Ben Lorica: coding agents that build the robot instead of controlling it

bigdata · x · 2026-10-28

In a Gradient Flow essay titled "I changed my mind about world models," Ben Lorica asks a new question: instead of one big neural network controlling the robot, what if AI builds the robot itself?

Key points

A large class of "variational automation" jobs in warehouses, factories and food processing—same machine and workspace, but packages, groceries and cables differ every time—remain expensive to automate.
Bottlenecks: robots lack the vast physical-experience data of language models; each application still needs engineers to assemble perception, grasping, motion planning and recovery; simulations are labor-intensive to build; and the commercial bar demands fast, reliable operation, not cherry-picked demos.
He outlines four approaches to building robots, from traditional modular systems to newer Vision-Language-Action (VLA) models.
The emerging paradigm: coding agents assemble programs from reusable robotics skills, test them in simulation, and fix what breaks—cutting the cost of testing and fixing robot software.

Lorica revises his earlier thesis that robotics is moving toward general-purpose foundation models: models are getting easier to buy, the hard work is making data useful—and now even the engineering of the robot itself may be automatable.

Original post →

Ben Lorica: coding agents that build the robot instead of controlling it

More from Embodied