s] 比如说, [1067.0s -> 1068.0s] 力的模态, [1068.0s -> 1069.0s] 就是一个很关键的模态, [1069.0s -> 1070.0s] 所以, [1070.0s -> 1071.0s] 当我们在做, [1071.0s -> 1072.0s] 预测能力模型的时候, [1072.0s -> 1073.0s] 我们姑且, [1073.0s -> 1074.0s] 也称之为, [1074.0s -> 1075.0s] 是这个, [1075.0s -> 1076.0s] 世界动作模型, [1076.0s -> 1077.0s] 当我们做, [1077.0s -> 1078.0s] 这个模型的时候呢, [1078.0s -> 1079.0s] 我们一方面, [1079.0s -> 1080.0s] 会去建模, [1080.0s -> 1081.0s] 这个原始的, [1081.0s -> 1082.0s] 这个图像, [1082.0s -> 1083.0s] 以及我的, [1083.0s -> 1084.0s] 动作, [1084.0s -> 1085.0s] 进来以后, [1085.0s -> 1086.0s] 我希望能够预测未来, [1086.0s -> 1087.0s] 我在预测未来, [1087.0s -> 1088.0s] 图像的同时, [1088.0s -> 1089.0s] 我也希望预测未来的力, [1089.0s -> 1090.0s] 因为, [1090.0s -> 1091.0s] 力的这个模态, [1092.0s -> 1094.0s] 眼睛可能换个角度, [1094.0s -> 1095.0s] 你得到的信息, [1095.0s -> 1096.0s] 是跟你的, [1096.0s -> 1097.0s] 本体, [1097.0s -> 1099.0s] 跟人是没有那么大关系的, [1099.0s -> 1100.0s] 最多是, [1100.0s -> 1101.0s] 角度有一个差异, [1101.0s -> 1102.0s] 但是力, [1102.0s -> 1103.0s] 力这个东西, [1103.0s -> 1104.0s] 它是, [1104.0s -> 1105.0s] 跟你的本体, [1105.0s -> 1106.0s] 施加到环境中的, [1106.0s -> 1107.0s] 完全决定于你, [1107.0s -> 1108.0s] 你自己的, [1108.0s -> 1109.0s] 就是说, [1109.0s -> 1110.0s] 你施加力大, [1110.0s -> 1111.0s] 你反馈的力就大, [1111.0s -> 1112.0s] 施加的小, [1112.0s -> 1113.0s] 就反馈的小, [1113.0s -> 1114.0s] 所以, [1114.0s -> 1115.0s] 它是一个, [1115.0s -> 1116.0s] 跟动作, [1116.0s -> 1117.0s] 极其绑定偶和的东西, [1117.0s -> 1118.0s] 所以, [1118.0s -> 1119.0s] 我们做的模型, [1119.0s -> 1120.0s] 一方面在, [1120.0s -> 1121.0s] 预测未来的图像, [1121.0s -> 1122.0s] 会把这个, [1122.0s -> 1123.0s] 预测的结果, [1123.0s -> 1124.0s] 会跟实际的, [1124.0s -> 1125.0s] 施加的一个动作, [1125.0s -> 1126.0s] 实测出来一个力, [1126.0s -> 1127.0s] 又会去比较, [1127.0s -> 1128.0s] 去做避缓, [1128.0s -> 1130.0s] 它不能只单做一个前馈, [1130.0s -> 1131.0s] 而说, [1131.0s -> 1132.0s] 当我预测出来, [1132.0s -> 1134.0s] 感受到多大的力之后, [1134.0s -> 1135.0s] 我也预测我的一个动作, [1135.0s -> 1137.0s] 那我施加我预测的动作, [1137.0s -> 1138.0s] 到环境中,

[1138.0s -> 1139.0s] 又会感觉到一个新的力, [1139.0s -> 1140.0s] 那拿这个新的力, [1140.0s -> 1142.0s] 跟我预测的力去比较, [1142.0s -> 1143.0s] 来反补, [1143.0s -> 1145.0s] 来纠正我的预测能力, [1145.0s -> 1147.0s] 所以我们形成这么一个避缓, [1147.0s -> 1148.0s] 就可以把这个, [1148.0s -> 1149.0s] 比解, [1149.0s -> 1150.0s] 直觉的一些物体,物体, [1150.0s -> 1151.0s] 交互规律, [1151.0s -> 1152.0s] 或者人, [1152.0s -> 1153.0s] 机器人, [1153.0s -> 1154.0s] 跟物体的交互规律, [1154.0s -> 1155.0s] 就能够, [1155.0s -> 1156.0s] 逐渐的随着数据增多, [1156.0s -> 1157.0s] 能够把这个, [1157.0s -> 1158.0s] 估计误差, [1158.0s -> 1159.0s] 把他的理解能力, [1159.0s -> 1161.0s] 就是提的越来越好。 [1161.0s -> 1162.0s] 因为我们也看, [1162.0s -> 1163.0s] 我们也投, [1163.0s -> 1164.0s] 你们也在实践, [1164.0s -> 1165.0s] 而且大家今天, [1165.0s -> 1167.0s] 各个机器人企业, [1167.0s -> 1168.0s] 多多少少, [1168.0s -> 1169.0s] 在这个行业里都来到了, [1169.0s -> 1170.0s] 要解决, [1170.0s -> 1172.0s] 跟整个环境, [1172.0s -> 1173.0s] 和状态相关的, [1173.0s -> 1175.0s] 这些操作和改变的问题, [1175.0s -> 1176.0s] 今天这些方向, [1176.0s -> 1177.0s] 就像我们刚才讲的, [1177.0s -> 1178.0s] 肯定是没有收敛的, [1178.0s -> 1179.0s] 就像你刚才讲, [1179.0s -> 1180.0s] 你已经涉及到了, [1180.0s -> 1181.0s] 这个, [1181.0s -> 1182.0s] 模式迁移过程当中的, [1182.0s -> 1184.0s] 或者名词的弱干都变化了, [1184.0s -> 1185.0s] VLA, [1185.0s -> 1186.0s] VLM, [1186.0s -> 1187.0s] 就word action, [1187.0s -> 1188.0s] 等等等等, [1188.0s -> 1189.0s] 还有WM, [1189.0s -> 1190.0s] 那, [1190.0s -> 1191.0s] 我们先不管, [1191.0s -> 1192.0s] 这些所有的名词, [1192.0s -> 1193.0s] 也不管他们没有收敛, [1193.0s -> 1195.0s] 大家都在积极探索, [1195.0s -> 1196.0s] 你觉得, [1196.0s -> 1197.0s] 从你的角度, [1197.0s -> 1198.0s] 当然这肯定只代表, [1198.0s -> 1199.0s] 个人观点了, [1199.0s -> 1200.0s] 最终大家在, [1200.0s -> 1201.0s] 接下来的一步, [1201.0s -> 1202.0s] 或两步, [1202.0s -> 1203.0s] 或三步, [1203.0s -> 1204.0s] 在探索, [1204.0s -> 1205.0s] 怎么能够, [1205.0s -> 1206.0s] 建模世界, [1206.0s -> 1207.0s] 或者理解世界的, [1207.0s -> 1208.0s] 这个过程当中, [1208.0s -> 1210.0s] 大家可能会, [1210.0s -> 1211.0s] 向哪, [1211.0s -> 1212.0s] 一两三个方向, [1212.0s -> 1213.0s] 进行收敛, [1213.0s -> 1214.0s] 我觉得这个呢, [1214.0s -> 1215.0s] 是需要从, [1215.0s -> 1216.0s] 最终, [1216.0s -> 1217.0s] 想要达到的, [1217.0s -> 1218.0s] 来倒推, [1218.0s -> 1219.0s] 最终, [1219.0s -> 1220.0s] 我们肯定是想, [1220.0s -> 1221.0s] 达到, [1221.0s -> 1222.0s] 灵样本, [1222.0s -> 1223.0s] 小样本, [1223.0s -> 1224.0s] 的方法, [1224.0s -> 1225.0s] 扔到一个新环境, [1225.0s -> 1226.0s] 这, [1226.0s -> 1227.0s] 一次展示, [1227.0s -> 1228.0s] 甚至灵展示, [1228.0s -> 1229.0s] 它就可以, [1229.0s -> 1230.0s] 动起来, [1230.0s -> 1231.0s] 它就可以, [1231.0s -> 1232.0s] 理解任务去做, [1232.0s -> 1233.0s] 第二点呢, [1233.0s -> 1234.0s] 它需要是, [1234.0s -> 1235.0s] 能够去, [1235.0s -> 1236.0s] 绝对的, [1236.0s -> 1237.0s] 安全, [1237.0s -> 1238.0s] 知道自己的边界, [1238.0s -> 1239.0s] 如果有异常, [1239.0s -> 1240.0s] 它是能够处理, [1240.0s -> 1241.0s] 这个肯定是, [1241.0s -> 1242.0s] 我们最终, [1242.0s -> 1243.0s] 希望达到的, [1243.0s -> 1244.0s] 就是, [1244.0s -> 1245.0s] 可翻发, [1245.0s -> 1246.0s] 且可靠, [1246.0s -> 1247.0s] 那这个倒推回来, [1247.0s -> 1248.0s] 那肯定是, [1248.0s -> 1249.0s] 希望, [1249.0s -> 1250.0s] 数据最高效的, [1250.0s -> 1251.0s] 一个方式, [1251.0s -> 1252.0s] 才能到, [1252.0s -> 1253.0s] 这种, [1253.0s -> 1254.0s] 程度, [1254.0s -> 1255.0s] 另外一个呢, [1255.0s -> 1256.0s] 它是需要, [1256.0s -> 1257.0s] 能够保证, [1257.0s -> 1258.0s] 自主的, [1258.0s -> 1259.0s] 安全性, [1259.0s -> 1260.0s] 所以需要, [1260.0s -> 1261.0s] 能够, [1261.0s -> 1262.0s] 有一个方式, [1262.0s -> 1263.0s] 达出的, [1263.0s -> 1264.0s] 不确定性, [1264.0s -> 1265.0s] 边界, [1265.0s -> 1266.0s] 所以, [1266.0s -> 1267.0s] 倒推回来, [1267.0s -> 1268.0s] 应该是,

[1268.0s -> 1269.0s] 沿着这两个路线走, [1269.0s -> 1270.0s] 按下来, [1270.0s -> 1271.0s] 我们正向的, [1271.0s -> 1272.0s] 这个世界模型, [1272.0s -> 1273.0s] 再往前, [1273.0s -> 1274.0s] 去, [1274.0s -> 1275.0s] 发展的路径上, [1275.0s -> 1276.0s] 大家也, [1276.0s -> 1277.0s] 逐渐的去, [1277.0s -> 1278.0s] 对世界模型, [1278.0s -> 1279.0s] 去做一些, [1279.0s -> 1280.0s] 各种各样的, [1280.0s -> 1281.0s] 方便的, [1281.0s -> 1282.0s] 尝试, [1282.0s -> 1283.0s] 去提高, [1283.0s -> 1284.0s] 数据效率, [1284.0s -> 1285.0s] 比如说, [1285.0s -> 1286.0s] 有去做这个, [1286.0s -> 1287.0s] 让数据采集的更便宜, [1287.0s -> 1288.0s] 对啊, [1288.0s -> 1289.0s] 一个, [1289.0s -> 1290.0s] Ego Century, [1290.0s -> 1291.0s] 对, [1291.0s -> 1292.0s] 出现的时候, [1292.0s -> 1293.0s] 其实, [1293.0s -> 1294.0s] 就是, [1294.0s -> 1295.0s] 已经在采集, [1295.0s -> 1296.0s] 大量的, [1296.0s -> 1297.0s] Ego Century, [1297.0s -> 1298.0s] 在这个, [1298.0s -> 1299.0s] 在这个, [1299.0s -> 1300.0s] 在这个, [1300.0s -> 1301.0s] 在这个, [1301.0s -> 1302.0s] 就是, [1302.0s -> 1303.0s] 这个方面, [1303.0s -> 1304.0s] 肯定是一个尝试, [1304.0s -> 1305.0s] 如何能够, [1305.0s -> 1306.0s] 把数据采集, [1306.0s -> 1307.0s] 变得更便宜, [1307.0s -> 1308.0s] 对, [1308.0s -> 1309.0s] 第二个呢, [1309.0s -> 1310.0s] 是如何能, [1310.0s -> 1311.0s] 让这些数据, [1311.0s -> 1312.0s] 榨取出来的信息, [1312.0s -> 1313.0s] 更, [1313.0s -> 1314.0s] 密度更高, [1314.0s -> 1315.0s] 更, [1315.0s -> 1316.0s] 能够去, [1316.0s -> 1317.0s] 贴近, [1317.0s -> 1318.0s] 直觉物理, [1318.0s -> 1319.0s] 比如说, [1319.0s -> 1320.0s] 现在, [1320.0s -> 1321.0s] 从, [1321.0s -> 1322.0s] 二地, [1322.0s -> 1323.0s] 或者多视角的二地中, [1323.0s -> 1324.0s] 隐释的推出来3D, [1324.0s -> 1325.0s] 因为3D, [1325.0s -> 1326.0s] 是一个更, [1326.0s -> 1327.0s] 本质的, [1327.0s -> 1328.0s] 对, [1328.0s -> 1329.0s] 环境的一个, [1329.0s -> 1330.0s] 所以像, [1330.0s -> 1331.0s] 这些方法, [1331.0s -> 1332.0s] 大家也都, [1332.0s -> 1333.0s] 逐渐在尝试, [1333.0s -> 1334.0s] 然后另外一个, [1334.0s -> 1335.0s] 唯独就是, [1335.0s -> 1336.0s] 刚才说的, [1336.0s -> 1337.0s] 如何保证安全这一点, [1337.0s -> 1338.0s] 就是过去呢, [1338.0s -> 1339.0s] 我们采的这些数据, [1339.0s -> 1340.0s] 通常都是, [1340.0s -> 1341.0s] 视觉为主, [1341.0s -> 1342.0s] 甚至只有视觉, [1342.0s -> 1343.0s] 然后, [1343.0s -> 1344.0s] 现在行业里面, [1344.0s -> 1345.0s] 也在采集, [1345.0s -> 1346.0s] 立觉, [1346.0s -> 1347.0s] 甚至触觉, [1347.0s -> 1348.0s] 对, [1348.0s -> 1349.0s] 也有做触觉手套, [1349.0s -> 1350.0s] 传感器的, [1350.0s -> 1351.0s] 这样的数据, [1351.0s -> 1352.0s] 越来越多, [1352.0s -> 1353.0s] 我们自己呢, [1353.0s -> 1354.0s] 自己在探索, [1354.0s -> 1355.0s] 怎么把视力, [1355.0s -> 1357.0s] 触觉都融合起来, [1357.0s -> 1358.0s] 也做了自己的, [1358.0s -> 1359.0s] 处于采集, [1359.0s -> 1360.0s] 连络, [1360.0s -> 1361.0s] 这个也是一个, [1361.0s -> 1362.0s] 比较显著的, [1362.0s -> 1363.0s] 一个趋势, [1363.0s -> 1364.0s] 那这个话题, [1364.0s -> 1365.0s] 我们先放一放, [1365.0s -> 1366.0s] 回到, [1366.0s -> 1367.0s] 原络跟中国的, [1367.0s -> 1368.0s] 机器人相关的话题, [1368.0s -> 1369.0s] 就是你觉得, [1369.0s -> 1370.0s] 在过去的这两年, [1370.0s -> 1371.0s] 或者三年的, [1371.0s -> 1372.0s] 创业里面, [1372.0s -> 1373.0s] 从, [1373.0s -> 1374.0s] 发展过程来看, [1374.0s -> 1375.0s] 在整个中国的, [1375.0s -> 1376.0s] 机器人行业里边, [1376.0s -> 1377.0s] 软的, [1377.0s -> 1378.0s] 我们讲软的, [1378.0s -> 1379.0s] 和硬的, [1379.0s -> 1380.0s] 硬的就是, [1380.0s -> 1381.0s] 这些具体的, [1381.0s -> 1382.0s] 硬件, [1382.0s -> 1383.0s] 就不管它是电机, [1383.0s -> 1384.0s] 关节, [1384.0s -> 1385.0s] 等等这些事情, [1385.0s -> 1386.0s] 控制, [1386.0s -> 1387.0s] 就是包括手啊, [1387.0s -> 1388.0s] 所有这些, [1388.0s -> 1389.0s] 软的和硬的,

[1389.0s -> 1390.0s] 看起来, [1390.0s -> 1391.0s] 比三年以前, [1391.0s -> 1392.0s] 就各自进步的, [1392.0s -> 1393.0s] 状态和, [1393.0s -> 1394.0s] 进展的速度, [1394.0s -> 1395.0s] 和进展的规模, [1395.0s -> 1396.0s] 是啥样? [1396.0s -> 1397.0s] 我是觉得, [1397.0s -> 1398.0s] 硬件上, [1398.0s -> 1399.0s] 看起来, [1399.0s -> 1400.0s] 发展是, [1400.0s -> 1401.0s] 没有那么快, [1401.0s -> 1402.0s] 但是实质上, [1402.0s -> 1403.0s] 工程上, [1403.0s -> 1404.0s] 取得的进展, [1404.0s -> 1405.0s] 是很快, [1405.0s -> 1406.0s] 软件上, [1406.0s -> 1407.0s] 大家直觉上, [1407.0s -> 1408.0s] 这个, [1408.0s -> 1409.0s] 演示的, [1409.0s -> 1410.0s] 漂亮程度, [1410.0s -> 1411.0s] 是, [1411.0s -> 1412.0s] 很高很高, [1412.0s -> 1413.0s] 变得特别特别快, [1413.0s -> 1414.0s] 但是, [1414.0s -> 1415.0s] 拉长时间纬度, [1415.0s -> 1416.0s] 来看, [1416.0s -> 1417.0s] 可能, [1417.0s -> 1418.0s] 进展没有想象, [1418.0s -> 1419.0s] 那么快, [1419.0s -> 1420.0s] 就是硬件上呢, [1420.0s -> 1421.0s] 很多人会说, [1421.0s -> 1422.0s] 三年前, [1422.0s -> 1423.0s] NRC, [1423.0s -> 1424.0s] 展示各种, [1424.0s -> 1425.0s] 这个, [1425.0s -> 1426.0s] 协作币, [1426.0s -> 1427.0s] 也有一些人生机器人, [1427.0s -> 1428.0s] 那现在还是这些, [1428.0s -> 1429.0s] 然后, [1429.0s -> 1430.0s] 用的关节, [1430.0s -> 1431.0s] 可能还是, [1431.0s -> 1432.0s] 写播, [1432.0s -> 1433.0s] 行星, [1433.0s -> 1434.0s] 然后现在, [1434.0s -> 1435.0s] 也有一些, [1435.0s -> 1436.0s] 想, [1436.0s -> 1437.0s] 包括设计上, [1437.0s -> 1438.0s] 包括系统机程上, [1438.0s -> 1439.0s] 实质的发展, [1439.0s -> 1440.0s] 还是比想象, [1440.0s -> 1441.0s] 快很多的, [1441.0s -> 1442.0s] 成本上, [1442.0s -> 1443.0s] 就自不必说了, [1443.0s -> 1444.0s] 然后, [1444.0s -> 1445.0s] 另外是, [1445.0s -> 1446.0s] 它的可耗性上, [1446.0s -> 1448.0s] 像新一代的机器人, [1448.0s -> 1449.0s] 这么一两年, [1449.0s -> 1450.0s] 已经慢慢能做到, [1450.0s -> 1451.0s] 像我们自己做, [1451.0s -> 1452.0s] 已经重新定义, [1452.0s -> 1453.0s] 绝对定义, [1453.0s -> 1454.0s] 已经能逼近, [1454.0s -> 1455.0s] 以往, [1455.0s -> 1456.0s] 工业机器人的级别了, [1456.0s -> 1457.0s] 就是这个, [1457.0s -> 1458.0s] 在两年前, [1458.0s -> 1459.0s] 不是那么敢想象的, [1459.0s -> 1460.0s] 所以, [1460.0s -> 1461.0s] 硬件的发展程度, [1461.0s -> 1462.0s] 我们国内, [1462.0s -> 1463.0s] 确实有这个优势, [1463.0s -> 1464.0s] 它的硬件迭代程度, [1464.0s -> 1465.0s] 甚至, [1465.0s -> 1466.0s] 已经到了, [1466.0s -> 1467.0s] 软件迭代速度, [1467.0s -> 1468.0s] 能够逼近到这种, [1468.0s -> 1470.0s] 发展的有曲线, [1470.0s -> 1471.0s] 而且, [1471.0s -> 1472.0s] 这些硬件的, [1472.0s -> 1473.0s] 也是得益于, [1473.0s -> 1475.0s] 国家在大量的投入, [1475.0s -> 1476.0s] 得益于, [1476.0s -> 1477.0s] 我们百花齐放, [1477.0s -> 1478.0s] 这么多企业, [1478.0s -> 1479.0s] 包括, [1479.0s -> 1480.0s] 高校也都在投入研究, [1480.0s -> 1481.0s] 所以, [1481.0s -> 1482.0s] 造成, [1482.0s -> 1483.0s] 上游一些, [1483.0s -> 1484.0s] 不管是, [1484.0s -> 1485.0s] 极加工, [1485.0s -> 1486.0s] 还是电磁的, [1486.0s -> 1487.0s] 传感器的, [1487.0s -> 1488.0s] 这方面, [1488.0s -> 1489.0s] 确实, [1489.0s -> 1490.0s] 上游很成熟, [1490.0s -> 1491.0s] 愿意去做这行, [1491.0s -> 1492.0s] 那人也越来越多, [1492.0s -> 1493.0s] 整个生态, [1493.0s -> 1494.0s] 做得越来越好了, [1494.0s -> 1495.0s] 然后, [1495.0s -> 1496.0s] 整个协议呢, [1496.0s -> 1497.0s] 整个标准, [1497.0s -> 1498.0s] 也逐渐在形成, [1498.0s -> 1499.0s] 这个, [1499.0s -> 1500.0s] 是比大家体感的, [1500.0s -> 1501.0s] 会快一些, [1501.0s -> 1502.0s] 然后, [1502.0s -> 1503.0s] 模型, [1503.0s -> 1505.0s] 上海这一侧呢, [1505.0s -> 1506.0s] 我们确实是, [1506.0s -> 1507.0s] 看到了, [1507.0s -> 1508.0s] 很多, [1508.0s -> 1509.0s] 新的成果, [1509.0s -> 1510.0s] 不管是, [1510.0s -> 1511.0s] 基于, [1511.0s -> 1512.0s] 纯, [1512.0s -> 1513.0s] 端道端像,

[1513.0s -> 1514.0s] VLA这种的, [1514.0s -> 1515.0s] 还是, [1515.0s -> 1516.0s] 基于世界模型这样, [1516.0s -> 1517.0s] 可能, [1517.0s -> 1518.0s] 学一个预测模型, [1518.0s -> 1519.0s] 然后, [1519.0s -> 1520.0s] 再做一个, [1520.0s -> 1521.0s] 模型预测控制这样的, [1521.0s -> 1522.0s] 资路, [1522.0s -> 1523.0s] 展示出来的能力, [1523.0s -> 1524.0s] 也都越来越强, [1524.0s -> 1525.0s] 但是, [1525.0s -> 1526.0s] 如果透过, [1526.0s -> 1527.0s] 这些, [1527.0s -> 1528.0s] 展示能力, [1528.0s -> 1529.0s] 看本质的话, [1529.0s -> 1530.0s] 研发的方式, [1530.0s -> 1531.0s] 其实是, [1531.0s -> 1532.0s] 没有特别大的颠覆, [1532.0s -> 1533.0s] 所以, [1533.0s -> 1534.0s] 还是在那儿以往的, [1534.0s -> 1535.0s] 比如说, [1535.0s -> 1536.0s] 你可以讲, [1536.0s -> 1537.0s] 我们在, [1537.0s -> 1538.0s] 超大圆模型, [1538.0s -> 1539.0s] 的作业, [1539.0s -> 1540.0s] 对, [1540.0s -> 1541.0s] 还是在, [1541.0s -> 1542.0s] 以这种, [1542.0s -> 1543.0s] 去预测未来, [1543.0s -> 1544.0s] 然后, [1544.0s -> 1545.0s] 去, [1545.0s -> 1546.0s] 用这样的数据, [1546.0s -> 1547.0s] 来去训我们的, [1547.0s -> 1548.0s] 具身质能模型, [1548.0s -> 1549.0s] 但是, [1549.0s -> 1550.0s] 我其实更期待的, [1550.0s -> 1551.0s] 学校, [1551.0s -> 1552.0s] 这边研究上, [1552.0s -> 1553.0s] 所做的, [1553.0s -> 1554.0s] 也是希望能找到, [1554.0s -> 1555.0s] 一个不一样的方式, [1555.0s -> 1556.0s] 因为, [1556.0s -> 1557.0s] 本质上, [1557.0s -> 1558.0s] 预言, [1558.0s -> 1559.0s] 它是一个, [1559.0s -> 1560.0s] 维度, [1560.0s -> 1561.0s] 就是, [1561.0s -> 1562.0s] 那么多, [1562.0s -> 1563.0s] 这个, [1563.0s -> 1564.0s] 文本的token space, [1564.0s -> 1565.0s] 是一个理想的空间, [1565.0s -> 1566.0s] 我的, [1566.0s -> 1567.0s] 观测空间, [1567.0s -> 1568.0s] 和我的预测空间, [1568.0s -> 1569.0s] 而且是一致的, [1569.0s -> 1570.0s] 对, [1570.0s -> 1571.0s] 它是一个, [1571.0s -> 1572.0s] 很特殊的问题, [1572.0s -> 1573.0s] 但是, [1573.0s -> 1574.0s] 在, [1574.0s -> 1575.0s] 机器人行业, [1575.0s -> 1576.0s] 我的输入空间, [1576.0s -> 1577.0s] 是如此之多, [1577.0s -> 1578.0s] 这个, [1578.0s -> 1579.0s] 对, [1579.0s -> 1581.0s] 不是视觉观测到的一些状态, [1581.0s -> 1582.0s] 比如说, [1582.0s -> 1583.0s] 这个水是满的, [1583.0s -> 1584.0s] 还是, [1584.0s -> 1585.0s] 给看不见的水是满的, [1585.0s -> 1586.0s] 还是空的, [1586.0s -> 1587.0s] 包括, [1587.0s -> 1588.0s] 一个电器, [1588.0s -> 1589.0s] 还是开的, [1589.0s -> 1590.0s] 还是关的, [1590.0s -> 1591.0s] 有些状态, [1591.0s -> 1592.0s] 就是各种各样的状态, [1592.0s -> 1593.0s] 都要融合进来, [1593.0s -> 1594.0s] 而你输出空间, [1594.0s -> 1595.0s] 就是, [1595.0s -> 1596.0s] 机器人自己的, [1596.0s -> 1597.0s] 一些控制, [1597.0s -> 1598.0s] 所以它, [1598.0s -> 1599.0s] 跟大原木星, [1599.0s -> 1600.0s] 其实, [1600.0s -> 1601.0s] 本质是不一样的一个问题, [1601.0s -> 1602.0s] 但是, [1602.0s -> 1603.0s] 我们现在呢, [1603.0s -> 1604.0s] 还是在大原木星的方式, [1604.0s -> 1605.0s] 再去推, [1605.0s -> 1606.0s] 所以这是我为什么说, [1606.0s -> 1607.0s] 如果, [1607.0s -> 1608.0s] 拉长时间, [1608.0s -> 1609.0s] 就是, [1609.0s -> 1610.0s] 机器人模型的, [1610.0s -> 1611.0s] 训练推理饭市场, [1611.0s -> 1612.0s] 可能还需要, [1612.0s -> 1613.0s] 更多的思考, [1613.0s -> 1614.0s] 因为我不是, [1614.0s -> 1615.0s] 相关, [1615.0s -> 1616.0s] 技术专业背景的, [1616.0s -> 1617.0s] 但是, [1617.0s -> 1618.0s] 我们也, [1618.0s -> 1619.0s] 做了不同方向, [1619.0s -> 1620.0s] 各种, [1620.0s -> 1621.0s] 类型的投资, [1621.0s -> 1622.0s] 从我的角度, [1622.0s -> 1623.0s] 有一个这样的, [1623.0s -> 1624.0s] 简单规纳, [1624.0s -> 1625.0s] 我不知道, [1625.0s -> 1626.0s] 从内行的角度来看, [1626.0s -> 1627.0s] 这边, [1627.0s -> 1628.0s] 多少正确, [1628.0s -> 1629.0s] 多少错误, [1629.0s -> 1630.0s] 就是, [1630.0s -> 1631.0s] 我们先, [1631.0s -> 1632.0s] 把大原模型的, [1632.0s -> 1633.0s] 技术和架构问题, [1633.0s -> 1634.0s] 先放在一边,

[1634.0s -> 1635.0s] 就比如说, [1635.0s -> 1636.0s] 它是个多少围, [1636.0s -> 1637.0s] 它是个几千围的, [1637.0s -> 1638.0s] 生成式的, [1638.0s -> 1639.0s] 文本内容, [1639.0s -> 1640.0s] 来打比方, [1640.0s -> 1641.0s] 那这种, [1641.0s -> 1642.0s] 因为语言和语言, [1642.0s -> 1643.0s] 或者叫字和词, [1643.0s -> 1644.0s] 和词之间的关系, [1644.0s -> 1645.0s] 我把它称之为, [1645.0s -> 1646.0s] 叫是个逻辑关系, [1646.0s -> 1647.0s] 相对不是, [1647.0s -> 1648.0s] 绝对零和一的关系, [1648.0s -> 1649.0s] 就是, [1649.0s -> 1650.0s] 简单来讲, [1650.0s -> 1651.0s] 咱们一句话, [1651.0s -> 1652.0s] 可以有很多个表达方法, [1652.0s -> 1653.0s] 也有很多个, [1653.0s -> 1654.0s] 用很多个词, [1654.0s -> 1655.0s] 或者用很多不同的词, [1655.0s -> 1656.0s] 都可以表达同一个意思, [1656.0s -> 1657.0s] 这是, [1657.0s -> 1658.0s] 所谓逻辑关系, [1658.0s -> 1659.0s] 但是它有一些, [1659.0s -> 1660.0s] 简单约束, [1660.0s -> 1661.0s] 比如说, [1661.0s -> 1662.0s] 通常动词, [1662.0s -> 1663.0s] 它是个, [1663.0s -> 1664.0s] 就是能, [1664.0s -> 1665.0s] 很多可能性, [1665.0s -> 1666.0s] 去达到同一个目的, [1666.0s -> 1668.0s] 或者完成同一个生成, [1668.0s -> 1669.0s] 在这个机器人上, [1669.0s -> 1670.0s] 它不是个逻辑关系, [1670.0s -> 1672.0s] 它是个纯物理关系, [1672.0s -> 1673.0s] 就比如说, [1673.0s -> 1674.0s] 你肯定不能, [1674.0s -> 1675.0s] 把手穿过桌面, [1675.0s -> 1676.0s] 对吧, [1676.0s -> 1677.0s] 就是, [1677.0s -> 1678.0s] 你也不能, [1678.0s -> 1679.0s] 直接把瓶子捏破, [1679.0s -> 1680.0s] 就大概, [1680.0s -> 1681.0s] 它有一些, [1681.0s -> 1682.0s] 非常零一零一的问题, [1682.0s -> 1683.0s] 就是, [1683.0s -> 1684.0s] 它不是说, [1684.0s -> 1685.0s] 什么样都可以, [1685.0s -> 1686.0s] 或者哪种执行方法都可以, [1686.0s -> 1687.0s] 它受到了一些, [1687.0s -> 1688.0s] 非常, [1688.0s -> 1689.0s] 严格的, [1689.0s -> 1691.0s] 物理和环境条件的约束, [1691.0s -> 1692.0s] 甚至为, [1692.0s -> 1693.0s] 有非常强的物理约束, [1693.0s -> 1694.0s] 就它有很多, [1694.0s -> 1695.0s] 零和一之间, [1695.0s -> 1696.0s] 没有零点一, [1696.0s -> 1697.0s] 零点三这样的, [1697.0s -> 1698.0s] 这个过程, [1698.0s -> 1699.0s] 那, [1699.0s -> 1700.0s] 所以回到, [1700.0s -> 1701.0s] 你刚才那个问题, [1701.0s -> 1702.0s] 就是, [1702.0s -> 1703.0s] 假定, [1703.0s -> 1704.0s] 之前经历过的, [1704.0s -> 1705.0s] 不管是VLAVLM, [1705.0s -> 1706.0s] 不管是哪一个, [1706.0s -> 1707.0s] 就是, [1707.0s -> 1708.0s] 从视觉, [1708.0s -> 1709.0s] 到语言, [1709.0s -> 1710.0s] 到再往下去, [1710.0s -> 1711.0s] 预测, [1711.0s -> 1712.0s] 机器人的这些, [1712.0s -> 1713.0s] 模型, [1713.0s -> 1714.0s] 都是过了一下, [1714.0s -> 1715.0s] 跟语言类似的, [1715.0s -> 1716.0s] 结构, [1716.0s -> 1717.0s] 演绎, [1717.0s -> 1718.0s] 但我刚才讲的这个, [1718.0s -> 1719.0s] 粗浅外行理解, [1719.0s -> 1720.0s] 它是两个逻辑上的事情, [1721.0s -> 1723.0s] 相对存关系和关系, [1723.0s -> 1725.0s] 一个是有硬的物理约束的, [1725.0s -> 1726.0s] 我不知道, [1726.0s -> 1727.0s] 从你的角度来看, [1727.0s -> 1729.0s] 因为它们存在这样的, [1729.0s -> 1730.0s] 我这个, [1730.0s -> 1731.0s] 这种外行理解的差别, [1731.0s -> 1732.0s] 它最终, [1732.0s -> 1733.0s] 有多大可能性, [1733.0s -> 1734.0s] 它会, [1734.0s -> 1736.0s] 借鉴大圆模型的, [1736.0s -> 1737.0s] 一部分能力, [1737.0s -> 1738.0s] 或者架构, [1738.0s -> 1739.0s] 或技术, [1739.0s -> 1740.0s] 但是有, [1740.0s -> 1742.0s] 多大一部分比例, [1742.0s -> 1743.0s] 最终会, [1743.0s -> 1744.0s] 超过, [1744.0s -> 1745.0s] 或者叫做, [1745.0s -> 1746.0s] 最少是, [1746.0s -> 1747.0s] 差异化和, [1747.0s -> 1748.0s] 语言, [1748.0s -> 1749.0s] 模型, [1749.0s -> 1750.0s] 的发展, [1750.0s -> 1752.0s] 这是一个挺难的问题, [1752.0s -> 1753.0s] 首先, [1753.0s -> 1754.0s] 我是完全赞同, [1754.0s -> 1755.0s] 对大圆模型的抽象, [1755.0s -> 1756.0s] 它, [1756.0s -> 1757.0s] 本质确实在学一个, [1757.0s -> 1758.0s] 逻辑关系, [1758.0s -> 1759.0s] 对, [1759.0s -> 1760.0s] 而且这个, [1760.0s -> 1761.0s] 约束, [1761.0s -> 1762.0s] 是, [1762.0s -> 1763.0s] 更简单的, [1763.0s -> 1764.0s] 它的解空间,

[1764.0s -> 1765.0s] 是更大的, [1765.0s -> 1766.0s] 以前上课的时候, [1766.0s -> 1767.0s] 我们会学习最优化, [1767.0s -> 1768.0s] 最优化的时候, [1768.0s -> 1769.0s] 你有可能是一个, [1769.0s -> 1770.0s] 不是全局, [1770.0s -> 1771.0s] 不是一个, [1771.0s -> 1772.0s] 担任, [1772.0s -> 1773.0s] 拖优化问题, [1773.0s -> 1774.0s] 有可能很多局部, [1774.0s -> 1775.0s] 最终解, [1775.0s -> 1776.0s] 大圆模型, [1776.0s -> 1777.0s] 有可能很多, [1777.0s -> 1778.0s] 局部, [1778.0s -> 1779.0s] 最终都是好的, [1779.0s -> 1780.0s] 就是物理世界方面, [1780.0s -> 1781.0s] 它, [1781.0s -> 1782.0s] 约束更多一些, [1782.0s -> 1783.0s] 然后, [1783.0s -> 1784.0s] 我认为它, [1784.0s -> 1785.0s] 不单单, [1785.0s -> 1786.0s] 不是逻辑, [1786.0s -> 1787.0s] 而说, [1787.0s -> 1788.0s] 它, [1788.0s -> 1789.0s] 也运寒的逻辑, [1789.0s -> 1790.0s] 当然, [1790.0s -> 1791.0s] 在此之上, [1791.0s -> 1792.0s] 它还有, [1792.0s -> 1793.0s] 物理和集合, [1793.0s -> 1794.0s] 对, [1794.0s -> 1795.0s] 就是, [1795.0s -> 1796.0s] 集合呢, [1796.0s -> 1797.0s] 是一些, [1797.0s -> 1798.0s] 运动学相关的, [1798.0s -> 1799.0s] 尺寸大小啊, [1799.0s -> 1800.0s] 包括, [1800.0s -> 1801.0s] 我这个关节, [1801.0s -> 1802.0s] 动多少度, [1802.0s -> 1803.0s] 相应的, [1803.0s -> 1804.0s] 我这个关节, [1804.0s -> 1805.0s] 带着这个鸽, [1805.0s -> 1806.0s] 动了多少, [1806.0s -> 1807.0s] 这个感的末端, [1807.0s -> 1808.0s] 动了多少, [1808.0s -> 1809.0s] 动力学相关的, [1809.0s -> 1810.0s] 施加多大力, [1810.0s -> 1811.0s] 那, [1811.0s -> 1812.0s] 这个东西, [1812.0s -> 1813.0s] 会被推多远, [1813.0s -> 1814.0s] 我擦是多少, [1814.0s -> 1815.0s] 相应的, [1815.0s -> 1816.0s] 会断多少度, [1816.0s -> 1817.0s] 所以, [1817.0s -> 1818.0s] 它是一个, [1818.0s -> 1819.0s] 更高级, [1819.0s -> 1820.0s] 更复杂的问题, [1820.0s -> 1821.0s] 当然, [1821.0s -> 1822.0s] 也包括, [1822.0s -> 1823.0s] 这个任务层面, [1823.0s -> 1824.0s] 就是逻辑问题, [1824.0s -> 1825.0s] 就是大象, [1825.0s -> 1826.0s] 放平产, [1826.0s -> 1827.0s] 分析不是那种, [1827.0s -> 1828.0s] 所以, [1828.0s -> 1829.0s] 是逻辑, [1829.0s -> 1830.0s] 集合和物理, [1830.0s -> 1831.0s] 对, [1831.0s -> 1832.0s] 所以, [1832.0s -> 1833.0s] 具身质呢, [1833.0s -> 1834.0s] 我觉得, [1834.0s -> 1835.0s] 可以给一个, [1835.0s -> 1836.0s] 不是那么, [1836.0s -> 1837.0s] 正式的定义的话, [1837.0s -> 1838.0s] 改变, [1838.0s -> 1839.0s] 目标物体的状态, [1839.0s -> 1840.0s] 对, [1840.0s -> 1841.0s] 那它, [1841.0s -> 1842.0s] 是一个, [1842.0s -> 1843.0s] 很多时候是有, [1843.0s -> 1844.0s] 精确解的, [1844.0s -> 1845.0s] 做到, [1845.0s -> 1846.0s] 就是做到, [1846.0s -> 1847.0s] 没做到, [1847.0s -> 1848.0s] 就是没做到, [1848.0s -> 1849.0s] 但是语言, [1849.0s -> 1850.0s] 就, [1850.0s -> 1851.0s] 肯定是可以, [1851.0s -> 1852.0s] 很模棱两可的, [1852.0s -> 1853.0s] 所以, [1853.0s -> 1854.0s] 这个好像是对的, [1854.0s -> 1855.0s] 这个是错的, [1855.0s -> 1856.0s] 之前这块, [1856.0s -> 1857.0s] 确实是, [1857.0s -> 1858.0s] 它是一个, [1858.0s -> 1859.0s] 比, [1859.0s -> 1860.0s] 大圆谋星, [1860.0s -> 1861.0s] 更难的一个问题, [1861.0s -> 1862.0s] 所以, [1862.0s -> 1863.0s] 如果这个假事对的话, [1863.0s -> 1864.0s] 我们肯定是可以借借, [1864.0s -> 1865.0s] 大圆谋星, [1865.0s -> 1866.0s] 很多东西过来的, [1866.0s -> 1867.0s] 应该是做到, [1867.0s -> 1868.0s] 这种方式, [1868.0s -> 1869.0s] 去学, [1869.0s -> 1870.0s] 过去看到的, [1870.0s -> 1871.0s] 和未来的一些关联, [1871.0s -> 1872.0s] 对, [1872.0s -> 1873.0s] 我们用这个关联, [1873.0s -> 1874.0s] 来近似因果, [1874.0s -> 1876.0s] 那这个关联近似因果的, [1876.0s -> 1877.0s] 范氏, [1877.0s -> 1879.0s] 一定是可以借借到, [1879.0s -> 1880.0s] 机权里面的, [1880.0s -> 1881.0s] 是, [1881.0s -> 1882.0s] 但是如果, [1882.0s -> 1883.0s] 只有这个, [1883.0s -> 1884.0s] 它是不够的, [1884.0s -> 1885.0s] 因为, [1885.0s -> 1886.0s] 在,

[1886.0s -> 1887.0s] 这个文字上, [1887.0s -> 1888.0s] 它其实没有时间概念, [1888.0s -> 1889.0s] 物理上, [1889.0s -> 1890.0s] 它是有时间纬度, [1890.0s -> 1891.0s] 然后, [1891.0s -> 1892.0s] 你就设计了几颗, [1892.0s -> 1893.0s] 它的约束也有很多, [1893.0s -> 1894.0s] 所以在, [1894.0s -> 1895.0s] 机权人领域, [1895.0s -> 1896.0s] 时间, [1896.0s -> 1897.0s] 大圆谋形的这个, [1897.0s -> 1898.0s] 训练, [1898.0s -> 1899.0s] 和数据, [1899.0s -> 1900.0s] 选择以后, [1900.0s -> 1901.0s] 推理的范氏, [1901.0s -> 1902.0s] 同时, [1902.0s -> 1903.0s] 我相信, [1903.0s -> 1904.0s] 如可白, [1904.0s -> 1905.0s] 因果, [1905.0s -> 1906.0s] 约束, [1906.0s -> 1907.0s] 加进来, [1907.0s -> 1908.0s] 时间纬度, [1908.0s -> 1909.0s] 加进来, [1909.0s -> 1910.0s] 这个是一个必须的, [1910.0s -> 1911.0s] 一个东西, [1911.0s -> 1912.0s] 那我们先把这个, [1912.0s -> 1913.0s] 稍微抽象一点的问题, [1913.0s -> 1914.0s] 这个讨论题, [1914.0s -> 1915.0s] 先越过, [1915.0s -> 1916.0s] 我们回到, [1916.0s -> 1917.0s] 言诺本身, [1917.0s -> 1918.0s] 最少两年半以前, [1918.0s -> 1919.0s] 在确定应用, [1919.0s -> 1920.0s] 场景方向的时候, [1920.0s -> 1921.0s] 你就选了一些, [1921.0s -> 1922.0s] 相对于中不同的, [1922.0s -> 1923.0s] 因为那个时候, [1923.0s -> 1924.0s] 大家都还在做各种各样, [1924.0s -> 1925.0s] 就像我们讲的, [1925.0s -> 1926.0s] 点衣服, [1926.0s -> 1927.0s] 什么倒咖啡之类的, [1927.0s -> 1928.0s] 你稍微选了一个, [1928.0s -> 1929.0s] 比较不同, [1929.0s -> 1930.0s] 而且是, [1930.0s -> 1932.0s] 相对难做的方向, [1932.0s -> 1933.0s] 那, [1933.0s -> 1934.0s] 你能给大家, [1934.0s -> 1935.0s] 解释和描述一下, [1935.0s -> 1936.0s] 当时你是怎么考虑的, [1936.0s -> 1937.0s] 以及那个方向, [1937.0s -> 1938.0s] 最终, [1938.0s -> 1939.0s] 它可能会, [1939.0s -> 1940.0s] 难在什么地方, [1940.0s -> 1941.0s] 对于当时的, [1941.0s -> 1942.0s] 极其人, [1942.0s -> 1943.0s] 和现在的, [1943.0s -> 1944.0s] 极其人能力来看, [1944.0s -> 1945.0s] 上次我们, [1945.0s -> 1946.0s] 跟佛如谈的时候, [1946.0s -> 1947.0s] 提到这个, [1947.0s -> 1948.0s] 不可能三角, [1948.0s -> 1949.0s] 通行性, [1949.0s -> 1950.0s] 可耗性, [1950.0s -> 1951.0s] 和速度, [1951.0s -> 1952.0s] 当时我们在, [1952.0s -> 1953.0s] 选场景的时候, [1953.0s -> 1954.0s] 说, [1954.0s -> 1955.0s] 我如果, [1955.0s -> 1956.0s] 同时达到, [1956.0s -> 1957.0s] 那我们牺牲哪一个点, [1957.0s -> 1958.0s] 我们当时的选择, [1958.0s -> 1959.0s] 就是, [1959.0s -> 1960.0s] 在速度上, [1960.0s -> 1961.0s] 不用那么苛求, [1961.0s -> 1962.0s] 只要达到, [1962.0s -> 1963.0s] 一定的预知, [1963.0s -> 1964.0s] 就可以了, [1964.0s -> 1965.0s] 然后在通行性上, [1965.0s -> 1966.0s] 我们, [1966.0s -> 1967.0s] 会把它拉下来一些, [1967.0s -> 1968.0s] 我们不追求, [1968.0s -> 1969.0s] 以下能解决, [1969.0s -> 1970.0s] 一万件事情, [1970.0s -> 1971.0s] 我们能解决, [1971.0s -> 1972.0s] 可能几十件, [1972.0s -> 1973.0s] 上百件, [1973.0s -> 1974.0s] 就是, [1974.0s -> 1975.0s] 会多一些事情, [1975.0s -> 1976.0s] 就够了, [1976.0s -> 1977.0s] 所以我们, [1977.0s -> 1978.0s] 基于这个假设, [1978.0s -> 1979.0s] 我们会去, [1979.0s -> 1980.0s] 寻找一些, [1980.0s -> 1981.0s] 场景领域, [1981.0s -> 1982.0s] 能够去, [1982.0s -> 1984.0s] 在可预见的时间内, [1984.0s -> 1985.0s] 去落地, [1985.0s -> 1986.0s] 去能摆, [1986.0s -> 1987.0s] 现在的, [1987.0s -> 1988.0s] 智能机器人, [1988.0s -> 1989.0s] 或者点点角, [1989.0s -> 1990.0s] 能够得到的一些, [1990.0s -> 1991.0s] 场景, [1991.0s -> 1992.0s] 但是同时呢, [1992.0s -> 1993.0s] 也不希望这个场景, [1993.0s -> 1994.0s] 会把我们的天花本哈死, [1994.0s -> 1995.0s] 会把我们, [1995.0s -> 1996.0s] 牵引到, [1996.0s -> 1997.0s] 另外一些, [1997.0s -> 1998.0s] 更短视的一些, [1998.0s -> 1999.0s] 技术路线上, [1999.0s -> 2000.0s] 所以我们, [2000.0s -> 2001.0s] 希望选一些, [2001.0s -> 2002.0s] 难的场景, [2002.0s -> 2003.0s] 那, [2003.0s -> 2004.0s] 怀穿着这个, [2004.0s -> 2005.0s] 我们是希望找这种, [2005.0s -> 2006.0s] 让又有长序列, [2006.0s -> 2007.0s] 要严格考打, [2007.0s -> 2008.0s] 这个逻辑,

[2008.0s -> 2009.0s] 能力的, [2009.0s -> 2010.0s] 然后, [2010.0s -> 2011.0s] 比较高精度的, [2011.0s -> 2012.0s] 对, [2012.0s -> 2013.0s] 要去, [2013.0s -> 2014.0s] 考问这个, [2014.0s -> 2015.0s] 几何和屋里, [2015.0s -> 2016.0s] 同时要高柔性, [2016.0s -> 2017.0s] 那, [2017.0s -> 2018.0s] 它一定是一个, [2018.0s -> 2019.0s] 智能的东西, [2019.0s -> 2020.0s] 去别于以往的, [2020.0s -> 2021.0s] 机器人, [2021.0s -> 2022.0s] 或者工业自动化, [2022.0s -> 2023.0s] 所以这是, [2023.0s -> 2024.0s] 我们希望, [2024.0s -> 2025.0s] 在一个, [2025.0s -> 2026.0s] 最难的, [2026.0s -> 2027.0s] 考场里面, [2027.0s -> 2028.0s] 去, [2028.0s -> 2029.0s] 去嵌延我们的, [2029.0s -> 2030.0s] 技术研发, [2030.0s -> 2031.0s] 去提升这个能力, [2031.0s -> 2032.0s] 如果能把, [2032.0s -> 2033.0s] 这些事情做好, [2033.0s -> 2034.0s] 那一定能够, [2034.0s -> 2035.0s] 解决, [2035.0s -> 2036.0s] 所有, [2036.0s -> 2037.0s] 未来走向, [2037.0s -> 2038.0s] 家用的, [2038.0s -> 2039.0s] 一万点事情, [2039.0s -> 2040.0s] 从本身来讲, [2040.0s -> 2041.0s] 它, [2041.0s -> 2042.0s] 一方面, [2042.0s -> 2043.0s] 具备了, [2043.0s -> 2044.0s] 我们, [2044.0s -> 2045.0s] 从做, [2045.0s -> 2046.0s] 军生智能, [2046.0s -> 2047.0s] 机器人的一些诉求, [2047.0s -> 2048.0s] 同时, [2048.0s -> 2049.0s] 从行业本身, [2049.0s -> 2050.0s] 它又是一个, [2050.0s -> 2051.0s] 真的需求, [2051.0s -> 2052.0s] 就是, [2052.0s -> 2053.0s] 整个, [2053.0s -> 2054.0s] 科业智能, [2054.0s -> 2055.0s] 算实验室, [2055.0s -> 2056.0s] 生化环采行业, [2056.0s -> 2057.0s] 生物医药行业, [2057.0s -> 2058.0s] 他们, [2058.0s -> 2059.0s] 确实是, [2059.0s -> 2060.0s] 需要这样的, [2060.0s -> 2061.0s] 高柔性, [2061.0s -> 2062.0s] 高精度, [2062.0s -> 2063.0s] 高智能的, [2063.0s -> 2064.0s] 这些, [2064.0s -> 2065.0s] 智能体, [2065.0s -> 2066.0s] 来做, [2066.0s -> 2067.0s] 这些物理操作, [2067.0s -> 2068.0s] 一方面, [2068.0s -> 2069.0s] 来做的, [2069.0s -> 2070.0s] 大家在做, [2070.0s -> 2071.0s] 实验的时候, [2071.0s -> 2072.0s] 通常, [2072.0s -> 2073.0s] 很大一个, [2073.0s -> 2074.0s] 碰点, [2074.0s -> 2075.0s] 就是, [2075.0s -> 2076.0s] 实验不可浮现, [2076.0s -> 2077.0s] 你就是, [2077.0s -> 2078.0s] 人做这些事情, [2078.0s -> 2079.0s] 就是, [2079.0s -> 2080.0s] 不可浮现的, [2080.0s -> 2081.0s] 即使是, [2081.0s -> 2082.0s] 实验流程, [2082.0s -> 2083.0s] 写好, [2083.0s -> 2084.0s] 我要把, [2084.0s -> 2085.0s] ABCD, [2085.0s -> 2086.0s] 在25度, [2086.0s -> 2087.0s] 去混云, [2087.0s -> 2088.0s] 那, [2088.0s -> 2089.0s] 有的人对它, [2089.0s -> 2090.0s] 理解, [2090.0s -> 2091.0s] 就是, [2091.0s -> 2092.0s] 我从, [2092.0s -> 2093.0s] 自度的冰箱里面, [2093.0s -> 2094.0s] 拿出来, [2094.0s -> 2095.0s] 拿到25度, [2095.0s -> 2096.0s] 放, [2096.0s -> 2097.0s] 一个小时, [2097.0s -> 2098.0s] 其实还是4度的, [2098.0s -> 2099.0s] 然后, [2099.0s -> 2100.0s] 倒完以后, [2100.0s -> 2101.0s] 发现, [2101.0s -> 2102.0s] 才想起, [2102.0s -> 2103.0s] 我要C, [2103.0s -> 2104.0s] 去冰箱, [2104.0s -> 2105.0s] 又拿, [2105.0s -> 2106.0s] 拿完以后, [2106.0s -> 2107.0s] 夸一刀, [2107.0s -> 2108.0s] 那反应条件, [2108.0s -> 2109.0s] 自然就是不一样, [2109.0s -> 2110.0s] 就是, [2110.0s -> 2111.0s] 类似这样的, [2111.0s -> 2112.0s] 细节很多, [2112.0s -> 2113.0s] 更别提, [2113.0s -> 2114.0s] 它做的时间, [2114.0s -> 2115.0s] 还可能打过困体, [2115.0s -> 2116.0s] 你就是, [2116.0s -> 2117.0s] 这种小细节, [2117.0s -> 2118.0s] 就导致, [2118.0s -> 2119.0s] 它, [2119.0s -> 2120.0s] 很难浮现, [2120.0s -> 2121.0s] 就是, [2121.0s -> 2122.0s] 所谓的, [2122.0s -> 2123.0s] 10年, [2123.0s -> 2124.0s] 10个亿, [2124.0s -> 2125.0s] 成功力, [2125.0s -> 2126.0s] 感知识, [2126.0s -> 2127.0s] 就是, [2127.0s -> 2128.0s] 我每一个步骤,

[2128.0s -> 2129.0s] 都是, [2129.0s -> 2130.0s] 忠诚的, [2130.0s -> 2131.0s] 细腻的做下来, [2131.0s -> 2132.0s] 今天能做, [2132.0s -> 2133.0s] 第二天还是能做, [2133.0s -> 2134.0s] 每个条件, [2134.0s -> 2135.0s] 参数, [2135.0s -> 2136.0s] 都能记录下来, [2136.0s -> 2137.0s] 这样, [2137.0s -> 2138.0s] 提升它的可靠性, [2138.0s -> 2139.0s] 另外一个点的, [2139.0s -> 2140.0s] 就是, [2140.0s -> 2141.0s] 现在的这些, [2141.0s -> 2142.0s] 大量的, [2142.0s -> 2143.0s] 实验室, [2143.0s -> 2144.0s] 研发, [2144.0s -> 2145.0s] 依赖仪器, [2145.0s -> 2146.0s] 这些仪器, [2146.0s -> 2147.0s] 是依赖于人去操作的, [2147.0s -> 2148.0s] 如果, [2148.0s -> 2149.0s] 一旦人不工作, [2149.0s -> 2150.0s] 人可能, [2150.0s -> 2151.0s] 就是, [2151.0s -> 2152.0s] 着酒晚午这样, [2152.0s -> 2153.0s] 那, [2153.0s -> 2154.0s] 一旦不工作, [2154.0s -> 2155.0s] 仪器, [2155.0s -> 2156.0s] 利用率可能, [2156.0s -> 2157.0s] 只有百分之三四十, [2157.0s -> 2158.0s] 这些仪器, [2158.0s -> 2159.0s] 往往都是很贵的, [2159.0s -> 2160.0s] 所以就, [2160.0s -> 2161.0s] 导致它的研发流程, [2161.0s -> 2162.0s] 就会被无限的拉长, [2162.0s -> 2163.0s] 这相比于, [2163.0s -> 2164.0s] 我们在集权机领域, [2164.0s -> 2165.0s] 在AI领域, [2165.0s -> 2166.0s] 我们是, [2166.0s -> 2167.0s] 能够把时间维度缩短, [2167.0s -> 2168.0s] 我们通常, [2168.0s -> 2169.0s] 会并行的, [2169.0s -> 2170.0s] 跑一大堆实验, [2170.0s -> 2171.0s] 但是, [2171.0s -> 2172.0s] 在这样的行业里面, [2172.0s -> 2173.0s] 它是很难去, [2173.0s -> 2174.0s] 做这些事情的, [2174.0s -> 2175.0s] 所以如果, [2175.0s -> 2176.0s] 我们一旦能有机器人, [2176.0s -> 2177.0s] 去做一个实验助理, [2177.0s -> 2178.0s] 去, [2178.0s -> 2179.0s] 天然地把, [2179.0s -> 2180.0s] 这些仪器, [2180.0s -> 2181.0s] 来调动起来, [2181.0s -> 2182.0s] 那就可以, [2182.0s -> 2183.0s] 把它研发效率, [2183.0s -> 2184.0s] 大大的提升, [2184.0s -> 2185.0s] 所以一方面, [2185.0s -> 2186.0s] 是在质量上, [2186.0s -> 2187.0s] 我们能够提升, [2187.0s -> 2188.0s] 另一方面, [2188.0s -> 2189.0s] 在这个, [2189.0s -> 2190.0s] 提高它的吞吐量上, [2190.0s -> 2191.0s] 体质又能提亮, [2191.0s -> 2192.0s] 这两块, [2192.0s -> 2193.0s] 我们都能够, [2193.0s -> 2194.0s] 产生更大的价值, [2194.0s -> 2195.0s] 更不用说, [2195.0s -> 2196.0s] 有一些是危险的, [2196.0s -> 2197.0s] 本身就不适合, [2197.0s -> 2198.0s] 人去做的事情, [2198.0s -> 2199.0s] 所以我们从, [2199.0s -> 2200.0s] 行业的需求来讲, [2200.0s -> 2201.0s] 也是希望, [2201.0s -> 2202.0s] 来去做, [2202.0s -> 2203.0s] 能够真正的, [2203.0s -> 2204.0s] 负能真正的, [2204.0s -> 2205.0s] 去产生一些, [2205.0s -> 2206.0s] 产业价值, [2206.0s -> 2207.0s] 我们现在, [2207.0s -> 2208.0s] AI已经, [2208.0s -> 2209.0s] 能够让我们, [2209.0s -> 2210.0s] 快速的走向, [2210.0s -> 2211.0s] 这个, [2211.0s -> 2212.0s] 好吃懒惰, [2212.0s -> 2213.0s] 至少是, [2213.0s -> 2214.0s] 现在已经, [2214.0s -> 2215.0s] 不用怎么思考了, [2215.0s -> 2216.0s] 也来帮我们思考, [2216.0s -> 2217.0s] 机器人呢, [2217.0s -> 2218.0s] 应该能够比较快的, [2218.0s -> 2219.0s] 帮助我们, [2219.0s -> 2220.0s] 不用怎么做事情, [2220.0s -> 2221.0s] 另外一个理想, [2221.0s -> 2222.0s] 产生不老, [2222.0s -> 2223.0s] 应该也能比较, [2223.0s -> 2224.0s] 快速的达到, [2224.0s -> 2225.0s] 如果, [2225.0s -> 2226.0s] 它的科研方式, [2226.0s -> 2227.0s] 研发方式, [2227.0s -> 2228.0s] 是能够有一个, [2228.0s -> 2229.0s] 比较好的变革的话, [2229.0s -> 2230.0s] 对, [2230.0s -> 2231.0s] 这个, [2231.0s -> 2232.0s] 跟你这个, [2232.0s -> 2233.0s] 方向相关, [2233.0s -> 2234.0s] 就是生化还材, [2234.0s -> 2235.0s] 当然生化还材, [2235.0s -> 2236.0s] 在我们念书的时候, [2236.0s -> 2237.0s] 都是所谓, [2238.0s -> 2239.0s] 那今天他们都变成, [2239.0s -> 2240.0s] 重要和, [2240.0s -> 2241.0s] 热门的方向了, [2241.0s -> 2242.0s] 那个, [2242.0s -> 2243.0s] 就跟你讲的这个, [2243.0s -> 2244.0s] 饭食变化里面, [2244.0s -> 2245.0s] 还有个, [2245.0s -> 2246.0s] 多多少少, [2246.0s -> 2247.0s] 最近的好消息, [2247.0s -> 2248.0s] 就是, [2248.0s -> 2249.0s] 因为就在这周,

[2249.0s -> 2250.0s] 原来在新冠期间, [2250.0s -> 2252.0s] 做这个MRI疫苗的这个, [2252.0s -> 2253.0s] Moderna这个公司, [2253.0s -> 2254.0s] Moderna这个公司, [2254.0s -> 2255.0s] 这三期零创, [2255.0s -> 2256.0s] 得到显著结果, [2256.0s -> 2257.0s] 虽然还没有最后, [2257.0s -> 2258.0s] 披露所有数据, [2258.0s -> 2259.0s] 但是, [2259.0s -> 2260.0s] 它已经claim, [2260.0s -> 2261.0s] 三期有显出结果, [2261.0s -> 2262.0s] 里面也包括了, [2262.0s -> 2263.0s] 药物联用, [2263.0s -> 2264.0s] 但总而言之, [2264.0s -> 2265.0s] 大家把它, [2265.0s -> 2266.0s] 总结为的, [2266.0s -> 2267.0s] 就是, [2267.0s -> 2268.0s] 个性化的, [2268.0s -> 2269.0s] 某种, [2269.0s -> 2270.0s] 肿瘤, [2270.0s -> 2271.0s] 疫苗, [2271.0s -> 2272.0s] 算严正成功, [2272.0s -> 2273.0s] 然后, [2273.0s -> 2274.0s] 这算是个大的突破, [2274.0s -> 2275.0s] 当然, [2275.0s -> 2276.0s] 在新闻里面, [2276.0s -> 2277.0s] 他们也说了, [2277.0s -> 2278.0s] 这件事情, [2278.0s -> 2279.0s] 在筛选的时候, [2279.0s -> 2280.0s] 用了AI, [2280.0s -> 2281.0s] 当然, [2281.0s -> 2282.0s] 看起来, [2282.0s -> 2283.0s] 不一定是, [2283.0s -> 2284.0s] 用了大圆模型, [2284.0s -> 2285.0s] 但是, [2285.0s -> 2286.0s] 用了一些AI, [2286.0s -> 2287.0s] 那所以, [2287.0s -> 2288.0s] 它导致的结果, [2288.0s -> 2289.0s] 就你讲的, [2289.0s -> 2290.0s] 科研前面的, [2290.0s -> 2291.0s] 发现方式, [2291.0s -> 2292.0s] 有了一些变化, [2292.0s -> 2293.0s] 然后, [2293.0s -> 2294.0s] 为什么这事, [2294.0s -> 2295.0s] 跟你有关呢, [2295.0s -> 2296.0s] 发现, [2296.0s -> 2297.0s] 就是, [2297.0s -> 2298.0s] 跟药物发现有关的, [2298.0s -> 2299.0s] 前面那部分, [2299.0s -> 2300.0s] 犯事的话, [2300.0s -> 2301.0s] 最大的瓶颈, [2301.0s -> 2302.0s] 就来到了湿试验, [2302.0s -> 2303.0s] 因为, [2303.0s -> 2304.0s] 只有湿试验, [2304.0s -> 2305.0s] 能提供, [2305.0s -> 2306.0s] 需要配套, [2306.0s -> 2307.0s] 这个发现效率, [2307.0s -> 2308.0s] 提供足够多, [2308.0s -> 2309.0s] 足够快, [2309.0s -> 2310.0s] 足够整齐的, [2310.0s -> 2311.0s] 大量的数据, [2311.0s -> 2312.0s] 来进行, [2312.0s -> 2313.0s] 执行, [2313.0s -> 2314.0s] 和数据循环, [2314.0s -> 2315.0s] 就是, [2315.0s -> 2316.0s] 不管它是, [2316.0s -> 2317.0s] 证明是对的, [2317.0s -> 2318.0s] 还是证明是错的, [2318.0s -> 2319.0s] 都要回到, [2319.0s -> 2320.0s] 那个预测过程里, [2320.0s -> 2321.0s] 去用AI, [2321.0s -> 2322.0s] 再来进行, [2322.0s -> 2323.0s] 好处, [2323.0s -> 2324.0s] 就是大家今天, [2324.0s -> 2325.0s] 开始慢慢, [2325.0s -> 2326.0s] 共识到说, [2326.0s -> 2327.0s] AI用在生物医疗里, [2327.0s -> 2328.0s] 显著提高了, [2328.0s -> 2329.0s] 发现效率, [2329.0s -> 2330.0s] AI用在生物医疗, [2330.0s -> 2331.0s] 和药物发现里, [2331.0s -> 2332.0s] 平静, [2332.0s -> 2333.0s] 接下来, [2333.0s -> 2334.0s] 就来到了, [2334.0s -> 2335.0s] 湿试验本身, [2335.0s -> 2336.0s] 湿试验怎么能够提高, [2336.0s -> 2337.0s] 我不知道, [2337.0s -> 2338.0s] 是不是就跟, [2338.0s -> 2339.0s] 刚才你讲的, [2339.0s -> 2340.0s] 你们的应用方向有关, [2340.0s -> 2341.0s] 但是问题是, [2341.0s -> 2342.0s] 虽然人做试验, [2342.0s -> 2343.0s] 有很多不可负陷, [2343.0s -> 2344.0s] 这就, [2344.0s -> 2345.0s] 让我想起了, [2345.0s -> 2346.0s] 痛苦的, [2346.0s -> 2347.0s] 我当时, [2347.0s -> 2348.0s] 在国外念研究生的时候, [2348.0s -> 2349.0s] 就是, [2349.0s -> 2351.0s] 每次和每次都不太一样, [2351.0s -> 2352.0s] 在化学实验, [2352.0s -> 2353.0s] 和化学合成里边, [2353.0s -> 2354.0s] 那我不知道, [2354.0s -> 2355.0s] 在这个过程当中, [2355.0s -> 2356.0s] 回到刚才我们那个问题的, [2356.0s -> 2357.0s] 第二个部分, [2357.0s -> 2359.0s] 它可能要迎来春天了, [2359.0s -> 2360.0s] 但是, [2360.0s -> 2361.0s] 到底今天机器人, [2361.0s -> 2363.0s] 能在这个湿试验, [2363.0s -> 2364.0s] 缓解和范围, [2364.0s -> 2365.0s] 和流程里, [2365.0s -> 2366.0s] 从能力上, [2366.0s -> 2367.0s] 它今天到底能解决到, [2367.0s -> 2368.0s] 哪些程度, [2368.0s -> 2369.0s] 能解决哪些问题, [2369.0s -> 2370.0s] 能做到, [2370.0s -> 2371.0s] 什么情况, [2371.0s -> 2372.0s] 能带来哪些好处, [2372.0s -> 2373.0s] 有哪些难的,

[2373.0s -> 2374.0s] 技术节点, [2374.0s -> 2375.0s] 被你们, [2375.0s -> 2376.0s] 正在克服, [2376.0s -> 2377.0s] 或克服了, [2377.0s -> 2378.0s] 如何去远离多, [2378.0s -> 2379.0s] 做的, [2379.0s -> 2380.0s] 这些, [2380.0s -> 2381.0s] 积累的, [2381.0s -> 2382.0s] 这些能力, [2382.0s -> 2383.0s] 至少在, [2383.0s -> 2384.0s] 常见的, [2384.0s -> 2385.0s] 一些, [2385.0s -> 2386.0s] 实验流程里边, [2386.0s -> 2387.0s] 比如说, [2387.0s -> 2388.0s] 如何去处理, [2388.0s -> 2389.0s] 液体, [2389.0s -> 2390.0s] 固体, [2390.0s -> 2391.0s] 粉末状态, [2391.0s -> 2392.0s] 或是颗粒状的, [2392.0s -> 2393.0s] 这些样品, [2393.0s -> 2394.0s] 样品是一个很大的, [2394.0s -> 2395.0s] 要素, [2395.0s -> 2396.0s] 然后, [2396.0s -> 2397.0s] 常见的仪器, [2397.0s -> 2398.0s] 理性机, [2398.0s -> 2399.0s] PCR, [2399.0s -> 2400.0s] 很多, [2400.0s -> 2401.0s] 实验室里边, [2401.0s -> 2402.0s] 常见的那些流程, [2402.0s -> 2403.0s] 我们会, [2403.0s -> 2404.0s] 都拆解出来, [2404.0s -> 2405.0s] 对应出来, [2405.0s -> 2406.0s] 相应那些, [2406.0s -> 2407.0s] 常见的, [2407.0s -> 2408.0s] 一些流程, [2408.0s -> 2409.0s] 80%以上, [2409.0s -> 2410.0s] 都是可以浮现的, [2410.0s -> 2411.0s] 至少是一些, [2411.0s -> 2412.0s] 举步, [2412.0s -> 2413.0s] 已经能做到, [2413.0s -> 2414.0s] 小避缓了, [2414.0s -> 2415.0s] 相当于, [2415.0s -> 2416.0s] 给一些, [2416.0s -> 2417.0s] PI, [2417.0s -> 2418.0s] 或者, [2418.0s -> 2419.0s] 甚至一些, [2419.0s -> 2420.0s] 可以享福的, [2420.0s -> 2421.0s] 博士生, [2421.0s -> 2422.0s] 也可以有这样的, [2422.0s -> 2423.0s] 研究助理, [2423.0s -> 2424.0s] 相当于, [2424.0s -> 2425.0s] 它可以, [2425.0s -> 2426.0s] 发一个指令, [2426.0s -> 2427.0s] 然后, [2427.0s -> 2428.0s] 机器人可以, [2428.0s -> 2429.0s] 调度, [2429.0s -> 2430.0s] 实验室里边的, [2430.0s -> 2431.0s] 这些, [2431.0s -> 2432.0s] 相关的要素, [2432.0s -> 2433.0s] 去完成一个, [2433.0s -> 2434.0s] 举步的, [2434.0s -> 2435.0s] 避缓, [2435.0s -> 2436.0s] 再一步, [2436.0s -> 2437.0s] 所以我们, [2437.0s -> 2438.0s] 现在是, [2438.0s -> 2439.0s] 一个辅助, [2439.0s -> 2440.0s] 科研人员, [2440.0s -> 2441.0s] 研发人员的一个, [2441.0s -> 2442.0s] 阶段, [2442.0s -> 2443.0s] 并不会去, [2443.0s -> 2444.0s] 剥夺, [2444.0s -> 2445.0s] 或者去替代, [2445.0s -> 2446.0s] 这样的科学家, [2446.0s -> 2447.0s] 或者研发人员, [2447.0s -> 2448.0s] 我们还是, [2448.0s -> 2449.0s] 在一个, [2449.0s -> 2450.0s] 先做局部的, [2450.0s -> 2451.0s] 可以类比, [2451.0s -> 2452.0s] 像自动驾驶, [2452.0s -> 2453.0s] 一个, [2453.0s -> 2454.0s] 上高速以后, [2454.0s -> 2455.0s] 开一个L2的这种, [2455.0s -> 2456.0s] 我们现在, [2456.0s -> 2457.0s] 是做这种, [2457.0s -> 2458.0s] 辅助驾驶的能力, [2458.0s -> 2459.0s] 能够去, [2459.0s -> 2460.0s] 让他们更, [2460.0s -> 2461.0s] 解脱出来, [2461.0s -> 2462.0s] 这样就避免, [2462.0s -> 2463.0s] 一个人, [2463.0s -> 2464.0s] 他做实验, [2464.0s -> 2465.0s] 像做一组实验, [2465.0s -> 2466.0s] 他不用想, [2466.0s -> 2467.0s] 半夜两点, [2467.0s -> 2468.0s] 他回来再换一次衣, [2468.0s -> 2469.0s] 那机器人, [2469.0s -> 2470.0s] 就可以干这个事情, [2470.0s -> 2471.0s] 保证, [2471.0s -> 2472.0s] 他第二天早上, [2472.0s -> 2473.0s] 九点来之后, [2473.0s -> 2474.0s] 结果是读出来的, [2474.0s -> 2475.0s] 就我们现在, [2475.0s -> 2476.0s] 这种能力, [2476.0s -> 2477.0s] 已经达到了, [2477.0s -> 2478.0s] 刚才, [2478.0s -> 2479.0s] 风叔也提到, [2479.0s -> 2480.0s] 我完全赞同, [2480.0s -> 2481.0s] 现在AI是, [2481.0s -> 2482.0s] 数字AI, [2482.0s -> 2483.0s] 已经能把, [2483.0s -> 2484.0s] 生成的这个问题, [2484.0s -> 2485.0s] 解决得很好了, [2485.0s -> 2486.0s] 可以, [2486.0s -> 2487.0s] 在一个, [2487.0s -> 2488.0s] 极大的搜索空间里, [2488.0s -> 2489.0s] 去提假设, [2489.0s -> 2490.0s] 去提一些, [2490.0s -> 2491.0s] 这个可能的, [2491.0s -> 2492.0s] 这个, [2492.0s -> 2493.0s] 流程,

[2493.0s -> 2494.0s] 但是, [2494.0s -> 2495.0s] 那当你提了, [2495.0s -> 2496.0s] 流程, [2496.0s -> 2497.0s] 越来越多的时候, [2497.0s -> 2498.0s] 我们不可能, [2498.0s -> 2499.0s] 还是靠人海, [2499.0s -> 2500.0s] 这还是说, [2500.0s -> 2501.0s] 去一个一个去验, [2501.0s -> 2502.0s] 所以, [2502.0s -> 2503.0s] 现在就存在, [2503.0s -> 2504.0s] 一个很大的, [2504.0s -> 2505.0s] GAF, [2505.0s -> 2506.0s] 一个红钩, [2506.0s -> 2507.0s] 机器人里面, [2507.0s -> 2508.0s] 经常讲, [2508.0s -> 2509.0s] SIMTRIAL GAF, [2509.0s -> 2510.0s] 这个听材, [2510.0s -> 2511.0s] 像是一个, [2511.0s -> 2512.0s] 生成到验证的, [2512.0s -> 2513.0s] 一个GAF, [2513.0s -> 2514.0s] 就是, [2514.0s -> 2515.0s] 实验的, [2515.0s -> 2516.0s] 一个确实, [2516.0s -> 2517.0s] 一直在努力, [2517.0s -> 2518.0s] 去提供物理的, [2518.0s -> 2519.0s] 界面, [2519.0s -> 2520.0s] 这个物理界面, [2520.0s -> 2521.0s] 就可以去, [2521.0s -> 2522.0s] 把这个红钩, [2522.0s -> 2523.0s] 去用这些数据, [2523.0s -> 2524.0s] 那我们的, [2524.0s -> 2525.0s] 物理的一个, [2525.0s -> 2526.0s] API可以去, [2526.0s -> 2527.0s] 被调度, [2527.0s -> 2528.0s] 然后能够把, [2528.0s -> 2529.0s] 整个流程, [2529.0s -> 2530.0s] 就实现了, [2530.0s -> 2531.0s] 一个避缓, [2531.0s -> 2532.0s] 所以这点, [2532.0s -> 2533.0s] 是我们过去, [2533.0s -> 2534.0s] 几年所积累的, [2534.0s -> 2535.0s] 而且, [2535.0s -> 2536.0s] 现在也跟很多客户, [2536.0s -> 2537.0s] 在出现达成, [2537.0s -> 2538.0s] 证明, [2538.0s -> 2539.0s] 以及开始规模化的, [2539.0s -> 2540.0s] 一些事情, [2540.0s -> 2541.0s] 对好, [2541.0s -> 2542.0s] 那, [2542.0s -> 2543.0s] 这个大概能做什么, [2543.0s -> 2544.0s] 大概听懂了, [2544.0s -> 2545.0s] 但是, [2545.0s -> 2546.0s] 今天我们, [2546.0s -> 2547.0s] 回过头来看, [2547.0s -> 2548.0s] 就说, [2548.0s -> 2549.0s] 在做的这些事情里面, [2549.0s -> 2550.0s] 对于今天的机器人, [2550.0s -> 2551.0s] 或者大部分机器人来讲, [2551.0s -> 2552.0s] 哪些是最难的, [2552.0s -> 2553.0s] 我们先问一个, [2553.0s -> 2554.0s] 大家容易, [2554.0s -> 2555.0s] 直觉化的, [2555.0s -> 2556.0s] 这个感受一下, [2556.0s -> 2557.0s] 比如说, [2557.0s -> 2558.0s] 你参加, [2558.0s -> 2559.0s] 这个WRC, [2559.0s -> 2560.0s] 这个世界机器人大会, [2560.0s -> 2561.0s] 你也肯定会, [2561.0s -> 2562.0s] 到处溜的溜的, [2562.0s -> 2563.0s] 看看同行们, [2563.0s -> 2564.0s] 都搞成啥样的, [2564.0s -> 2565.0s] 都做了哪些东西, [2565.0s -> 2566.0s] 在你在, [2566.0s -> 2568.0s] WRC里满世界溜的的时候, [2568.0s -> 2569.0s] 你有看见, [2569.0s -> 2570.0s] 跟你们, [2570.0s -> 2571.0s] 做相关方向, [2571.0s -> 2572.0s] 类似尝试的公司吗, [2572.0s -> 2573.0s] 这是第一个问题, [2573.0s -> 2574.0s] 第二个问题, [2574.0s -> 2575.0s] 是就回过头来, [2575.0s -> 2576.0s] 假设这样的公司, [2576.0s -> 2577.0s] 很少的话, [2577.0s -> 2578.0s] 这些事情, [2578.0s -> 2579.0s] 难到底, [2579.0s -> 2580.0s] 对于今天的机器人, [2580.0s -> 2581.0s] 它难在了哪些地方, [2581.0s -> 2582.0s] 所以, [2582.0s -> 2583.0s] 我确实, [2583.0s -> 2584.0s] 走了很快, [2584.0s -> 2585.0s] 非常大, [2585.0s -> 2586.0s] 我觉得, [2586.0s -> 2587.0s] 今年一好处, [2587.0s -> 2588.0s] 视觉上, [2588.0s -> 2589.0s] 它的负担, [2589.0s -> 2590.0s] 没有那么大, [2590.0s -> 2591.0s] 虽然, [2591.0s -> 2592.0s] 场馆很大, [2592.0s -> 2593.0s] 走了很多, [2593.0s -> 2594.0s] 但是看到的信息, [2594.0s -> 2595.0s] 其实, [2595.0s -> 2596.0s] 用伤来衡量的话, [2596.0s -> 2597.0s] 其实, [2597.0s -> 2598.0s] 没有那么大, [2598.0s -> 2599.0s] 就是, [2599.0s -> 2600.0s] 很多是, [2600.0s -> 2601.0s] 相似的, [2601.0s -> 2602.0s] 对, [2602.0s -> 2603.0s] 很多是相似, [2603.0s -> 2604.0s] 所以, [2604.0s -> 2605.0s] 没有发现说, [2605.0s -> 2606.0s] 跟, [2606.0s -> 2607.0s] 我们做很, [2607.0s -> 2608.0s] 相关的事情, [2608.0s -> 2609.0s] 至少在, [2609.0s -> 2610.0s] 还是比较, [2610.0s -> 2611.0s] 独特的, [2611.0s -> 2612.0s] 在, [2612.0s -> 2613.0s] 死壳这个, [2613.0s -> 2614.0s] 双方才,

[2614.0s -> 2615.0s] 科学智能这个领域, [2615.0s -> 2616.0s] 这些, [2616.0s -> 2617.0s] 高精度, [2617.0s -> 2618.0s] 高柔性的事情, [2618.0s -> 2619.0s] 这个我稍微, [2619.0s -> 2620.0s] 插了一句, [2620.0s -> 2621.0s] 就是, [2621.0s -> 2622.0s] 从解决方案, [2622.0s -> 2623.0s] 从你们做到今天, [2623.0s -> 2624.0s] 已经, [2624.0s -> 2625.0s] 实现了, [2625.0s -> 2626.0s] 这个精度的结果, [2626.0s -> 2627.0s] 来看, [2627.0s -> 2628.0s] 就是, [2628.0s -> 2629.0s] 我们讲, [2629.0s -> 2630.0s] 智能和硬件, [2630.0s -> 2631.0s] 各自能, [2631.0s -> 2632.0s] 对, [2632.0s -> 2633.0s] 精度的进展到, [2633.0s -> 2634.0s] 这么精确的级别, [2634.0s -> 2635.0s] 各自要, [2635.0s -> 2636.0s] 起头大作用, [2636.0s -> 2637.0s] 我觉得, [2637.0s -> 2638.0s] 硬件上, [2638.0s -> 2639.0s] 使得, [2639.0s -> 2640.0s] 我们有可能, [2640.0s -> 2641.0s] 去解决, [2641.0s -> 2642.0s] 这些问题, [2642.0s -> 2643.0s] 明白, [2643.0s -> 2644.0s] 刚刚成立的时候, [2644.0s -> 2645.0s] 我们也会, [2645.0s -> 2646.0s] 外采一些, [2646.0s -> 2647.0s] 已有的必, [2647.0s -> 2648.0s] 但是, [2648.0s -> 2649.0s] 精度确实, [2649.0s -> 2650.0s] 没有到这个程度, [2650.0s -> 2651.0s] 然后, [2651.0s -> 2652.0s] 我们自言的, [2652.0s -> 2653.0s] 最初期, [2653.0s -> 2654.0s] 也是斗动很厉害, [2654.0s -> 2655.0s] 也是, [2655.0s -> 2656.0s] 倒表这个程度, [2656.0s -> 2657.0s] 但是, [2657.0s -> 2658.0s] 现在确实, [2658.0s -> 2659.0s] 已经能到了, [2659.0s -> 2660.0s] 这, [2660.0s -> 2661.0s] 只有, [2661.0s -> 2662.0s] 到了这个程度, [2662.0s -> 2663.0s] 才能, [2663.0s -> 2664.0s] 有可能, [2664.0s -> 2665.0s] 去解决, [2665.0s -> 2666.0s] 软件, [2666.0s -> 2667.0s] 模型这一块, [2667.0s -> 2668.0s] 去, [2668.0s -> 2669.0s] 不管是, [2669.0s -> 2670.0s] 做一业, [2670.0s -> 2671.0s] 还是操作理性, [2671.0s -> 2672.0s] 这些操作, [2672.0s -> 2673.0s] 所以, [2673.0s -> 2674.0s] 这个是, [2674.0s -> 2675.0s] 一个上限的问题, [2675.0s -> 2676.0s] 我们现在, [2676.0s -> 2677.0s] 比较乐观, [2677.0s -> 2678.0s] 比较好的一个, [2678.0s -> 2679.0s] 健康的一个状态, [2679.0s -> 2680.0s] 是说, [2680.0s -> 2681.0s] 我们下限, [2681.0s -> 2682.0s] 是比较不得, [2682.0s -> 2683.0s] 然后, [2683.0s -> 2684.0s] 我们上限, [2684.0s -> 2685.0s] 也到了, [2685.0s -> 2686.0s] 一个可不设可交付的状态, [2686.0s -> 2687.0s] 当然, [2687.0s -> 2688.0s] 我们会持续的, [2688.0s -> 2689.0s] 进一步的, [2689.0s -> 2690.0s] 去, [2690.0s -> 2691.0s] 提升这个上限, [2691.0s -> 2692.0s] 它所需要的, [2692.0s -> 2693.0s] 对于处理新任务, [2693.0s -> 2694.0s] 所需要的, [2694.0s -> 2695.0s] 要能量, [2695.0s -> 2696.0s] 进一步减少, [2696.0s -> 2697.0s] 在持续提升的地方, [2697.0s -> 2698.0s] 对, [2698.0s -> 2699.0s] 所以, [2699.0s -> 2700.0s] 回到刚才, [2700.0s -> 2701.0s] 你接着讲, [2701.0s -> 2702.0s] 我把它打断了一下, [2702.0s -> 2703.0s] 就是, [2703.0s -> 2704.0s] 在, [2704.0s -> 2705.0s] 涉及到, [2705.0s -> 2706.0s] 克服, [2706.0s -> 2707.0s] 这个过程当中, [2707.0s -> 2708.0s] 以及, [2708.0s -> 2709.0s] 它能做到的, [2709.0s -> 2710.0s] 这个精度, [2710.0s -> 2711.0s] 难度上, [2711.0s -> 2712.0s] 除了精度, [2712.0s -> 2713.0s] 是一个问题之外, [2713.0s -> 2714.0s] 还有别的, [2714.0s -> 2715.0s] 在, [2715.0s -> 2716.0s] 从机器人能力上, [2716.0s -> 2717.0s] 来看, [2717.0s -> 2718.0s] 在过去两年, [2718.0s -> 2719.0s] 需要克服的困难和, [2719.0s -> 2720.0s] 挑战吗? [2720.0s -> 2721.0s] 另外一个, [2721.0s -> 2722.0s] 其实是, [2722.0s -> 2723.0s] 泛化性这部分, [2723.0s -> 2724.0s] 泛化性这部分呢, [2724.0s -> 2725.0s] 还是物体的, [2725.0s -> 2726.0s] 一些物理状态, [2726.0s -> 2727.0s] 这个, [2727.0s -> 2728.0s] 有没有拧紧, [2728.0s -> 2729.0s] 这个, [2729.0s -> 2730.0s] 有没有插紧, [2730.0s -> 2731.0s] 这个, [2731.0s -> 2732.0s] 颜色合同的对齐, [2732.0s -> 2733.0s] 类似于, [2733.0s -> 2734.0s] 这种微小的,

[2734.0s -> 2735.0s] 状态估计, [2735.0s -> 2736.0s] 以及, [2736.0s -> 2737.0s] 对于这些状态, [2737.0s -> 2738.0s] 如何去反馈控制, [2738.0s -> 2739.0s] 这一部分, [2739.0s -> 2740.0s] 也都, [2740.0s -> 2741.0s] 都是要去克服的, [2741.0s -> 2742.0s] 就是, [2742.0s -> 2743.0s] 相当于, [2743.0s -> 2744.0s] 精度解决了, [2744.0s -> 2745.0s] 我们, [2745.0s -> 2746.0s] 知道到哪儿, [2746.0s -> 2747.0s] 一定能到, [2747.0s -> 2748.0s] 这部分呢, [2748.0s -> 2749.0s] 我们就知道, [2749.0s -> 2750.0s] 我们下一步, [2750.0s -> 2751.0s] 应该去什么地方, [2751.0s -> 2752.0s] 所以, [2752.0s -> 2753.0s] 这个, [2754.0s -> 2755.0s] 包括训练的, [2755.0s -> 2756.0s] 架构上, [2756.0s -> 2757.0s] 我们也调整了很多, [2757.0s -> 2758.0s] 这块也是, [2758.0s -> 2759.0s] 克服积累了, [2759.0s -> 2760.0s] 很多能力, [2761.0s -> 2762.0s] 所以说, [2762.0s -> 2763.0s] 回到刚才讲, [2763.0s -> 2764.0s] 就讲那个软的, [2764.0s -> 2765.0s] 就像你刚才讲的, [2765.0s -> 2766.0s] 各, [2766.0s -> 2767.0s] 在过程当中, [2767.0s -> 2768.0s] 有不同的, [2768.0s -> 2769.0s] 阶段性进展, [2769.0s -> 2770.0s] 当然他们可能是, [2770.0s -> 2771.0s] A一下, [2771.0s -> 2772.0s] B一下, [2772.0s -> 2773.0s] A一下, [2773.0s -> 2774.0s] B一下, [2774.0s -> 2775.0s] 从今天来看, [2775.0s -> 2776.0s] 我不知道, [2776.0s -> 2777.0s] 就在应用场景上, [2777.0s -> 2778.0s] 你们已经有了, [2778.0s -> 2779.0s] 应用的证明吗? [2779.0s -> 2780.0s] 比较, [2780.0s -> 2781.0s] 现在是之前, [2781.0s -> 2782.0s] 很早就能跟, [2782.0s -> 2783.0s] 做的合作伙伴, [2783.0s -> 2784.0s] 去推这些事情, [2784.0s -> 2785.0s] 包括, [2785.0s -> 2786.0s] 华大制造, [2786.0s -> 2787.0s] 包括, [2787.0s -> 2788.0s] 专会里面, [2788.0s -> 2789.0s] 也都有, [2789.0s -> 2790.0s] 比较深入的合作, [2790.0s -> 2791.0s] 所以, [2791.0s -> 2792.0s] 我们也能够, [2792.0s -> 2793.0s] 在研发的过程中, [2793.0s -> 2794.0s] 已经把, [2794.0s -> 2795.0s] 这些, [2795.0s -> 2796.0s] 能力, [2796.0s -> 2797.0s] 在真实的, [2797.0s -> 2798.0s] 流程里面去验证, [2798.0s -> 2799.0s] 所以, [2799.0s -> 2800.0s] 这个反馈, [2800.0s -> 2801.0s] 是一个最直接的, [2801.0s -> 2802.0s] 就像一些, [2802.0s -> 2803.0s] 简单的, [2803.0s -> 2804.0s] 细胞盘样, [2804.0s -> 2805.0s] 包括一些这个, [2805.0s -> 2806.0s] 一个, [2806.0s -> 2807.0s] 独性检测, [2807.0s -> 2808.0s] 这些实验, [2808.0s -> 2809.0s] 我们已经是, [2809.0s -> 2810.0s] 能够处理这个, [2810.0s -> 2811.0s] 那已经能在, [2811.0s -> 2812.0s] 在应用问题上, [2812.0s -> 2813.0s] 我们就像, [2813.0s -> 2814.0s] 刚才最早, [2814.0s -> 2815.0s] 问, [2815.0s -> 2816.0s] 外国有人一样, [2816.0s -> 2817.0s] 你觉得在, [2817.0s -> 2818.0s] 今天有多少, [2818.0s -> 2819.0s] 可能的, [2819.0s -> 2820.0s] 场景, [2820.0s -> 2821.0s] 或者是, [2821.0s -> 2822.0s] 具体的, [2822.0s -> 2823.0s] 行业, [2823.0s -> 2824.0s] 或方向, [2824.0s -> 2825.0s] 可能能用得上, [2825.0s -> 2826.0s] 你们了, [2826.0s -> 2827.0s] 乐观的讲, [2827.0s -> 2828.0s] 不是说, [2828.0s -> 2829.0s] 就是不严谨的讲, [2829.0s -> 2830.0s] 都是可以做, [2830.0s -> 2831.0s] 但是, [2831.0s -> 2832.0s] 会付出的, [2832.0s -> 2833.0s] 能力, [2833.0s -> 2834.0s] 付出的, [2834.0s -> 2835.0s] 投入会不一样, [2835.0s -> 2836.0s] 我们现在硬件上, [2836.0s -> 2837.0s] 已经到了, [2837.0s -> 2838.0s] 可以, [2838.0s -> 2839.0s] 支撑, [2839.0s -> 2840.0s] 这些应用的程度了, [2840.0s -> 2841.0s] 举个, [2841.0s -> 2842.0s] 不是这个行业的, [2842.0s -> 2843.0s] 比如说, [2843.0s -> 2844.0s] 家里吃完饭, [2844.0s -> 2845.0s] 你觉得生菜, [2845.0s -> 2846.0s] 还是扔掉, [2846.0s -> 2847.0s] 还是存起来, [2847.0s -> 2848.0s] 假设, [2848.0s -> 2849.0s] 我们要把它存起来, [2849.0s -> 2850.0s] 放冰箱, [2850.0s -> 2851.0s] 要撕保鲜膜, [2851.0s -> 2852.0s] 放上, [2852.0s -> 2853.0s] 撕保鲜膜, [2853.0s -> 2854.0s] 这件事情, [2854.0s -> 2855.0s] 就是一个, [2855.0s -> 2856.0s] 我们接到的需求,

[2856.0s -> 2857.0s] 这个需求, [2857.0s -> 2858.0s] 确实是, [2858.0s -> 2859.0s] 实实在在存在的, [2859.0s -> 2860.0s] 但是, [2860.0s -> 2861.0s] 我们现在的技能, [2861.0s -> 2862.0s] 能做吗, [2862.0s -> 2863.0s] 应做也能做, [2863.0s -> 2864.0s] 但是, [2864.0s -> 2865.0s] 真的, [2865.0s -> 2866.0s] 特别硬着头皮上, [2866.0s -> 2867.0s] 对于, [2867.0s -> 2868.0s] 这个, [2868.0s -> 2869.0s] 实实的决策能力, [2869.0s -> 2870.0s] 哦, [2870.0s -> 2871.0s] 撕保鲜膜, [2871.0s -> 2872.0s] 对于机器人来讲, [2872.0s -> 2873.0s] 可太难了, [2873.0s -> 2874.0s] 因为它从, [2874.0s -> 2875.0s] 它从感知上, [2875.0s -> 2876.0s] 就很难, [2876.0s -> 2877.0s] 就说, [2877.0s -> 2878.0s] 因为它是, [2878.0s -> 2879.0s] 又高透, [2879.0s -> 2880.0s] 又高反, [2880.0s -> 2881.0s] 还薄, [2881.0s -> 2882.0s] 还一丝就破, [2882.0s -> 2883.0s] 就是对力的要求, [2883.0s -> 2884.0s] 还高。 [2884.0s -> 2885.0s] 会有, [2885.0s -> 2886.0s] 类似这样的, [2886.0s -> 2887.0s] 需求, [2887.0s -> 2888.0s] 但是, [2888.0s -> 2889.0s] 真的有, [2889.0s -> 2890.0s] 我们一定也能做, [2890.0s -> 2891.0s] 我们可以做一个, [2891.0s -> 2892.0s] 设备, [2892.0s -> 2893.0s] 特有的一个东西, [2893.0s -> 2894.0s] 来解决这个问题, [2894.0s -> 2895.0s] 而现在呢, [2895.0s -> 2896.0s] 我们还是, [2896.0s -> 2897.0s] 在, [2897.0s -> 2898.0s] 以自己的一个, [2898.0s -> 2899.0s] 一个单点, [2899.0s -> 2900.0s] 我们希望, [2900.0s -> 2901.0s] 慢慢地图宽, [2901.0s -> 2902.0s] 逐渐地有机地, [2902.0s -> 2903.0s] 图宽到哪个地, [2903.0s -> 2904.0s] 所以你最希望, [2904.0s -> 2905.0s] 今天, [2905.0s -> 2906.0s] 可能, [2906.0s -> 2907.0s] 能跟你, [2907.0s -> 2908.0s] 合作上的, [2908.0s -> 2909.0s] 如果是商业化的话, [2909.0s -> 2910.0s] 它大概, [2910.0s -> 2911.0s] 需要一个, [2911.0s -> 2912.0s] 什么样的商业, [2912.0s -> 2913.0s] 环境和条件, [2913.0s -> 2914.0s] 就是, [2914.0s -> 2915.0s] 它把你用哪儿, [2915.0s -> 2916.0s] 它既能用得起, [2916.0s -> 2917.0s] 你也能做得出, [2917.0s -> 2918.0s] 并且对它来讲, [2918.0s -> 2919.0s] 也可用和好用。 [2919.0s -> 2920.0s] 我们现在, [2920.0s -> 2921.0s] 其实提醒, [2921.0s -> 2922.0s] 积极的希望和, [2922.0s -> 2923.0s] 提到的生化环台, [2923.0s -> 2924.0s] 检验检测, [2924.0s -> 2926.0s] 包括Fancy, [2926.0s -> 2927.0s] 材料学科也好, [2927.0s -> 2928.0s] 是, [2928.0s -> 2929.0s] 只要学科也好, [2929.0s -> 2930.0s] 它要持续地, [2930.0s -> 2932.0s] 去避缓这件事情, [2932.0s -> 2933.0s] 就是, [2933.0s -> 2934.0s] 机器人带来好处, [2934.0s -> 2935.0s] 它可以去, [2935.0s -> 2937.0s] 兼容以往的, [2937.0s -> 2938.0s] 所有的, [2938.0s -> 2939.0s] 易购的一切, [2939.0s -> 2940.0s] 老的还是新的, [2940.0s -> 2942.0s] 不需要任何接口, [2942.0s -> 2943.0s] 最好的一个, [2943.0s -> 2944.0s] 不恰当类比, [2944.0s -> 2945.0s] 就是, [2945.0s -> 2946.0s] 人就是一个最好的API, [2946.0s -> 2947.0s] 就是, [2947.0s -> 2948.0s] 人可以调度, [2948.0s -> 2949.0s] 所有的接口, [2949.0s -> 2950.0s] 那我们现在, [2950.0s -> 2951.0s] 智能机器人, [2951.0s -> 2952.0s] 是提供了这么一个, [2952.0s -> 2953.0s] 接口, [2953.0s -> 2954.0s] 所以, [2954.0s -> 2955.0s] 不管, [2955.0s -> 2956.0s] 过去的, [2956.0s -> 2957.0s] 有很多程度, [2957.0s -> 2958.0s] 对它的改造是, [2958.0s -> 2959.0s] 比较, [2959.0s -> 2960.0s] 轻量的, [2960.0s -> 2961.0s] 甚至是零, [2961.0s -> 2962.0s] 零成本去改造, [2962.0s -> 2963.0s] 我们能够带来价值, [2963.0s -> 2964.0s] 一方面是, [2964.0s -> 2965.0s] 能够去, [2965.0s -> 2966.0s] 避缓这个事情, [2966.0s -> 2967.0s] 就是, [2967.0s -> 2968.0s] 避免, [2968.0s -> 2969.0s] 人去引入不确定性, [2969.0s -> 2970.0s] 人引入那些, [2970.0s -> 2971.0s] 效率的问题, [2971.0s -> 2972.0s] 同时, [2972.0s -> 2973.0s] 更好, [2973.0s -> 2974.0s] 更大的, [2974.0s -> 2975.0s] 带来的价值, [2975.0s -> 2976.0s] 是说, [2976.0s -> 2977.0s] 机器人一旦, [2977.0s -> 2978.0s] 开始去, [2978.0s -> 2979.0s] 避缓, [2979.0s -> 2980.0s] 那这些数据,

[2980.0s -> 2981.0s] 自然是, [2981.0s -> 2982.0s] 标准的话, [2982.0s -> 2983.0s] 可信赖的, [2983.0s -> 2984.0s] 数据, [2984.0s -> 2985.0s] 其实, [2985.0s -> 2986.0s] 比基座, [2986.0s -> 2987.0s] 类比于, [2987.0s -> 2988.0s] 之前我们做, [2988.0s -> 2989.0s] AI做计算, [2989.0s -> 2990.0s] GPU, [2990.0s -> 2991.0s] 它是一个, [2991.0s -> 2992.0s] 计算平台, [2992.0s -> 2993.0s] 有了它, [2993.0s -> 2994.0s] 可能不光是说, [2994.0s -> 2995.0s] 我解决这些计算, [2995.0s -> 2996.0s] 我比以前快多少倍, [2996.0s -> 2997.0s] 而说, [2997.0s -> 2998.0s] 有了GPU以后, [2998.0s -> 2999.0s] 我可以, [2999.0s -> 3001.0s] 做以前不该想的事情, [3001.0s -> 3002.0s] 我们现在可以, [3002.0s -> 3003.0s] 大规模的去处理, [3003.0s -> 3004.0s] 图像的视频, [3004.0s -> 3005.0s] 现在处理语言, [3005.0s -> 3006.0s] 处理更多的, [3006.0s -> 3007.0s] 又可以, [3007.0s -> 3009.0s] 解锁一些新的, [3009.0s -> 3010.0s] 方式, [3010.0s -> 3011.0s] 那, [3011.0s -> 3012.0s] 机器人引入, [3012.0s -> 3013.0s] 这个, [3013.0s -> 3014.0s] 办实验室的环境, [3014.0s -> 3015.0s] 能实话还餐, [3015.0s -> 3016.0s] 领域之后, [3016.0s -> 3018.0s] 它不光是说, [3018.0s -> 3019.0s] 我去, [3019.0s -> 3020.0s] 避免了人来操作, [3020.0s -> 3021.0s] 而说, [3021.0s -> 3022.0s] 它, [3022.0s -> 3023.0s] 由于它带来的, [3023.0s -> 3024.0s] 环境数据化, [3024.0s -> 3025.0s] 由于它带来的, [3025.0s -> 3026.0s] 标准化, [3026.0s -> 3027.0s] 和可负线性, [3027.0s -> 3028.0s] 它可以解锁, [3028.0s -> 3029.0s] 过去, [3029.0s -> 3030.0s] 我们不敢, [3030.0s -> 3031.0s] 设计那些实验, [3031.0s -> 3032.0s] 我觉得, [3032.0s -> 3033.0s] 这个是带来的, [3033.0s -> 3034.0s] 很大的一个优势, [3034.0s -> 3035.0s] 所以, [3035.0s -> 3036.0s] 之前提了, [3036.0s -> 3037.0s] 很模糊地提到, [3037.0s -> 3038.0s] 我们可以,