Robotics has an unwritten rule. Hardware first. Deployment later. A small team in Brooklyn ignored that rule and still shipped a working robot in months. Here is what they did differently 🧵:

The devastating truth of warehouse automation: it's built for giants. It’s rigid, costs millions, and takes weeks or months to integrate. If a task changes, the robot is a paperweight. This has left thousands of 3PLs stuck with a 1.3 million-worker labor shortage.

Enter @Ultraroboticsco from the YC S24 batch. The founders aren't new to this. This is the same core team that built Voodoo Manufacturing, one of the world's largest 3D printing factories. They don't just know code. They know scaling real-world hardware.

The team is a powerhouse: Jon Miller Schwartz (CEO) - @JonMSchwartz Max Friefeld (COO) Oliver Ortlieb (CTO) Chetan Parthiban (Chief Scientist) - @chetan_ This isn't a team of academics; it's a team of serial entrepreneurs who have scaled companies.

While everyone else is chasing $100k+ humanoids or complex custom hardware... Ultra is doing the opposite. They realized the hardware is "mostly" solved. The real problem (the one holding everyone back) is the software: The "brain."

Here's the magic. Ultra uses reliable, off-the-shelf robot arms. Their innovation is a neural network control system. Instead of hiring robotics PhDs to write code for every single task... They use remote human operators to train the AI by example. This is the key. Deployed from Day 1, providing value (and collecting data) via teleop.

But the most impressive part? The velocity. Ultra was founded in 2024. They didn't spend a year in a lab.

They are already deployed in 3PL warehouses in Brooklyn, Chicago, and Austin. Their first robots are already picking and packing e-commerce orders for real customers. This is not a simulation. [-8-hr shift, 1,000+ orders packed-]

Let me repeat that: They are deploying robots in hours, not weeks. Real zero integration. Drop it in and the operator starts working while the AI learns. Listen to my full interview with @JonMSchwartz:

Why this matters: This approach breaks the "robotics-is-rigid" curse. It creates a data flywheel. 1. More deployments 2. More real-world data 3. Smarter AI 4. Faster, more autonomous robots They're building a data moat from actual warehouse chaos.

One of the most compelling robotics stories this year. Ultra was part of the YC S24 batch, a record setter for robotics. This is the receipt. A veteran team solving a real labor crisis with an AI and data first approach that delivers value from day one. This is startup speed. Credit:

Solving the chicken egg problem of robotics, need smarter AI to deploy, need to deploy to get high-quality data for smarter AI. By deploying more robots and collecting real-world data to train AI, they are breaking the cycle.

I like these guys because they have real revenue and are pricing it as a RaaS just cheap enough compared to a human to get customers. They are also working with small companies (why burden yourself with a F100 giant when small customers move faster). They aren’t yet profitable since reliable robot hardware isn’t cheap enough, but it’s obviously going to get there. Their IP is everything you need for a RaaS to work. Training in house AI, teleop built into the production workflow, sales, service, etc. In a way, these guys are a robot integration company. Even in the Tesla nirvana future of cheap performant humanoids, someone is going to need to integrate AI bots into factory workflows and then maintain them. And AI factory bots won’t all be humanoids, there will always be a market for specialized form factor AI bots that companies like this will be nimble enough to use.

Ship first, perfect later. The best way to learn is by getting real hardware into the field and iterating.

They produce a great machine. It’s a somewhat placeholder tech / mindset as framed here. Most of the others are not focused on the stop gap semi human replacement automation. Any automation that does not require human interaction will gravitate towards efficiency which is not the human form. The others are chasing the ‘Human Butler’ market. And they are chasing autonomous software and hardware to accomplish that. A quasi R2D2 form with assorted attachments is a much better Idea for personal use IMO but the majority want a C-3PO variant🤷‍♂️

Bullish on this team ;)

Impressive! What’s the speed up on these videos?

Great approach, I have always wondered why robots need to wander the plant floor when most often they are replacing stationary workers such as this.

Intentionally speeding up video? Nikola truck vibes.

This is basically the usual way to design robots get the hardware right first, then build the software around it.

There's absolutely no such rule for new experimental robots. It doesn't even make sense. It's very unclear what's your thesis.

Thanks for the explanation, but to me, this is similar to the robots in the car factory, except that one is heavy industry, and what you showed is light industry.

@growthhub_ @TheUpsiderAI Save this post

I wonder how their datasets port to updated hardware. Ie, the arms/vision combo in tbr videos is super janky, and likely to need an upgrade soon — does the trained model transfer well to new hardware kit?

that's our robot! 👀😍

Please explain the difference, lest the Chinese little pink say that your American robots are played by real people.

Robotics takeover for real 🤖🦾

But nobody is still thinking about the battery

I love real data

Head moving is a weird teleoperation artifact? Or is this teleOp now? Companies have to STOP speeding up videos. Real human see it and think “so it’s really slow. Cool.”

I have a set of applications which I would like to apply the same intelligence flywheel to

Love when teams skip the endless planning and just ship. Real progress happens in the trenches, not in theory. The Brooklyn crew nailed it.

What do you think of this?

Johnny 5 is alive....wow. Very nice.

Do you think dynamic/adaptive or self recovering palletizing can be developed via this approach?

Excellent post. Thanks for sharing.

@lukas_m_ziegler Hm, I am confused. What is deployment in this context (real question, I want to know)?

Impressive and productive.

They found the alpha a working robot means they already built the node for the AI L2!

Working robots in months is impressive.

Love this! I hope they become profitable

Gud tec. Love their approach. Thx for making that interview

Those are off the shelf cobot manipulators. No wonder they don't have to worry about hardware first.