参数化与人工智能,从计算机辅助到计算机决策,同济大学DigitalFuture演讲记录

点击加入mixlab社区

15000+跨学科人群,机器学习、机器视觉、自然语言处理、知识图谱、量化交易、物联网、区块链、产品经理、交互设计、建筑设计、服装设计、珠宝设计、平面设计、音乐、艺术等。

作者:

ML258-郑豪-机器学习/机械臂技术/混合现实技术/生成式设计

这是他在同济大学DigitalFuture演讲稿,为我们介绍了人工智能在建筑领域的应用。欢迎大家关注他的公众号(见文末

很高兴收到同济大学建筑与城市规划学院的邀请,能给大家讲讲我之前的研究,从参数化到人工智能,从计算机辅助到计算机决策。

It's my honor to give a speech here in the College of Architecture and Urban Planning, Tongji University. My previous researches were mainly about two fields, computer-aided design and computer-decided design.

那么计算机能为建筑学做到什么?首先,它可以进行快读的迭代式的运算,来辅助建筑师描述复杂的形态和数据。在柔性强度这个项目中,我们使用了rhinovault设计了一个复杂曲面,以及它的曲面细分和非标准节点细节。原先难以被图纸描述设计,在计算机运算力的辅助下,得以通过三维模型数据进行表述。

So what can we do with computers? First, it can be used to carry out mass calculations to help architectsdescribe complex patterns and data. In soft rigidity 2016, we used rhinnovaultto design a complex surface, along with its details of tessellation and non-standard nodes. Designs, which were formerly difficult to describe by traditional architectural drawings, with the assistance of computing power, now can be expressed through three-dimensional model data.

同样的,在编织鸟巢的项目中,一个鸟巢的近似曲面被迭代细分。计算机辅助实现了设计细节的精确描述,这里的每一块单元体都是严格通过算法计算曲面,然后3D打印而成的。

Also, in the Bird Nest2017, the approximate surface of the bird's nest was subdivided into hundreds of panels. Each panel was precisely generated and described by a certain algorithm, and then 3D printed.

另外,在玻璃纤维弯曲这个项目中,我们通过对材料特性的力学模拟,来重现材料弯曲时的最优形态,最终形成一个自受力平衡的构筑物。这三个项目都是利用了计算机的运算能力,来帮助建筑师实现自己的设计理念,所以我们称之为计算机辅助设计。

In fiberglass bending 2017, we simulated the materialproperties of resins and fiberglass, to optimize the shape of the pavilion, so that after bending, the pavilion will support itself and become a structure with self-balanced forces. Soin all three projects above, the usage of the computer is just to help architects achieve their own design concepts, like anassistant without creative work, so we call it computer-aided design.

另一个概念是计算机辅助建造,通过对机械生产工具的高精度控制,来实现建造精度上的巨大提升。在集群行为这个项目中,机械臂控制下生产出来的单元构件都是几乎没有误差的。计算机在这里扮演了监工的角色,辅助一个构筑物的精密建造。

Another concept is computer-aided construct, which achieves a great improvement inconstruction accuracy through the precise control of mechanical production tools. In the swarm behavior 2015, we used a robotic team to fabricate the components of the pavilion, so there were almost no errors in the components produced by the robotic arms. The computer playedthe role of a supervisor here to lead the precise production of a pavilion.

而随着混合现实技术的发展,计算机不仅能将建造数据交给机械臂,还可以将数据通过图像嵌入的方式表达在视觉数据上,并实时反馈给人眼,指导人类进行精确的建造。计算机在辅助建造的过程中,同样只是帮助建筑师来实现设计理念,并没有实际的创造性成分在内。

With the development of mixed reality technology, the computer can not only deliverthe construction data to the robotic arm, but also visualize the datathrough the image embedding, and feedback to the human eyes in real time to guide the workers to make the accurate construction. By looking through the VR glasses, we found a clear digital model shown in front of the target image. According to it, we assembled all 22 wooden blocks into the right positions. It only took 6 minutes and one untrained worker with a smart phone to build. Compared torobotic assembly, it provides accuracy while cutting down the cost in a very high degree. But again, in the process of computer aided construct, the computer still works to help architects to build their designs without creative jobs.

这么看来,计算机显然已经替代了人类的劳动力工作,那么下一步,计算机是否有可能不仅作为辅助者,而是作为决策者,取代建筑师呢?

It seems that the computer has the potential to fully replace the human labor. So in the next step, is it possible that the computer may not only be an assistant, but also be a decision maker todo creative jobs, and finally replace mental workers, like architects?

答案是肯定的。在人工智能高速发展的背景下,计算机已经具备学习的能力。它可以学习两种数据之间的对应关系,并找出最合适的方法来转译输入和输出数据。我们使用大量的建筑平面图训练了一个预测模型,在给定外边界的情况下,计算机就可以生成可能的内部设计。

Of course it will. Under the rapid development of AI technology, the computer already has the ability to learn like human. It can learn the correspondence between two kinds of data and find the most suitable method to translate between input and output data. We trained a machine learning model to map one image into the other image. So the network will take in a boundary image as pure white background and black block, then output the predicted architectural drawings inside the boundary.

我们还建立了一个城市图像模型,并配合一套绘图系统,来预测相同的城市配置情况下,不同的城市肌理表现。在系统中,用户可以自由的指定城市元素的信息,比如建筑物的位置,道路的位置等等。程序将针对用户选择的城市,生成预测的卫星图像。

Also based on the same system but different training data from city maps, we built models to predict the most possible satellite images of a given colored map. So in this application, users can draw colorful images to represent a simplified city map. Then by inputting the sketches into the program, designers can get the predicted satellite photos as a preview of city images.

之后,我们将研究重点放在了室内设计中。同样使用图像处理的人工智能技术,我们先建立了一个平面图识别系统,对于给定的室内装修平面图,系统能够识别出不同房间的位置,用不同的颜色来表示。

Later, we focused our research on interior design. Using the same technology of image-to-image training, we first built a floor plan recognition system. For a given interior plan, the system can identify the positions of different rooms and use different colors to represent them.

同样的反过来,对于给定的平面颜色区分,系统能自动判别每个房间的属性,并设计出最合理的室内布局,比如家具的位置。而这些学习模型仅仅需要100张左右的平面图,就能达到一定的效果。

In the same way, for a given colorful map of rooms distribution, the system can automatically determine the properties of each room and design the most reasonable interior layout, such as the position of the furniture. This model was trained only by 100 floor plans and it already has the ability to generate drawings for preview.

我们还开发了另一套基于矢量数据的学习系统,来更精确地预测室内布局。相比于之前的图像模型,基于CAD文件的人工智能技术提高了训练速度2000倍,预测精度10000倍,并能给出多个合理的设计供用户选择。

Next, we developed another vector-based machine learning system to predict the interior layout more accurately. Compared to previous image models, vectorized model based on CAD files speeds up the training process 2000 times and increases the prediction accuracy 10000 times, and the model can give multiple solutions for designers to choose.

同时,我们也迁移了室内家具排布的矢量化模型,来自动设计卧室中的床和床头柜还有电视机的摆放位置。在完全没有人工干预的情况下,计算机决策设计能够根据给出的既定条件,快速得出不亚于人类设计的解。

At the same time, we also transplanted vectorized model of furniture layout to automatically design beds and bedsides as well as TV placements in bedrooms. Without any intervention from human, computer-decided design can quickly output possible predictions to a certain design question.

而下一步将会是计算机决策建造。波士顿动力公司已经研发出了人形机器人,配合人工智能的算法,全功能机器人的出现已经指日可待。未来甚至整个城市的设计到建设都将由机器人团队完成。

And the next step will be computer-decided construct. Boston Dynamics has developed humanoid robots. With the machine learning algorithm, the emergence of fully-functional robots is just around the corner. In the future, even the design and construction of the entire city will be completed by the robot teams.

在人工智能高速发展的今天,我们建筑师应该思考自己的定位,找到合适的价值观和突破点,应对即将到来的冲击。未来,人类,建筑师,这将是我们之后需要探讨的问题。

So nowadays, with the rapid improvement of AI technology, our architects should think about our own positions, find new values and breakthrough points, and defense ourselves from the incoming impact. Future, humans, architects, will be the issues we need to discuss.

谢谢!

Thankyou for listening!

郑豪,宾夕法尼亚大学设计学院博士生,程序和设计研究者,专攻机器学习,机械臂技术,混合现实技术,生成式设计。他毕业于加州大学伯克利分校(UCB),在Simon Schleicher教授的指导下获得建筑学硕士学位,本科毕业于上海交通大学,在刘士兴教授的引导下获得建筑学学士学位。他曾工作于清华大学,先后于徐卫国工作室研究机械臂辅助施工,并于黄蔚欣工作室研究机器学习。他还曾在加州大学伯克利分校担任研究助理,和Maria Paz Gutierrez教授研究仿生材料3D打印,和Kyle Steinfeld教授研究人工智能。

Hao Zheng is currently an incoming Ph.D. student at the University of Pennsylvania, School of Design. He is a programmer and design researcher, specializing in machine learning, robotic technology, mixed reality, and generative design. He holds a Master of Architecture degree from the University of California, Berkeley, advised by Prof. Simon Schleicher, and a Bachelor of Architecture degree from Shanghai Jiao Tong University, advised by Prof. Shixing Liu. Before joining UPenn, Hao worked as a research assistant at Tsinghua University with a concentration on the robotic assembly under the supervision of Prof. Weiguo Xu and machine learning under the supervision of Prof. Weixin Huang. In addition, he worked as a researcher assistant at UC Berkeley, researching bio-material 3D printing supervised by Prof. Maria Paz Gutierrez and machine learning supervised by Prof. Kyle Steinfeld.

开启建筑领域的AI之旅

——公众号回复留言即可加入社群。

以上为全文。

近期活动推荐

快!TensorFlow、黑客马拉松、DevFest千人大会 ,GDG DevFest 2018在召唤你

--------------------------------------

mixlab 无界社区 是一所面向未来的实验室,它提倡“跨界创新,开放成长”的理念。

——跨界 开放 互助 学习 思维 创新。

目前社区汇集了15000+跨学科人群,主要来自GoogleBrain、微软、华为、阿里鲁班、腾讯、旷视、三角兽、物灵科技、众安保险、美团、360等科技人才、设计师及CEO和投资人;高校分布MIT、Oxford、Cambridge、CMU、UoM、清华、上交大、同济等学生及教师群体;学科跨越机器学习、自然语言处理、量化交易、物联网、区块链、前端、后端、产品经理、UI设计、建筑设计、服装设计、珠宝设计、音乐、艺术等。

社区典型的用户技能为:

ML00:小程序开发、AI+设计、区块链技术

ML04: 算法作曲

ML05:AI变脸

ML07:WebGL+Three.js数据可视化

ML14:增长黑客

ML16:机器学习

ML37:AR

ML41:人工智能+音乐

ML44:珠宝设计+参数化设计

ML83:量化交易、数字货币交易

ML109:Arduino、processing

ML113:人工智能+服装设计

ML206:VR/AR

ML245:建筑设计智能化

ML258:机器学习、机械臂技术、混合现实技术、生成式设计

ML260:AI+设计、无人机

详细介绍点击以下卡片:

mixlab期待您的加入!

原文发布于微信公众号 - 无界社区mixlab(Design-AI-Lab)

原文发表时间:2018-10-11

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏大数据文摘

用机器学习的经验指导人生:如何实现学习效率最大化

1193
来自专栏人工智能头条

【CSDN AI 周刊】第11期 | 周志华提出深度森林 引发持续热议

1613
来自专栏新智元

精品 | 2016三大深度学习峰会:顶级科学家360度剖析技术重点、应用关键及投资机会

【新智元导读】Re-Work上放出了本年度波士顿、伦敦、柏林的三大深度学习峰会的全部演讲视频,156位行业精英带来精彩演讲,其中包括 Yoshua Bengio...

43419
来自专栏腾讯研究院的专栏

何为实在

image.png 推荐语: 霍金的《大设计》(The Grand Design)与其说是本科普读物,莫如说是相关“实在与终极”的著作。过去几千年间,...

2565
来自专栏吉浦迅科技

深度学习协助机器人学习人类的行走方式

机器人 Darwin 行走时身体仍会摇摇晃晃,有时候还会跌倒,但它跟大多数机器人不一样,它会从错误中学习(这点跟人类很像),飞快地调整技术。 它的踉跄学步导致新...

4558
来自专栏AI科技大本营的专栏

这个Kaggle三项排行榜的“顶级大师”,今年17岁

Kaggle,这是个坐拥 100 万会员的数据建模和数据分析竞赛平台,目前已被 Google 收购。企业和研究者可在这个平台上发布数据,而统计学者和数据挖掘专家...

1664
来自专栏专知

这是一篇论文评审吗?

【导读】面对如今每年数量暴增的AI领域顶级会议论文投稿量,作者Zachary C. Lipton在Approximately correct上发表了对论文评审的...

1312
来自专栏专知

走近Hinton:AI教父传奇人生

【导读】Geoffrey Hinton花费了30年的时间默默无闻,直到2012年,他证明了其研究的价值,并驳回了大多数其他科学家所谓的正确观点。如今,这个被称为...

4006
来自专栏AI科技评论

金句频出,16位大神在ICML上展开了一场机器学习的Great Debates

AI 科技评论按:机器学习圈子这几天发生了一件有趣的事,一群知名学者在 ICML 2018 上抱着吵架(辩论)的心思坐到一起,结果却没能吵起来。

1534
来自专栏新智元

斯坦福医疗ImageNet发布,如何评价PB级医疗影像数据集?

【新智元导读】斯坦福大学医学院与 Langlotzlab 合作创建的一个 PB 级的大型医疗影像数据集 Medical ImageNet 最近发布,从官方网页的...

4697

扫码关注云+社区

领取腾讯云代金券