内容来源：https://aibusiness.com/robotics/lack-training-data-stifling-humanoid-bot-development

内容总结：

当前，生成式人工智能技术的快速发展正在推动全球机器人产业加速演进，人形机器人及其他商用机器人的测试与部署不断提速。然而，训练数据的严重短缺正成为制约行业进一步突破的关键瓶颈。

据加拿大科技公司Telus Digital人工智能增长与解决方案副总裁斯凯·派克介绍，大语言模型之所以能够快速迭代，是因为“基本上把整个互联网的数据都消化了”。但机器人无法照搬这一路径。她援引前Meta首席人工智能科学家扬·勒昆的比喻指出，人工智能模型的实际智能水平仅相当于学龄前儿童，而一个4岁儿童对世界的理解能力实际上是大语言模型的五倍。儿童通过视觉和感官输入来理解物理规律、重力以及因果关系，仅靠阅读文本无法获得这些能力。

派克指出，训练当今先进机器人及其世界模型面临的最大障碍之一，是机器人开发者所需的视频素材数量庞大，其计算和数据存储成本极为高昂。此外，用于采集物理数据的不同摄像头、激光雷达设备和红外传感器种类繁杂、标准不一，也构成另一大障碍。她表示，这“产生了大量噪声”。

目前，派克正致力于解决人形机器人领域最重大的难题之一——物理安全。她表示，糟糕的数据在现实世界中造成的问题远比聊天机器人给出错误信息严重得多。聊天机器人出错可能只是提供错误信息，但在物理世界中，一旦机器人在拥挤的商场环境中摔倒并失控挥舞，后果将不堪设想。

中文翻译：

赞助内容 由 Google Cloud 呈现
如何选择你的首批生成式 AI 应用场景
要上手生成式 AI，先聚焦于那些能借助信息改善人类体验的领域。
AI 模型用互联网上的信息来训练。机器人可做不到这一点。
机器人技术面临一个大问题。
过去四年，生成式 AI 技术的迅猛发展给全球机器人行业注入了强大动力，也加快了人形机器人及其他商用机器人的测试和部署，但机器人开发者如今却在某种程度上陷入了僵局，原因正是训练数据短缺。
“大型语言模型……基本上，它们把整个互联网都吞了进去，”加拿大科技公司 Telus Digital 负责 AI 增长与解决方案的副总裁斯凯·派克在 AI Business 的《Targeting AI》播客中说道。
派克目前正牵头 Telus Digital 的机器人及世界模型项目。她提到了前 Meta AI 首席科学家扬·勒昆一个广为人知的类比：把 AI 模型的实际智能比作学龄前儿童。
勒昆“认为物理 AI 世界几乎超出了仅靠互联网数据所能触及的范围，”派克接着说。“他把它比作一个 4 岁孩子，而这个 4 岁孩子的理解力实际上大约是大语言模型的五倍。4 岁孩子学习的方式是通过视觉输入。4 岁孩子用他们的视觉神经和感官输入去理解物理如何运作、重力如何运作、因果关系如何运作。光靠读文本，你根本得不到这些。”
派克说，训练当今先进机器人以及指导它们在现实世界中行动的世界模型，最大的障碍之一就是机器人开发者所需的视频素材数量惊人，而且计算和数据存储成本高得离谱。另一个障碍则是用于采集物理数据、训练机器人 AI 模型的各种摄像头、激光雷达设备和红外传感器五花八门、互不兼容。
“这会产生太多噪声，”派克说。
派克正在攻关人形机器人领域最大的难题之一——物理安全。
“坏数据在现实世界里造成的问题，肯定比聊天机器人给你错误信息要严重得多，当然，聊天机器人给错信息本身也可能非常麻烦，”她说。“但在物理世界里……一个巨大的问题是，如果那台机器人在拥挤的商场环境中摔倒，还在地上乱挥乱动。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
AI models train on information available on the internet. Robots can’t do that.
Robotics has a big problem.
While the rapid development of generative AI technology over the last four years has supercharged the global robotics industry and accelerated testing and deployment of humanoid and other commercial bots, robot developers are at somewhat of an impasse because of the shortage of training data.
“Large language models … basically, they ingested all of the internet,” said Sce Pike, vice president of AI growth and solutions at Canada-based technology company Telus Digital, on the Targeting AI podcast from AI Business.
Pike, who is leading Telus Digital’s robotics and world model projects, referred to former Meta AI chief scientist Yann LeCun’s well-known analogy comparing the actual intelligence of an AI model to a preschooler’s.
LeCun “looks at the physical AI world as something that is almost beyond reach of just internet-based data,” Pike continued. “He compares it to a 4-year-old, who actually has about five times more comprehension than what an LLM has. The way that a 4-year-old learns is through visual input. A 4-year-old is using their visual ocular nerves and sensory input to understand how physics works, how gravity works, how cause and effect works. You just don't get that by just reading text.”
One of the biggest barriers to training today’s advanced robots and the world models that guide them in the real world is the sheer quantity and exorbitant compute and data storage cost of video footage robot developers require, Pike said. Another impediment is the clashing array of different cameras, lidar devices and infrared sensors used to collect physical data to train AI models for robots.
“It is causing so much noise,” Pike said.
Pike is working on one of the biggest problems in humanoid robotics -- physical safety.
“Bad data definitely can cause a lot more issues in the real world than with a chatbot where it just gives you the wrong information, which could be very problematic as well.,” she said. “But in the physical world … A huge issue is if that robot falls and is flailing around and it's in a crowded mall environment.”