What robots can (and can’t) do in 2025
08.29.2025 | By: Joanne Chen
Ken Goldberg is a professor of engineering at UC Berkeley and the co-founder of Ambi Robotics, a company applying AI-enabled robotics to the logistics industry. Ken has spent over four decades working on one of the hardest problems in robotics: how machines perceive and manipulate the physical world. We spoke about why tasks that seem effortless to humans – like picking up a glass or folding laundry – are still incredibly difficult for robots.
Our conversation also covers:
- What it would take to reach a “ChatGPT moment” in robotics
- Why simulation data isn’t enough without real-world grounding
- And why the next decade of robotics depends on combining cutting-edge models with good old-fashioned engineering
Chapters:
- 00:00 Cold open: Why robotics still needs good old-fashioned engineering
- 03:46 Hype cycles and winters in robotics
- 05:08 Why folding laundry is still hard for robots
- 10:38 What robots are good at today
- 15:00 Automation and the rise of warehouse robotics
- 19:39 Can LLMs and generative AI work for robotics?
- 26:52 The limits of simulation data and the sim-to-real gap
- 29:44 Why humanoids are still far from practical
- 36:34 What founders need to know about robotics timelines
- 37:08 Why robots need grounding and exploration
- 39:00 Combining the power of LLMs with traditional engineering
- 40:42 Why Ken is optimistic about the future of robotics
Published on August 29, 2025
Written by Joanne Chen
