Bob McGrew tried starting a robotics company in 2016. 

He built a system that taught a robot to play checkers by watching the board through vision.

When he looked at what it would take to make robots work in factories, he concluded the technology was "very fun and super cool and really far away from any form of commercialization."

He joined OpenAI instead and spent eight years leading development of GPT-3, GPT-4, and ChatGPT.

Now he's raising $70 million at a $700 million valuation for Arda, a startup that uses video-based AI to automate factories.

👋🏻 I'm Leonardo Ubbiali. This week we're looking at what changed between 2016 and 2025 that made McGrew go back to the problem he walked away from.

In 1985, General Motors Roger Smith spent $45 billion automating factories with robots.

The robots painted each other, welded doors shut, and smashed into equipment because they were programmed for specific tasks and couldn't adapt when parts arrived with slight variations.

Honda kept the least automated factories with the most people on production lines. When Honda needed to change a production line it took days. 

At GM it took months.

Tesla repeated GM's mistake in 2018 and on April 13, 2018, Musk admitted: "Humans are underrated."

That was the pattern McGrew saw in 2016. Robots were rigid, manufacturing requires flexibility, and the gap was too wide.

McGrew watched the robotics team spend years teaching a robot hand to manipulate a Rubik's Cube.

Years for one specific task.

Now companies like Physical Intelligence teach robots diverse problems like folding laundry and packing egg crates in months.

McGrew explained in a Sequoia interview: "Now you have LLMs, this language interface to the robot so you can describe tasks much more cheaply. And you have really strong vision encoders."

The robot doesn't need programming for every variation because it understands natural language and learns from watching demonstrations.

OpenAI spent years on the Rubik's Cube building the foundation.

Now companies can build on top of it.

McGrew's assessment shifted from "very far away" in 2016 to "end stages of being a research challenge" in 2025.

Arda uses a video-based AI model that watches factory floor footage and trains robotic systems to handle manufacturing autonomously.

The system learns from observation by watching the task, understanding the pattern, and applying it to variations.

McGrew co-founded with Augustus Odena from Adept AI and Palantir alumni.
Founders Fund and Accel are co-leading the $70 million round.

The goal is making Western manufacturing cost-effective.

US manufacturing labor costs $25-30 per hour versus China's $6-7.
Overall the costs in the US are 30-50% higher.

244,000 jobs were announced through reshoring in 2024, but 500,000 jobs remain unfilled.

The Reshoring Initiative noted US costs remain 10-50% higher than offshore competitors.

When subsidies run out, the math has to work on its own.

Physical Intelligence raised $600 million at $5.6 billion.
While Skild AI raised $1.4 billion at $14 billion.
Figure AI reached $39 billion with robots at BMW.

Arda at $700 million is building specifically for manufacturing rather than general-purpose robotics.

General Motors had the best robots available and they failed because General Motors didn't account for variability.

BMW tested Figure AI's humanoid robots at its Spartanburg plant for 11 months where the robots ran 10-hour shifts loading sheet metal parts into welding fixtures. 

The robots contributed to production of 30,000 BMW X3 vehicles and moved over 90,000 components, but BMW deliberately started by testing in the body shop where automation levels were already high and workers were familiar with integrating new technologies.

A supplier I spoke to last week is deploying robotic welding cells and they started by running the system on scrap parts for two weeks before touching production components.

They documented every failure mode: what happens when parts arrive with rust, when positioning is off by 2mm, when the wrong material grade shows up.

They also mapped volume fluctuations because most automation is designed for peak production, and during low-volume months the setup time for automated cells takes longer than doing the work manually.

Test before deploying, plan for failures, document manual procedures. 

That's what separated Honda from General Motors in the 1980s and what BMW is doing now.

The problem: You're considering automation but don't know which processes are good candidates.

Assessment of which processes automation handles well versus which need flexibility, volume impact, failure modes, and manual procedures.

How AI breakthroughs in vision and robotics create more intelligent industrial machines. Documents results from Amazon and Foxconn. Explains the shift from rule-based to training-based robotics.

Time to value: 25 minutes

How many of your automation plans account for what GM and Tesla learned the hard way?

Hit reply. I read every email.
Leo