网站地图 - XML地图 - 设为首页 - 加入收藏
您的当前位置:主页 > 国内 > 正文

???ڇ?;

不做硅谷follower:几个读博的年轻人,押注双足人形的一体化大脑_我的网站

我的前半生

一 |     

GOP front-runner and four times-indicted Donald Trump will not be physically present at Wednesday evening's first 2024 presidential primary debate. But he'll be there anyway as the troubled candidate remains the dominant figure – and issue – of the nominating contest.
Typically, early debates are a chance for poll-laggers to make an impression, either by delivering a well-placed zinger at a top-tier candidate or shining with a smart response to a question. The front-runners tend to have targets on their backs, which could result in a weakened position or a stronger one, if they handle attacks from primary rivals well.
But in the case of the debate in Milwaukee on Wednesday, the participants are in an awkward, perhaps no-win situation: The front-runner won't even be there, and the rest of the field can't attack him without alienating the very voters they'd need to sew up the nomination, should the legally imperiled Trump be forced to suspend his campaign.
Trump is "going to be the huge, spray-tanned elephant in the room. He's going to be what everybody is thinking about, even though he's not there," says Dave Cohen, a political science professor at the University of Akron.
That means the eight candidates who have qualified for the debate – Florida Gov. Ron DeSantis, tech entrepreneur Vivek Ramaswamy, Sen. Tim Scott of South Carolina, former U.N. Ambassador Nikki Haley, former Vice President Mike Pence, North Dakota Gov. Doug Burgum, former New Jersey Gov. Chris Christie and former Arkansas Gov. Asa Hutchinson – have few choices, and none is terribly appealing.
They can attack Trump, defend him or try to ignore him – which debate moderators are not likely to let pass, Cohen says. None of those approaches helps a candidate emerge as a better choice than Trump without angering Trump's rank-and-file loyalists.
"They're in a difficult spot, strategically," says Stephen Smith, professor at the School of Politics and Global Studies at Arizona State University. But "it's one they've earned," he adds.
Only a few have been willing to criticize Trump directly, and it's not appeared to have endeared them to GOP primary voters.
Christie – who worked on Trump's transition team before turning on him – has been the most aggressive in his attacks on Trump, saying the former president lacks the "character" to serve in the Oval Office and needling Trump for skipping the debate, which Christie last week called "completely disrespectful" to GOP voters.
The tactic – essentially making himself the marquee foil of Trump – is based on the gamble that if Trump ends up leaving the race, Christie will emerge as the most prominent truth-teller and default new front-runner.
It's not going so well, so far: National polls have Christie in the low single digits. A recent Des Moines Register/NBC poll in Iowa puts Christie in sixth place with 5% support among GOP caucus-goers.
Former Arkansas Gov. Asa Hutchinson – who arguably on paper has the most complete and traditional resume for a Republican presidential contender – is also polling in the low single digits, his criticism of Trump not moving the needle at all.
Former Vice President Mike Pence has stepped up his criticism of Trump's behavior – saying the former president asked him to put him above the Constitution by asking him to refuse to certify the 2020 election results – but hasn't made it as personal.
Pence recently said he was not aware of any "broad-based effort" to declassify records before Trump left office, undercutting the former president's claim that the classified documents he is accused of taking and improperly keeping at his Mar-a-Lago resort were, in fact, declassified. But Pence has also been reluctant to say Trump should be treated like any other politician who has been convicted of a felony – meaning, ousted from office.
"I think that needs to be left to the American people. Look, let’s let the former president have his day in court. Let’s maintain a presumption of innocence in this matter," Pence told ABC on Sunday.
Former U.S. Ambassador Nikki Haley – the only woman who will be onstage Wednesday night – has cautioned that Trump's legal troubles could cost Republicans the general election. She hasn't directly criticized the former president but didn't shy away earlier this month by reminding people of how much worse Trump's situation is getting.
"Like most Americans, I’m tired of commenting on every Trump drama. I’ve lost track of whether this indictment is the third or fourth or the fifth," Haley said during an appearance on “Good Morning New Hampshire” radio show with Jack Heath – before the fourth indictment out of Georgia’s Fulton County.
Others have responded to questions about Trump with attacks on the criminal justice system and its prosecutors themselves.
Florida Gov. Ron DeSantis, who is running a distant second to Trump in most polls, decried the "weaponization of government" after Trump's third indictment – though he conceded he had not actually read the document. DeSantis also acknowledged recently that the 2020 election was not stolen, telling NBC "of course he lost" when asked about Trump's discredited claims of a fraudulent 2020 contest – a comment not likely to endear him to Trump's most devoted supporters.
Sen. Tim Scott of South Carolina called the recent indictment of Trump "un-American and unacceptable. We need a better system than that."
Tech entrepreneur Vivek Ramaswamy, who has edged upward in a couple of state polls, has walked a similar line.
"These are politicized persecutions through prosecution,” he said at a town hall meeting last week. “It would be a lot easier for me if Donald Trump were not in this primary, but that is not how I want to win this election," he added.
And North Dakota Gov. Doug Burgum is just deflecting the matter entirely.
"I do trust the voters. ... They want presidential elections to be about the future, a vision for the future," Burgum told NBC's "Meet the Press" on Sunday after being asked about Trump's troubles.
In reality, it is the candidates not named Trump who will have the most to prove or explain Wednesday night, analysts say. DeSantis, whose campaign has been flailing, might need to explain why he suggested some Republicans backing Trump are "listless vessels" – a cringe-worthy similar remark to 2016 Democratic presidential nominee Hillary Clinton reference to the "basket of deplorables" in Trump's base.
As the front-runner among the also-rans, DeSantis will likely take the heaviest rhetorical artillery from the GOP field at the debate. Most of the rest will be struggling to get out of single digits – with the possible exception of Burgum, who is still trying to get the American public to recognize his name.
Trump, meanwhile, said he plans to counter-program with the airing of a taped one-on-one interview with former Fox News host Tucker Carlson. Thursday – a day when the political news would typically be dominated by post-debate analysis – Trump will be the star player in an unprecedented legal and political drama, turning himself in for processing at a Fulton County jail.
"He's cutting right into the shadow of the debate" with that timing, Smith says. As he undoubtedly will during the debate itself.
。    
乔不迟 发自 凹非寺量子位 | 公众号 QbitAI周末在家刷社交平台,一个视频惊艳到我了。一个机器人驾驶卡丁车在赛道上飞驰,过弯、加速、打方向,一镜到底。没想到我们的机器人现在已经可以在卡丁车赛道炫技了呢!

等等?这和我在WRC看到的画风不一样啊!展会上的机器人跑步、还有火花加闪电,没想到视频里的机器人什么时候背着人类偷偷报名赛车培训班了?它完全是个老司机,高速过弯、灵活避障,跑一圈下来一气呵成。和以往那些机器人慢吞吞走过去、弯腰拿东西的demo相比,这段视频里的机器人真的自己坐进了车里,手眼脚同步协同,在全速状态下完成多接触点平衡和精细力控。这种人形双足机器人能够眼、手、足,全身协同,在高速运动、复杂姿态的变化下的全身协调完成细节操作,在此前的人形机器人演示里几乎没出现过。放出这段视频的,是一家叫共生知行的公司。共生知行做的,是双足人形机器人的端到端感控一体化「大脑」,也就是基座模型。创始人丁鹏翔96年出生,核心团队其余成员全是00后,全员博士在读,公司成立刚满两个月。

二 | 对于这个看上去有点「不务正业」的Demo,丁鹏翔这样解释:我希望机器人能更像人,而不是只去做一些很机械的服务功能。好比人能开卡丁车,机器人也能干。一个问题随之而来:为什么一群连毕业证都还没拿到的年轻人,敢去碰具身智能里最难、最前沿的那条赛道?

押最难的路:研究过分层,现在要挑战端到端「端到端」双足人形机器人模型这个赛道新到什么程度?新到市面上找不到现成的人才。按丁鹏翔的说法,这个方向近两年才热起来,懂技术的人还在读博士,他们几乎没办法从人才库里找到符合需求的人。已毕业多年的跟不上节奏,没读到博士的,对底层的理解又差一口气。

三 | 最前沿的突破,就握在这一批顶尖博士生手里。共生知行的团队就是这样一批年轻又前沿的人。

四 | 他们斩获了国内具身智能领域最具分量的学术荣誉——两项 Best Paper,并有一项入选 Best Paper Candidate;也打造了国内首个 GitHub Stars 突破 2K 的具身基座模型。更重要的是,他们并非风口到来后的追随者,而是国内最早一批探索具身基座模型的先行者,也是这个赛道最坚定的长期主义者。团队累计发表40余篇顶会论文,技术版图覆盖感知、决策、控制、数据与系统全栈;从双系统、轻量化到多构型,一系列行业首创工作,都出自这支年轻团队。

他们的判断很明确:机器人最终最广泛进入人类生活的形态是双足;按照自动驾驶和大模型的进化路径,端到端是效率最高的技术模式。这个判断源自他们的一线研究经验。

五 | 早在2023年12月,丁鹏翔就开始给四足机器人装「大脑」。他以一作身份发布QUAR-VLA,是首个面向四足机器人的VLA任务范式,让上层模型(大脑)负责理解视觉、 语言和任务意图,再把决策交给底层运动系统(小脑)执行。这套「大脑管理解决策、小脑管运动执行」的分层思路,就是行业里说的分层,今天 Figure这样的公司都在沿用方案。在随后的研究中,他逐渐发现了问题。他把分层描述为「理论上可行、但是不那么优雅的方案」:大脑和小脑各做各的,每接一个新任务,都要对上游输出做定制化微调,才能把效果调好。更深一层,是结构上的先天缺陷。大脑和小脑分开训练,部署时再拼起来,每个局部更好,也不能保证整体最优;两层之间每翻译一次,就增加一次级联误差和信息损失。

六 | 丁鹏翔判断,「只要分层,中间就有信息瓶颈,这种接口设计决定了分层架构无法实现真正意义上的Scaling。」这一判断也来自他对自动驾驶和大模型发展的长期观察。早期自动驾驶依赖感知、预测、规划、控制多个模块协同,后来开始探索让模型直接从数据中学习驾驶能力。

特斯拉FSD转向端到端后,性能大幅提升。

七 | 在丁鹏翔看来,机器人可能也会经历类似变化,让模型从数据中学习完整的「感知—理解—行动」过程。他的逻辑是,从2023年开始已经把分层路线反复做过,如果判断未来的数据规模化最终会把系统推向端到端,就没有必要再绕一次弯。

八 | 这一步在2026年正在发生。谷歌的Gemini Robotics 2已经用单一策略把从脚到指尖的全身动作统一进一个模型;RSS 2026的顶会共识却写明,端到端单模型无法覆盖全身复杂动态,分层模块化是现阶段最优落地方案。共生知行押的纯端到端,是一条更激进、更少人走的路。

为什么偏偏是双足?丁鹏翔的答案是第一性原理:人类社会的建筑、工具、环境,全是照着人的身体构型造的。轮式只能固定在产线和商场里干活,双足能出门、上下楼、开车,还能跨场景完成不同任务。通用性意味着成本摊薄,双足还有拟人化的亲和感。时机同样关键。2026年4月之前,双足人形连通用遥操作模型都没有,没有数据来源就训不出基座。直到英伟达Sonic开源,行业才有了大规模生产数据的基石。这个方向已经不再处于「想训也没有数据」的早期阶段,又还没有成熟到路线完全收敛,共生知行赌的就是中间这个窗口。

技术分水岭:别人在学运动学,他们在啃动力学先弄清一个分水岭:现在绝大多数机器人基座模型,学的其实是运动学。运动学关心从A点到B点,手移动到哪里、抓没抓住,固定基座机械臂这么干没问题,因为它不会倒。

九 | 双足机器人伸手去够一个东西,身体重心会同时发生变化;下蹲拿物体时,腰部要前倾,踝关节和腿部要重新分配力度、保持平衡;再叠加摩擦、碰撞、惯性和接触,模型面对的已经是一套全身动力学。据丁鹏翔介绍,一台g1双足人形有29个自由度,比机械臂的7个自由度复杂得多。

十 | 过去机器人学的是「动作轨迹」,人形机器人要学「身体如何在物理世界中行动」

从运动学到动力学,从完成任务到保持稳定,这是双足机器人模型真正的技术分水岭。这能解释共生知行为什么把模型一路做到Joint Target,也就是关节目标层。分层方案里,大脑先输出运动学目标,下游小脑再解算成关节动作;共生知行把这个中间环节拿掉,让模型直接面对几十个关节的身体状态。问题随之而来:把小脑拿掉,机器人还站得稳吗?

光靠模仿学习,模型只学会标准动作,没见过执行中各种踉跄的身体状态,遇到陌生状态就可能直接摔倒。它缺的是从失败里站稳的能力。共生知行的核心,是一套「任务行为建模—运动先验蒸馏」的双域协同优化机制。

十一 | 一条优化通路延续行为克隆,保证任务精度与动作拟合能力;另一条通过DriftDistill,将底层控制器积累的稳定性、抗扰动与Failure Recovery能力,转化为统一模型的内生运动先验。它并非简单叠加两个Loss,而是将「准确完成任务」与「稳定控制身体」两类能力融入同一模型,让大模型同时获得任务智能与身体智能。如果DriftDistill能随模型和数据量继续扩展,它试图解决的就是一个更大的问题:在认知Scaling之外,机器人能不能连运动控制能力也一起Scaling。

十二 | 丁鹏翔将这一运动能力汇聚模块称为Motion Expert(运动专家模型),目前规模接近1B,并通过蒸馏不断吸收不同运控模型的能力。

再往前一步,端到端大模型还要解决机器人输出的力。

十三 | 今天很多机器人系统主要控制位置,但机器人真正进入现实世界后,位置正确并不等于任务完成。手伸到抽屉把手上,却不知道该用多大力,门依然打不开;给人递东西时位置没问题,力量过大同样不安全。共生知行的技术路线是先把Force Expert作为安全接触专家,让它学习施力、柔顺、阻抗和接触反馈,再通过策略蒸馏逐步融入Motion Expert和最终的端到端模型。位置决定机器人能不能「到那里」,力决定它能不能真正「把事情做好」。这也是共生「全身物理世界模型」的含义:让模型理解身体怎样受力、失衡、接触,再把这些约束直接变成动作。

数据、涌现,和一张冷静的时间表模型真正训练起来,首先撞上的就是数据荒。全球合规机器人数据截至2026年初大约50万小时,不足大语言模型的两万分之一。

十四 | 尴尬的是,很多数据采的是站定操作,机器人在格子间里原地干活,对全身移动操作帮助有限。共生知行的Blog放出了约1万多小时数据,自采其中几千小时。全身协调模型真正稀缺的是真实生活里的强耦合全身移动操作数据,比如骑单车、开卡丁车、给篮球打气等。

十五 | 他们的解法是TrajBooster:从机械臂或轮式机器人身上提取末端轨迹,让人形机器人在仿真里追踪这条轨迹并保持平衡,把不同来源的动作统一到同一个空间里,复用机械臂已经定义好的任务多样性,低成本给预训练提供数据。谈到模型能力,丁鹏翔几乎不用「泛化」这个词。用他的话说,「这个词没什么营养,像说一个人好一样空。」他要的是「涌现」:在关节角层面,比如给模型训练的是走和跳的数据,它在测试的时候自己组合出跑这个新的技能。因为只有到关节层面,部分动作能力才可能组合出全新动作。

时间表上,他的判断很冷静。他推断,十万小时左右数据更多还是实验室级Demo,真正有大规模商业意义,可能要等到百万小时之后。现在行业甚至连最合适的数据形式都没有收敛,肌电、外骨骼、动捕、Pico遥操等路线仍在竞争。

十六 | 在他看来,人形机器人当前最难的三个问题,第一是感知和控制究竟应该怎样Combine,第二是应该用什么类型的数据,第三才是数据规模本身。丁鹏翔希望,有一天由中国团队设计出来的机器人模型架构,也能反过来被北美同行广泛采用。我们不希望当Follower,我们希望去引领技术的发展。回到那辆卡丁车。让它全速过弯的,是一群还没毕业的年轻人。他们押的,是具身智能里最难、最前沿的方向。赌的是端到端终将取代分层。行业还远没有到可以证明终局路线已经跑通的时候,甚至丁鹏翔自己也承认,整个双足基础模型行业都还没有真正收敛。但一支创业团队的前沿性,有时候恰恰不在于把一个成熟答案做得更快,而在于它能不能比同行更早看见下一道真正重要的问题,并且敢在答案还不明朗时动手去解。

Current article:http://6zr.saorongriwazhaofutouzhaojie.cfd/uzudk/20260826/327655.html

Published on:07:42:57


注:凡本网注明来源非本站的作品,均转载自其它媒体,并不代表本网赞同其观点和对其真实性负责。
本站致力于帮助文章传播,希望能够建立合作关系。
若有任何不适的联系以下方式我们将会在24小时内删除。联系方式:
Copyright © 2018 我的网站 版权所有