• Hard Startups
  • Posts
  • Automatic CAD Instructions, ImageNet for Robots, Lockheed Mines An Asteroid

Automatic CAD Instructions, ImageNet for Robots, Lockheed Mines An Asteroid

Hey y’all in this week’s post we have:

- Automatic Assembly Instructions from CAD drawings
- Largest dataset of it’s kind enables new robotics capabilities
- Lockheed mines an asteroid
- $32M raised for robotics craftsman
- our first sponsor, check them out below!

from our sponsor

Build better products faster.

Working on complex hardware? Had enough of JIRA? Us too.

Motivated by our experience developing Formula One cars, semiconductors, and large-scale energy infrastructure, we built Anneal: a digital workspace for engineering teams. Intended to complement existing PLM/PDM solutions, it lets engineers collaboratively plan, discuss, and review engineering concepts, designs, and drawings.

Our free tier is now in public beta—and we're excited to invite Hard Startups subscribers to join. Visit getanneal.com to sign up.

Ikea (Assembly Instructions) for Everybody Else

On Monday, Dirac emerged from stealth. The NY-based startup is building a tool that helps automate the tedious job of writing assembly instructions.

The traditional way to create these type of instructions is by having some (probably unhappy) engineer take tons of CAD screenshots, manually write step-by-step instructions, and compile those into an instruction manual.

Dirac accelerates manufacturers by automating this process. Their software takes a CAD file, generates 3D assembly animations, and most of the assembly instructions automatically.

Writing instructions might seem like a niche thing to build a product for but the replies on Fil’s announcement paint a picture of how big of a problem this is for manufacturers. Imagine manually writing assembly instructions for a system with 10,000 parts.

I caught up with Fil (CEO of Dirac) post-launch and things look like they’re off to a strong start.

ImageNet for Robots

Deepmind has collaborated with > 30 robotics research labs around the world to build the biggest robotics dataset of it’s kind. The dataset contains over 1M robot trajectories, spanning more than 500 skills across 22 different robots.

They then trained their RT-1 and RT-2 models on this new dataset and saw incredible results.

Some industry experts think that this is the ImageNet moment for Robotics
(ImageNet is a huge dataset of over 14 million hand-annotated images launched in 2009. It’s creation helped kick off the deep-learning industry and enabled incredible progress in Computer Vision.)

An ImageNet type dataset for robotics could enable the industry to build more capable robots faster than ever possible before.

from Deepmind’s blog post. A robot is tackling an unseen task using RT-2-X

Deepmind was able to triple the performance of their RT-2 model by training on this new dataset. One surprising result is the diversity of robots used in the dataset helped improve individual robot performance. This is promising progress toward a general model for robotics.

In this work, we show training a single model on data from multiple embodiments leads to significantly better performance across many robots than those trained on data from individual embodiments.

ImageNet kicked off a revolution in deep learning, spawning many startups in the process. Let’s hope this work has the potential to do the same for robotics.

Quick Links

Lockheed Martin mines an asteroid and returns sample to earth after 7-year mission.

Boom Supersonic makes progress with the FAA for their commercial supersonic airplane Overture.

Cameron Schiller of Rangeview shares a cool picture of their robotic foundry. 🇺🇸

Machina Labs raises $32M to build robotics craftsmen.

Thanks for reading 🫡

We’ll see y’all next week.

I’d be glad to hear what you thought of this newsletter, just reply back with your thoughts!