Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Y Combinator hosts founders from Waddle Labs and RoboCurve to discuss how general-purpose LLMs like Claude and GPT-4o may beat specialized models at controlling robots.
The gist
-
LLMs write Python programs instead of single actions, letting robots handle repetitive steps while using vision to reason through decision points.
-
In-context learning tops out after 20 to 40 examples, so learned skills must be compressed into reusable code libraries to keep improving.
-
Training data from computer use, CAD tools, and GUI screenshots turns out to help LLMs understand 3D space and physical robot control.
-
The founders suggest general-purpose robots capable of human-level manipulation could arrive within 2 years if diverse, high-quality data keeps scaling.
Y, Riley, Greg and Peter agree on 38 things
See what they all use, in their own words. Paid ads are left out.
See what they all swear by Or pick your own peopleAlso mentioned
0:40
Waddle Labs
· "I'm Jaime. I'm from Waddle Labs together with Vincent. We work on building LLMs that control robots and we..."
9:37
ChatGPT
· "What was the capability of those compared to what you can see now? Those code-as-policies papers came out..."
20:57
Platonic Representation Hypothesis
· "Inspired by this paper also from Philip Isola that he titles the Platonic Representation Hypothesis, where..."
◎
Related episodes
Episodes covering similar topics
More from Y Combinator
Inside Theranos with ex-Employee Tyler Shultz
Nov 11, 2026
How To Build The Future: Paul Graham & Amjad Masad
Nov 07, 2026
Bringing Supersonic Flight To The People
Nov 06, 2026
The $1B Company Fixing AI’s Software Bottleneck
Nov 02, 2026
What Happens When Everyone Can Build Software
Oct 30, 2026