2025-08-27 カリフォルニア大学バークレー校 (UCB)
A robot demonstrates the skill of grasping fruit at China pavilion during the 7th China International Import Expo in Shanghai, China. Teaching robots to grasp and manipulate objects is still a major challenge and one of the reasons we shouldn’t expect humanoid robots to become a fixture of our homes or workplaces within the next few years, says UC Berkeley engineer Ken Goldberg. Jia Tianyong/China News Service/VCG via AP
＜関連情報＞
- https://news.berkeley.edu/2025/08/27/are-we-truly-on-the-verge-of-the-humanoid-robot-revolution/
- https://www.science.org/doi/10.1126/scirobotics.aea7390
- https://www.science.org/doi/10.1126/scirobotics.aea7897
古き良き工学技術がロボット工学における10万年の「データギャップ」を埋める Good old-fashioned engineering can close the 100,000-year “data gap” in robotics
Ken Goldberg
Science Robotics Published:27 Aug 2025
DOI:https://doi.org/10.1126/scirobotics.aea7390
Large vision-language models (VLMs) based on internet-scale data can now pass the Turing test for intelligence. In this sense, data have “solved” language, and many claim that data have solved speech recognition and computer vision.
Will data also solve robotics? Rich Sutton points out in “The Bitter Lesson” (1) that data and black-box “end-to-end” models have surpassed all the best-laid analytic work in artificial intelligence (AI). I accept that this trend will eventually produce general-purpose robots. But the question is… when?
Using commonly accepted metrics for converting word and image tokens into time, the amount of internet-scale data (texts and images) used to train contemporary VLMs is on the order of 100,000 years—it would take a human that long to read or view these data (2). However, the data needed to train robots are a combination of video inputs with robot motion commands: Those data do not exist on the internet.
「データがロボット工学と自動化を解決する：真か偽か？」：討論 “Data will solve robotics and automation: True or false?”: A debate
Nancy M. Amato, Seth Hutchinson, Animesh Garg, Aude Billard, […] , and Ken Goldberg
Science Robotics Published:27 Aug 2025
DOI:https://doi.org/10.1126/scirobotics.aea7897
Abstract
Leading researchers debate the long-term influence of model-free methods that use large sets of demonstration data to train numerical generative models to control robots.
