• 亿邦公众号
    • APP下载

北大数院戈登贝尔奖得主下场:破解物理AI“准与快”死结

作者:簪竹 文章来源:投中网

2026-11-03 12:02

作者:簪竹 文章来源:投中网

2026-11-03 12:02

邦小白快读

EN
全文速览

总:物理AI落地最核心的判断是“准”和“快”两道硬门槛,文章给出了从底层技术到商业落地的破局路径。

1.落地难根源:虚拟环境训练完美,一上真机就失效;依赖云端算力会遇到延迟与成本;物理AI的误差会变成物料损耗、产线停摆甚至安全风险。

2.技术方案:壹零壹壹由北大数院教授、首位戈登贝尔奖得主杨超创办,打造高精度物理仿真加全栈高性能计算双根技术;仿真引擎谐物做到资产真、引擎真、轨迹真,覆盖刚体、柔体、布料线缆、液体、颗粒等数千个可复用物理资产;端侧推理框架扶摇降低硬件依赖。

3.实操启示:用仿真低成本扩展样本空间、生成最优轨迹,再用真机数据修正仿真参数;端侧跨硬件适配可降低部署成本、匹配作业节拍。公司2026年6月成立后短期完成过亿元融资,并已在零售等场景落地。

总:对品牌商来说,文章的价值在产品研发逻辑、成本定价竞争力和渠道生态信号。

1.产品研发:物理AI不能只追demo,要从底层仿真和计算重构;谐物引擎把材料参数、物性属性和场景边界条件做进物理资产,轨迹满足动力学约束与执行器极限,提升方案向真机迁移的可靠性。

2.成本与价格竞争:端侧推理框架扶摇不堆高端硬件也能跑实时,机器人企业不必盲目升级硬件,降低部署成本;高保真仿真降低真机数据采集和调试成本,缩短新场景新物体适配周期。

3.渠道与生态:上游与头部大模型厂商做仿真数据合作,参与具身智能评测榜单仿真环境共建,进入多家云厂商仿真引擎供应商清单,并与头部芯片厂商联合适配;下游零售客户深度合作,诉求是部署周期短、综合成本可控、匹配真实作业节拍。

4.趋势与品牌信任:投资逻辑从场景故事转向底层基础设施长期价值,品牌应关注可衡量提质增效,用产业落地验证技术价值,而非依赖前沿概念放大估值。

总:卖家可从物理AI落地难点中识别风险、机会与合作入口。

1.风险提示:仿真不准会造成Sim2Real鸿沟;真机数据成本高、周期长,危险场景难复现且质量受操作人员水平限制;云端算力有传输延迟和网络波动,端侧算力不足会导致模型跑不动、时延压不住;工业节拍和零售成本约束不容等待。

2.机会提示:物理AI规模化应用空间打开,制造、能源、交通、航空航天等被寄望创造可衡量提质增效价值;上游仿真数据合作、评测榜单仿真环境共建、云厂商供应商清单、芯片联合适配、下游零售深度合作都是可关注合作方式。

3.可学习点:采用虚实融合,仿真低成本扩展样本空间并生成超越人类操作习惯的最优轨迹,真机数据修正仿真参数;不追短期场景故事,转向底层基础设施长期价值,构建复用性强的通用底座。

4.商业模式:双根技术基础设施赋能全场景,靠产业落地验证技术价值;政策与产业需求共同推动规模化应用。

总:工厂读者应关注物理AI对产线节拍、部署成本和数字化改造的直接影响。

1.生产需求:真实场景的物体滑落、接触碰撞、产线流转发生在毫秒级,核心控制闭环必须在本地完成;端侧推理框架扶摇可跨架构、跨算力层级适配,不依赖云端也能跑出满足工业节拍的实时性能。

2.设计与仿真:谐物引擎覆盖刚体、铰接结构、柔体、布料线缆、液体、颗粒等数千个可复用物理资产,并附材料参数、物性属性和场景边界条件;机械臂抓取柔性物件、灵巧手操作海绵、物体在液体中浮沉等可高置信度模拟。

3.商业机会与数字化启示:高保真仿真降低真机数据采集和方案调试成本,缩短新场景新物体适配周期;用仿真扩展样本、生成最优轨迹,再用真机数据修正参数,可减少物料损耗、产线停摆和安全风险;机器人企业不必盲目升级硬件即可满足节拍,降低部署成本。

总:服务商可从行业趋势、客户痛点与可交付解决方案三方面提炼干货。

1.趋势:物理AI从实验室demo走向真实生产力,竞争重心从场景故事转向底层基础设施;仿真与计算被投资人视为通用底座,不依赖单一赛道却能赋能全场景,具备复用性和产业穿透力。

2.客户痛点:仿真不准,图形学引擎重画面轻物理;工业仿真准但慢,单次复杂仿真动辄耗时数天;真机数据贵、周期长、危险场景难复现;云端部署有延迟和网络波动,端侧算力不足、成本高。

3.解决方案:谐物物理仿真引擎以资产真、引擎真、轨迹真控制误差,并用自适应建模、AI加速数值方法、隐式求解和大规模并行计算实现少算、快算、大步算、并行算;扶摇端侧推理框架做跨架构、跨算力层级适配;虚实融合形成仿真扩展样本与真机校准的正向循环。

4.生态合作:可与头部大模型厂商、云厂商、芯片厂商及零售客户合作。

总:平台商可关注物理AI平台对仿真环境、端侧推理和供应链适配的需求,以及文章给出的生态做法。

1.平台需求与问题:物理AI落地需要高保真仿真环境支撑评测和训练,也需要端侧推理保障现场作业效率与稳定性;客户普遍要求部署周期短、综合成本可控、匹配真实作业节拍,云端延迟和网络波动是不可控风险。

2.最新做法:壹零壹壹与头部大模型厂商开展仿真数据合作,获邀参与具身智能评测榜单的仿真环境共建,进入多家云厂商的仿真引擎供应商清单;端侧推理框架与国内头部芯片厂商联合适配,优化边缘算力下运行性能;下游与零售客户深度合作。

3.运营与风向规避:平台应接入具备双根技术的供应商,避免只堆场景故事和概念估值;投资逻辑正转向底层基础设施长期价值。仿真与计算是通用底座,可赋能全场景,具备复用性与产业穿透力。平台招商可围绕仿真引擎、端侧推理、芯片适配和评测环境共建展开。

总:研究者可从产业新动向、新问题、商业模式和政策启示四个层面提取。

1.新动向:物理AI与数字AI存在本质分野,误差会转化为真实物料损耗、产线停摆和安全风险;行业公认死结是仿得不准、算得太贵,需要从算法、算力、场景的系统性失配入手重构技术范式。壹零壹壹以高精度物理仿真和全栈高性能计算双根技术切入。

2.新问题:真机数据成本高、周期长、危险场景难复现且质量受操作人员水平限制;图形学仿真不严格遵循物理定律造成Sim2Real鸿沟,工业仿真精度足够但效率低数个数量级;云端集中算力有延迟和网络波动,端侧算力不足。

3.商业模式与启示:谐物以资产真、引擎真、轨迹真做高保真仿真,扶摇做端侧推理;虚实融合被视为根本性终局。投资逻辑从短期场景故事转向底层基础设施长期价值。政策与产业需求共同推动规模化应用,制造、能源、交通、航空航天等是提质增效方向。硬科技创业是马拉松,底层技术厚度决定产业走多远。

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

Overall: The core test for physical AI deployment is two hard thresholds—accuracy and speed. The article outlines a path from underlying technology to commercial implementation.

1. Root causes of deployment difficulty: Training in virtual environments can be perfect, but systems fail once deployed on real machines; relying on cloud computing introduces latency and cost; errors in physical AI translate into material loss, production line downtime, or even safety risks.

2. Technology solution: 1011 was founded by Yang Chao, a professor at Peking University's School of Mathematical Sciences and the first Gordon Bell Prize winner, and builds dual-root technologies: high-precision physical simulation and full-stack high-performance computing. Its Xiewu simulation engine achieves asset fidelity, engine fidelity, and trajectory fidelity, covering thousands of reusable physical assets such as rigid bodies, soft bodies, cloth and cables, liquids, and granular materials; its Fuyao edge inference framework reduces hardware dependence.

3. Practical implications: Use simulation to expand the sample space at low cost and generate optimal trajectories, then use real-machine data to calibrate simulation parameters; cross-hardware adaptation at the edge can lower deployment costs and match operating cadence. Since its founding in June 2026, the company has completed more than RMB 100 million in financing in a short period and has already deployed in retail and other scenarios.

Overall: For brands, the article's value lies in product R&D logic, cost and pricing competitiveness, and channel ecosystem signals.

1. Product R&D: Physical AI should not chase demos alone; it must be rebuilt from underlying simulation and computing. The Xiewu engine incorporates material parameters, physical properties, and scenario boundary conditions into physical assets, and its trajectories satisfy dynamic constraints and actuator limits, improving the reliability of transferring solutions to real machines.

2. Cost and price competition: The Fuyao edge inference framework can run in real time without stacking high-end hardware, so robotics companies need not blindly upgrade hardware, lowering deployment costs; high-fidelity simulation reduces the cost of real-machine data collection and debugging and shortens the adaptation cycle for new scenarios and objects.

3. Channel and ecosystem: Upstream, it partners with leading large-model vendors on simulation data, participates in co-building simulation environments for embodied intelligence evaluation benchmarks, enters the simulation engine supplier lists of multiple cloud vendors, and jointly adapts with leading chip vendors; downstream, it works deeply with retail customers whose requirements are short deployment cycles, controllable total costs, and alignment with real operating cadence.

4. Trends and brand trust: Investment logic is shifting from scenario stories to the long-term value of underlying infrastructure. Brands should focus on measurable quality and efficiency improvements and use industrial deployment to validate technological value, rather than relying on frontier concepts to inflate valuations.

Overall: Sellers can identify risks, opportunities, and entry points for cooperation from the deployment challenges of physical AI.

1. Risk alerts: Inaccurate simulation creates a Sim2Real gap; real-machine data is costly and slow to collect, dangerous scenarios are hard to reproduce, and data quality is limited by operator skill; cloud computing has transmission latency and network fluctuations, while insufficient edge computing power can make models unable to run and fail to contain latency; industrial cadence and retail cost constraints cannot wait.

2. Opportunity alerts: The space for large-scale application of physical AI is opening up. Manufacturing, energy, transportation, and aerospace are expected to create measurable quality and efficiency value. Upstream simulation data cooperation, co-building simulation environments for evaluation benchmarks, cloud vendor supplier lists, joint chip adaptation, and deep downstream retail cooperation are all cooperation models worth watching.

3. Takeaways: Adopt virtual-real integration: simulation expands the sample space at low cost and generates optimal trajectories beyond human operating habits, while real-machine data calibrates simulation parameters; do not chase short-term scenario stories, but shift toward the long-term value of underlying infrastructure and build a highly reusable general foundation.

4. Business model: Dual-root technology infrastructure empowers all scenarios and validates technological value through industrial deployment; policy and industrial demand jointly drive large-scale application.

Overall: Factory readers should focus on the direct impact of physical AI on production line cadence, deployment costs, and digital transformation.

1. Production requirements: In real scenarios, object slippage, contact collisions, and production line flow occur at the millisecond level, and the core control loop must be completed locally; the Fuyao edge inference framework can adapt across architectures and computing power tiers and deliver real-time performance that meets industrial cadence without relying on the cloud.

2. Design and simulation: The Xiewu engine covers thousands of reusable physical assets such as rigid bodies, articulated structures, soft bodies, cloth and cables, liquids, and granular materials, with material parameters, physical properties, and scenario boundary conditions attached; it can simulate with high confidence tasks such as robotic arms grasping flexible objects, dexterous hands handling sponges, and objects floating and sinking in liquids.

3. Business opportunities and digitalization implications: High-fidelity simulation reduces the cost of real-machine data collection and solution debugging and shortens the adaptation cycle for new scenarios and objects; using simulation to expand samples and generate optimal trajectories, then using real-machine data to calibrate parameters, can reduce material loss, production line downtime, and safety risks; robotics companies can meet cadence without blindly upgrading hardware, lowering deployment costs.

Overall: Service providers can extract key points from three aspects: industry trends, customer pain points, and deliverable solutions.

1. Trends: Physical AI is moving from lab demos to real productivity, and the focus of competition is shifting from scenario stories to underlying infrastructure; investors regard simulation and computing as a general foundation that does not depend on a single track yet can empower all scenarios, with strong reusability and industrial penetration.

2. Customer pain points: Simulation is inaccurate, and graphics engines prioritize visuals over physics; industrial simulation is accurate but slow, with a single complex simulation often taking days; real-machine data is expensive, slow to collect, and dangerous scenarios are hard to reproduce; cloud deployment has latency and network fluctuations, while edge computing power is insufficient and costly.

3. Solutions: The Xiewu physical simulation engine controls errors through asset fidelity, engine fidelity, and trajectory fidelity, and uses adaptive modeling, AI-accelerated numerical methods, implicit solving, and large-scale parallel computing to achieve less computation, faster computation, larger steps, and parallel computation; the Fuyao edge inference framework adapts across architectures and computing power tiers; virtual-real integration creates a positive cycle of simulation-expanded samples and real-machine calibration.

4. Ecosystem cooperation: It can cooperate with leading large-model vendors, cloud vendors, chip vendors, and retail customers.

Overall: Marketplace platforms can focus on the demands that physical AI platforms place on simulation environments, edge inference, and supply chain adaptation, as well as the ecosystem approaches described in the article.

1. Platform needs and problems: Physical AI deployment requires high-fidelity simulation environments to support evaluation and training, and edge inference to ensure on-site operational efficiency and stability; customers generally require short deployment cycles, controllable total costs, and alignment with real operating cadence, while cloud latency and network fluctuations are uncontrollable risks.

2. Latest practices: 1011 conducts simulation data cooperation with leading large-model vendors, has been invited to participate in co-building simulation environments for embodied intelligence evaluation benchmarks, and has entered the simulation engine supplier lists of multiple cloud vendors; its edge inference framework is jointly adapted with leading domestic chip vendors to optimize performance under edge computing power; downstream, it works deeply with retail customers.

3. Operations and trend risk avoidance: Platforms should onboard suppliers with dual-root technologies and avoid merely stacking scenario stories and concept valuations; investment logic is shifting toward the long-term value of underlying infrastructure. Simulation and computing are a general foundation that can empower all scenarios and offer reusability and industrial penetration. Platform recruitment can focus on simulation engines, edge inference, chip adaptation, and co-building evaluation environments.

Overall: Researchers can extract insights at four levels: new industry directions, new problems, business models, and policy implications.

1. New directions: Physical AI is fundamentally different from digital AI, as errors translate into real material loss, production line downtime, and safety risks; the industry-recognized deadlock is inaccurate simulation and expensive computation, requiring a reconstruction of the technology paradigm starting from the systemic mismatch among algorithms, computing power, and scenarios. 1011 enters through dual-root technologies: high-precision physical simulation and full-stack high-performance computing.

2. New problems: Real-machine data is costly, slow to collect, difficult to reproduce in dangerous scenarios, and limited in quality by operator skill; graphics-based simulation does not strictly follow physical laws, creating a Sim2Real gap, while industrial simulation is accurate enough but several orders of magnitude less efficient; centralized cloud computing has latency and network fluctuations, and edge computing power is insufficient.

3. Business model and implications: Xiewu delivers high-fidelity simulation through asset fidelity, engine fidelity, and trajectory fidelity, while Fuyao handles edge inference; virtual-real integration is regarded as the fundamental endgame. Investment logic is shifting from short-term scenario stories to the long-term value of underlying infrastructure. Policy and industrial demand jointly drive large-scale application, with manufacturing, energy, transportation, and aerospace as directions for quality and efficiency improvement. Hard-tech entrepreneurship is a marathon, and the depth of underlying technology determines how far the industry can go.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

精准击穿了行业长期悬而未决的 “准” 与 “快” 两大核心痛点。

当具身智能与物理AI的产业热度持续攀升,落地难却成了全行业无法回避的尴尬:虚拟环境里训练完美的策略,一上真机就频频失效;依赖云端算力的部署方案,到了真实应用就被延迟与成本击碎……“仿得不准、算得太贵”成为物理AI从实验室走向现实、从demo演示走向真实生产力的最大死结,把物理AI的落地路径和规模化商用卡在了“最后一公里”。

近期,这个行业公认的死结有望迎来最有力的破局者。由北京大学数学科学学院教授、我国首位戈登・贝尔奖获得者杨超创办的壹零壹壹科技(1011.AI),带着计算数学与高性能计算领域二十年的深耕积累下场。在杨超教授看来,物理AI的落地困境本质上是“算法—算力—场景”三者之间的系统性失配,整个行业需要从底层重构物理AI技术范式。

这支曾登顶全球超算之巅的团队,把国际顶尖的数值方法与并行计算能力转化到物理AI,着力打造“高精度物理仿真+全栈高性能计算”双根技术,精准击穿了行业长期悬而未决的 “准” 与 “快” 两大核心痛点。

资本的嗅觉往往最为敏锐。壹零壹壹公司自2026年6月成立以来短期内已完成过亿元融资,成为物理AI基础设施赛道最受关注的新锐玩家。天使轮投资方包括峰瑞资本、华控基金、北京信息产业基金、上海未来产业基金、道禾长期投资等头部市场化及国资机构,同时另有多家产业资本参与本轮投资。据悉,天使+轮也已经在陆续交割中。

一、物理AI的本质考题:不准不快,都是空谈

在杨超看来,物理AI与数字AI有着本质分野:数字AI的推演发生在虚拟信息域,误差只是输出结果的偏差;而物理AI要驱动实体机器人与真实世界交互,每一次计算误差,都会转化为真实的物料损耗、产线停摆甚至安全风险。

从第一性原理出发,“准” 与 “快” 不是优化方向,而是物理AI落地的硬性入场券。

准,是结果可靠的底线。抓取力度差一牛、运动轨迹偏一毫米,任务就可能失败;接触、形变、摩擦的计算偏离真实规律,仿真训练出的策略就无法在真机复用。

快,是实时执行的前提。真实场景中,物体滑落、接触碰撞、产线流转都发生在毫秒级,核心控制闭环必须在本地完成,无法完全依赖云端算力的传输与调度。

但行业现状恰好卡在这两个关口。“准” 的层面,真机数据采集是当前主流路径,但成本高、周期长,极端与危险场景无法复现,且数据质量高度依赖操作人员水平,人的能力就是数据质量的天花板。物理仿真本应是破局答案,却陷入了两难:主流AI仿真大多脱胎于图形学引擎,天生以视觉观感为核心,接触、形变、流体等物理过程大量采用近似处理,并不严格遵循物理定律,直接导致Sim2Real(仿真到真实)的鸿沟难以跨越;

传统工业仿真精度足够,但计算效率低了数个数量级,单次复杂仿真动辄耗时数天,根本无法支撑AI模型大规模训练的数据需求。“图形学仿真不够准,工业仿真不够快。物理AI要往前走,必须有一套原生的、同时兼顾高精度与高效率的仿真引擎”, 杨超的判断直指行业本质。

“快” 的层面,多数落地方案依赖集中式云端算力,可传输延迟、网络波动都是不可控风险。更现实的问题是,工业产线有固定节拍、零售场景有成本约束,端侧算力不足、模型跑不动、时延压不住,就永远匹配不了真实作业的节奏。

准度决定能不能用,效率决定能不能规模化。这两个死结不解开,物理AI就永远走不出实验室Demo阶段。

二、破解 “准” 的难题:从底层重做原生物理仿真引擎

一支做超算的团队为何能做好物理仿真?答案是能力同源。十年前杨超带领团队在国产 “神威・太湖之光” 超级计算机上,实现全球首个千万核可扩展非静力大气动力学全隐式模拟,求解规模超过7700亿未知数,核心要求就是:既准,又快。这套全隐式求解、多物理场耦合、大规模并行计算的底层能力,平移到物理AI场景,是天然的技术降维。

针对仿真失“准”的行业痛点,壹零壹壹自研了物理仿真引擎 “谐物”,从诞生之初就跳出 “重画面、轻物理” 的行业惯性,把误差控制贯穿资产构建、物理求解到轨迹输出的全流程,目标只有一个:让数字世界的每一次交互,都无限贴近真实物理规律。

这套高保真能力,拆解为三个可落地的维度:

第一是资产真。不同于传统仿真只做几何外观的3D模型,谐物搭建了覆盖刚体、铰接结构、柔体、布料与线缆、液体、颗粒等类别的数千个可复用物理资产,每一项都附带完整的材料参数、物性属性与场景边界条件,是能真正发生力学交互的数字实体,从源头消减仿真偏差。

第二是引擎真。团队从底层自研求解器,严格遵循物理定律计算接触碰撞、摩擦形变、刚柔耦合、刚流耦合等复杂物理过程。机械臂抓取柔性物件、灵巧手操作海绵、物体在液体中浮沉等典型场景,都能实现高置信度模拟,不再靠视觉效果 “凑数”。

第三是轨迹真。依赖严谨的数学计算,仿真输出的每一条动作轨迹,都严格满足动力学约束与执行器极限,确保训练出的策略可直接向真机迁移,从根本上避免 “仿真满分、真机零分” 的尴尬。

精度拉满的同时,成本并没有水涨船高。团队通过自适应建模、AI加速数值方法、隐式求解与大规模并行计算多维度协同优化,进一步实现了“少算、快算、大步算、并行算”等核心技术,在保证精度的前提下大幅压低计算成本。目前公司的仿真平台,在求解精度跻身行业第一梯队。更重要的是,团队没有陷入“纯仿真”的极端路径。在杨超看来,仿真追求极致精度,但真机数据仍有不可替代的校准价值。两者形成正向循环:用仿真低成本扩展样本空间、生成超越人类操作习惯的最优轨迹,用真机交互数据修正仿真参数、验证落地效果,才能真正缩小Sim2Real的差距。“虚实融合,可能是这个行业所需达到的根本性终局”,杨超说。

三、打通 “快” 的堵点:把超算能力沉到端侧

仿真解决“训练对不对”的问题,高性能计算则解决 “落地贵不贵、跑不跑得动” 的问题。 壹零壹壹没有把高性能计算当成配套工具,而是作为独立的根技术主线,打造了端侧高性能推理框架 “扶摇”,直戳“算得太贵”要害:不用堆砌昂贵的高端硬件,也能跑出满足工业节拍的实时性能。

杨超有一个明确的判断:实现端侧高速推理、摆脱对云端算力的单一依赖,是物理AI走向真实可落地的一道硬门槛。这个判断的依据完全来自产业现实:工业产线的节拍不会等云端传输延迟,零售场景的成本扛不住专属算力盒子,更广泛的普惠场景,更不可能为每一台机器人配置高端边缘算力。物理AI要规模化,必须能在不同层级的硬件上,以可接受的成本跑出实时性能。

这种跨架构、跨算力层级的深度适配能力,价值远不止于 “提速”:对机器人企业而言,不用盲目升级硬件就能满足作业节拍要求,直接降低部署成本;对芯片厂商而言,软件优化能充分释放硬件潜力,快速拓宽物理AI的应用场景。既能为客户节省硬件成本,又能为客户提生产效率,这也正是底层技术的核心产业价值。

四、商业化跑通:从生态到场景,全产业链验证落地

技术好不好,最终要看客户愿不愿意买单。成立短短数月,壹零壹壹的技术底座已经快速拿到全产业链的落地验证。

在上游生态侧,公司已与头部大模型厂商开展仿真数据合作,获邀参与具身智能评测榜单的仿真环境共建,进入多家云厂商的仿真引擎供应商清单;端侧推理框架也与国内头部芯片厂商推进联合适配,共同优化边缘算力下的运行性能。

在下游产业侧,这套全栈能力已经落地零售客户,开展深度合作。对场景方而言,物理AI落地的诉求高度一致:部署周期短、综合成本可控、能匹配真实作业节拍。而壹零壹壹的技术组合恰好精准命中需求:

高保真仿真大幅降低了真机数据采集和方案调试的成本,新场景、新物体的适配周期大幅缩短;

端侧推理优化保障了现场作业的效率与稳定性,不用依赖云端就能跑通生产流程。

相较于demo展示的行业普遍,壹零壹壹已经迈出产业落地的实质一步。

五、硬核且差异化的Infra能力收获多家头部市场化机构加持

物理AI赛道从不缺概念与热度,但潮水退去,终究要回归产业本质。

“政策与产业需求的共同推动,正在为中国物理AI打开规模化应用空间,其发展重心将进一步转向理解真实工况、可靠执行任务,并在制造、能源、交通、航空航天等产业中创造可衡量的提质增效价值。壹零壹壹选择的不是追求快速堆叠场景的故事性,而是沉下心打磨仿真与计算两项根技术;不依赖前沿概念放大估值迅速融资,而是用产业落地验证技术价值。这条路不喧嚣,但足够扎实;这条路不讨巧,但足够接近物理AI的本质。”——天使轮领投方,知名投资人峰瑞资本创始合伙人李丰总结道。

天使轮核心投资方,华控基金董事长张扬补充道:“对于一级市场而言,物理AI的投资逻辑也正在发生变化:从追逐场景应用的短期故事,转向布局底层基础设施的长期价值。仿真与计算是物理AI的通用底座,不依赖单一赛道,却能赋能全场景,具备极强的复用性与产业穿透力。从用数学公式描述物理规律,到用超算能力求解复杂系统,再到用技术赋予机器人真实行动能力,壹零壹壹走的是一条 ‘从底层向上击穿’ 的路径。当行业都在向上卷应用、卷场景时,向下扎根底层技术的玩家,反而会构建起最深的护城河”。

硬科技创业从来不是一场短跑,而是一场需要耐心与定力的马拉松。可以预见,随着物理AI从实验室涌向千行百业,底层技术的厚度,最终将决定产业能走多远。而壹零壹壹正在做的,就是为中国物理AI产业,筑牢可信、高效的数字地基。

注:文/簪竹,文章来源:投中网(公众号ID:China-Venture),本文为作者独立观点,不代表亿邦动力立场。

【版权提示】亿邦动力倡导尊重与保护知识产权。未经许可,任何人不得复制、转载、或以其他方式使用本网站的内容。如发现本站文章存在版权问题,烦请提供版权疑问、身份证明、版权证明、联系方式等发邮件至run@ebrun.com,我们将及时沟通与处理。

已点赞

+1

{{ praise_text }}

好文章,值得被更多人看到

FAQ回顾

物理AI落地面临哪些核心难题?

物理AI落地卡在“准”与“快”:真机数据采集成本高、周期长,图形学仿真重视觉轻物理导致Sim2Real鸿沟,传统工业仿真精度够但太慢;云端算力又有传输延迟和成本约束,端侧算力不足难以满足毫秒级实时控制。

壹零壹壹科技是一家什么公司?

壹零壹壹科技(1011.AI)由北京大学数学科学学院教授、我国首位戈登·贝尔奖获得者杨超创办,2026年6月成立,专注物理AI基础设施,打造高精度物理仿真引擎“谐物”和端侧高性能推理框架“扶摇”,已完成过亿元融资。

谐物仿真引擎如何提升物理AI仿真精度?

谐物从资产、引擎、轨迹三方面提升精度:构建数千个带材料参数和物性属性的可复用物理资产;底层自研求解器严格遵循物理定律计算接触碰撞、摩擦形变、刚柔耦合和刚流耦合;输出轨迹满足动力学约束与执行器极限,可直接向真机迁移。

端侧推理框架扶摇能解决物理AI的什么问题?

扶摇针对“算得太贵、跑不动”的问题,通过跨架构、跨算力层级适配,让物理AI不依赖云端和昂贵高端硬件,也能在端侧跑出满足工业节拍的实时性能,降低机器人企业部署成本,并帮助芯片厂商释放硬件潜力。

物理AI如何缩小仿真到真实世界的差距?

可采用虚实融合路径:用高精度仿真低成本扩展样本空间、生成超越人类操作习惯的最优轨迹,再用真机交互数据修正仿真参数、验证落地效果,形成正向循环,从而缩小Sim2Real差距,提高策略在真机上的可复用性。

广告

下一篇

李佳琦半小时卖出2万亿Token 阿里激进推动算力零售化

Token这个过去只出现在开发者后台和企业采购合同里的计量单位,第一次以商品的身份完成零售首秀。