- The use of 3D printers in manufacturing is expected to create a net positive of jobs in the US and negatively impact countries like China.
- The cost of robots has decreased, making them more accessible for integration into workplaces.
- Collaborative robots have safety features that prevent harm to humans.
- The software used for robotics is still developing, with the Robot Operating System (ROS) being a popular choice.
- The integration of robots in the 3D printing process allows for continuous production and scalability.
- Cheaper 3D printers are preferred for their reliability and the ability to run multiple jobs simultaneously.
- To get started in robotics, one can begin with a cheap robotic arm and utilize open source libraries for programming.
- Simulation tools are available to visualize and simulate robot movements and environments.
- Simulating robots is seen as the future of robotics, allowing for self-teaching and more complex tasks.

- The goal is to build a dexterous arm for assembly line tasks
- Current robot arms lack human dexterity and precision
- Two paths in robotics: research and industrial automation
- AI Grant focuses on simulation-based robotics projects
- Training in simulation involves using cameras and teaching computers to understand images
- End-to-end systems in the physical world have not been fully demonstrated
- Safety measures and constraints are actively being researched in AI
- No clear solution to the AGI problem and its potential risks has been found
- Humans have a history of mistreating less intelligent beings, raising concerns about AI ethics.

- The use of AI by humans to create damage is a more immediate threat than AI itself causing harm to humans.
- Industries are moving towards non-deterministic machine learning-based infrastructure, which makes them more susceptible to attacks by AI.
- The power grid being regulated by an algorithm instead of a human could be easier to interrupt by a smarter algorithm.
- The acquisition of Kiva by Amazon slowed down innovation in the field of mobile bases for robots.
- Scaling a factory of robots is easier than scaling a factory of humans due to hiring difficulties.
- Voodoo's vision is to move past 3D printing and onboard digital manufacturing technologies that can handle a range of materials.
- Currently, 3D printers lack the capabilities to produce most objects in a house, but eventually, there may be a future where objects can be made by a 3D printer in homes.
- 3D printing is not cost-effective for larger objects, but the next wave of 3D printed parts will likely be smaller consumer packaged goods.

- The next wave of 3D printed and digital manufactured parts will focus on smaller consumer packaged goods rather than larger objects like tables and houses.
- The cost of 3D printing large objects is currently too expensive due to the slower extrusion rate compared to injection molding machines.
- Intellectual property (IP) is not as protected in China, making it difficult to keep designs confidential when manufacturing there.
- 3D printed solar panels are not well-known or developed yet.
- 3D printing is expected to be incorporated into the educational system in the future, similar to how computers were introduced in classrooms.
- YC (Y Combinator) helped the founders of Voodoo Robotics position themselves as a robotics company and think bigger about their company's potential.
- Being a solo founder in YC is different from having a larger founding team, but Voodoo Robotics has been able to avoid common pitfalls and work well together.

- The founders of Voodoo met in college and convinced one of them to quit their job and join them.
- They pitched the idea of starting a company and convinced their co-founder to join.
- The technical capabilities of the team were engineering-focused, so they were looking for someone with a software background.
- The founders emphasize the importance of finding a co-founder who could also be a good friend.
- After YC, they faced challenges in keeping up momentum and had to focus on motivating the team.
- They suggest using PR and media as a forcing function and setting up demo days to create internal pressure and accountability.
- The founders express their goal of changing the future of manufacturing in America.

- The use of 3D printers in manufacturing is expected to create more jobs in the US and impact China negatively.
- Robots are becoming more affordable and can be integrated into workplaces.
- Collaborative robots have safety features to protect humans.
- Robot software is still developing, with the Robot Operating System (ROS) being popular.
- Integration of robots in 3D printing allows for continuous production and scalability.
- Cheaper 3D printers are preferred for reliability and running multiple jobs simultaneously.
- Starting in robotics can be done with a cheap robotic arm and open source libraries for programming.
- Simulation tools are available to visualize and simulate robot movements and environments.
- Simulating robots is seen as the future for self-teaching and complex tasks.
- The goal is to build a dexterous arm for assembly line tasks.
- Current robot arms lack human dexterity and precision.
- Two paths in robotics: research and industrial automation.
- AI Grant focuses on simulation-based robotics projects.
- Training in simulation involves using cameras and teaching computers to understand images.
- Safety measures and constraints are actively researched in AI.
- No clear solution to the AGI problem and its risks has been found.
- Concerns about AI ethics due to mistreatment of less intelligent beings in history.
- The immediate threat is humans using AI to cause damage.
- Industries moving towards non-deterministic machine learning-based infrastructure are more vulnerable to AI attacks.
- The power grid regulated by an algorithm could be interrupted by a smarter algorithm.
- Amazon's acquisition of Kiva slowed down innovation in mobile bases for robots.
- Scaling a factory of robots is easier than hiring humans.
- Voodoo Robotics aims to move past 3D printing and onboard digital manufacturing technologies.
- 3D printers currently cannot produce most objects in a house but may be able to in the future.
- The next wave of 3D printed parts will focus on smaller consumer packaged goods.
- 3D printing large objects is currently expensive compared to injection molding machines.
- Intellectual property protection is a challenge in China.
- 3D printed solar panels are not well-developed.
- 3D printing is expected to be incorporated into the educational system.
- YC helped Voodoo Robotics position themselves as a robotics company.
- Solo founders in YC face different challenges but Voodoo Robotics has worked well together.
- The importance of finding a co-founder who is also a good friend is emphasized.
- PR, media, and demo days can help maintain momentum and motivate the team.
- The founders aim to change the future of manufacturing in America.

Oliver Ortlieb and Max Friefeld are two of Voodoo Manufacturing's four cofounders - http://voodoomfg.com

Voodoo went through YC in the W17 batch. They bridge the gap between prototype and mass production with 3D printing. 

We're also joined by Daniel Gross, a Partner at YC - https://twitter.com/danielgross

YC's podcast is hosted by Craig Cannon - http://twitter.com/craigcannon

So I guess the easiest way to put it is Manufacturing using 3d printers will it create a net positive of jobs and will it be moved back to the US?

It will in the US I think it is it's a net positive for places that have seen manufacturing jobs leave their country, right?

I would say China is probably going to be in trouble.

I read a blog post recently about how Foxconn is looking to automate 30% of their workforce in the next 3 years, I think.

So that's 300,000 people that might lose their jobs due to automation.

The hope, obviously, is that things are growing, right?

They've been the low cost option, essentially.

So they're getting beaten at their own game, essentially.

And help us understand why, in particular, this is happening today.

How are the robots of today any better in terms of their technology or capabilities than the robots we've seen for a long time?

Yeah I think there's a few trends that are converging.

So just focusing on the hardware side first, we actually use a collaborative robot, UR10 it's called, and this is a $35,000 robot.

So depending on your perspective, that's either really expensive or really cheap.

It's cheap if you're used to spending money on $150,000 robots.

The other advantage of the collaborative bots is that you can just plop them into a work environment and you don't have to spend a ton of money on safety equipment.

Robots are really dangerous because if you're like working around them, it will, it'll just kill you honestly if you're in the wrong place at the wrong time.

So these are designed to accommodate that by, you know, they have force sensing and stuff like that so they don't hurt people.

This started as a research trend in Boston with this company called Rethink Robotics and they released their first arm in like 2011.

So not only did the price come down but they also kind of got easier to integrate into the workplace.

You can actually train them now just by like taking an arm and dragging it to a place and saying like move here and then like move here.

Well go into that a little bit deeper because we talked about it when you were giving me the tour.

Like I'm very curious about how you actually program an arm.

The software side of it, I think is the other interesting part of the industry, which is pretty, pretty young.

There's that Ollie can probably talk to this even more, but there's kind of like software of old with robotics, which is very rigid and structured and focused around safety.

And then there's like software of new, which is all 5 years old or less.

And most of that has been developed around this 1 system called ROS, Robot Operating System.

And ROS is really exciting for the industry because it's very accessible, you can program it, it's open source.

But it obviously has a lot of things to be improved before it's quite ready for highly structured environments where reliability is key.

I don't know, you could probably talk a little bit more about what we've been planning on doing here.

Yeah, I mean, the cooperative arms are really nice because even before you get to the level of, we need to integrate ROS to add vision or something else like that.

The UR10 is actually at the point where you just pop open a socket and send it some ASCII commands and you have it moving around immediately.

So the interface there is really easy and actually we've been able to treat most of our robotics projects to this point just as software projects.

We're definitely interested in bringing people in with robotics automation experience who can sort of handle or think about the corner cases more and things like recovering from failure.

But as far as getting a prototype together, it's basically just a software project.

And so to make this more clear, how are you guys actually integrating robots?

Because on the surface, you're doing 3D printing, and those are confined boxes.

But we're definitely adding more to the mix.

So I think the cool thing for us, it's actually a pretty straightforward application of the arm and it's something that people have been using robotics for for a while, which is just your classic machine tending.

So basically, you know, you have some high value piece of equipment that's churning out these jobs for you.

So you just have an arm sit there, and your machine is so high value that every second you can keep it producing is worth money to your company.

We actually have a slightly more exaggerated case of that, where we have 160 3D printers.

So it's not 1 arm tending 1 machine, like I think you tend to see in the traditional deployment of this technology.

We have 1 arm tending 50 or 100 machines.

So it's very valuable to us, essentially, to be able to do something like this with the arm.

And so why have you guys opted for the machines, the 3d printers that you've gone with?

And we actually used to work at MakerBot, the manufacturer of the printer.

And so when we were there, we obviously got a lot of experience with these.

They're kind of older technology at this point.

So we're definitely looking to integrate the next thing pretty soon.

And because of like, so people understand the landscape, there's a whole range of 3D printers, right?

Like in the past, I mean, when did it start?

Like 20 years ago, people started doing 3D printing?

83, I think was like the first like patent around 3d printing so in the past 30 years expensive printers like $100,000 printers the cheapest printer was maybe like $20,000 printers have been the name of the game And these are all used for prototypes.

Starting in around 2007, those original patents expired and people started developing desktop printers, which is what really spurred the 3D printing explosion of like 2012 and fall from maybe, you know, it was kind of this curve of like, Oh my God, we're really excited about 3d printing.

And then the reality set in, there's a lot of work that needs to be done before we're there.

So cheap printers, which is the desktop printers that we use are just low-end commoditized hardware.

The technology was invented a few decades ago, and people just figured out how to make them really cheaply and scale, and that's what we're built on.

And you guys have opted into those cheaper printers just so you can have more jobs running simultaneously?

Google-Amazon method to reliability, essentially.

And it sort of flips the current or old style 3d service bureau on its head.

Very valuable machines creating very valuable parts in small quantities.

For us, it's all about scale and cheap parts.

So going back to robotics for a second, if I'm listening to this podcast, today, And I'm kind of interested in getting into robotics, but I don't really know where to start.

That's a great question There are really cheap robotic arms out there that you would start with so like Craig and I were actually talking about this earlier.

You can buy like a 5 axis arm, which uses servos.

They use servos, they can lift a few hundred grams And you can just like put on your desk and have it like, you know, move your pen from 1 point to the other, you know, and you can actually obviously add more complex tasks on top of that.

So pick up 1 of those and then, start using 1 of the open source libraries out there.

Programming a robotic arm is a lot easier than maybe it was a long time ago because they handle all of the complicated stuff like how do you get from point A to point B?

You can kind of just tell it move to these coordinates in XYZ.

And then sometimes you tweak it a little bit for like how it moves there.

But the details are handled right now by software that's been developed over the last 10, 20 years.

And when you guys are programming your larger cooler arm, Are you doing any of it in simulation?

Are there any good software packages to be had there?

I think we had an intern working on that project over the summer, and I think just at the end, he discovered some simulation tools.

And yes, basically you can get a simulation that shows your environment, shows, actually this is 1 of the built-in features to ROS.

And it will let you simulate and just sort of show you exactly where the bot is, or if you're running simulations, where you think it is.

So the tooling around this, I think, has been improving.

And actually, today, it has a lot of things that we could use.

I think that's also an excellent point to bring up because that is going to be the future of robotics, simulating so that the robot can teach itself how to do things.

And, basically humans can train robots to do more kind of like objective oriented tasks, and then it can figure out the details for you.

Our goal is to eventually build a system for, let's say, assembling a product that we're making on an assembly line.

We would love an arm that is dexterous enough to replicate actions that human hands can do, and maybe just have like a video camera and have a human assemble 1, and then the arm will teach itself how to do that from a video.

I think that requires simulation is what we're working towards.

And does anyone manufacture today anything that resembles human dexterity in terms of, you know, fingers?

So there's, there's definitely some robot arms that are meant to like grab objects and they have fingers and, they're slow.

There's probably 2, 2 movements when it comes to, like grippers in the robotics world.

There's the hard, rigid, 1 or two-axi actuators.

So the hard actuators, you design to do 1 task.

And it can do that over and over again and it's cheap and simple.

The soft actuators which are not as mature are kind of designed to like pick up objects that are weird like this without breaking it.

A lot of them even work with like air sacks that just kind of like close around things.

And those really aren't dexterous enough or precise enough.

I haven't seen anything yet that is like as good as a human hand.

Following the line of training yourself to learn robotics, Where are the areas that you see developing right now that are super interesting and high value to companies where people can start contributing to open source projects?

So if I were in undergrad or getting my master's right now, I would be learning ROS and I would be focusing on research that was towards that training problem.

This is also a problem that OpenAI is working on and has published some cool videos and papers about.

If you're really just looking to get a job, there's kind of 2 flavors of robotics right now.

There's like the research side, which is what we're talking about.

And then there's the industrial automation side.

Industrial automation is about performing 1 task over and over again really efficiently and cost effectively.

Companies, most manufacturing companies will invest hundreds of thousands of dollars building a machine that just does 1 thing really efficiently.

So it's not flexible, is the main difference.

To do that, you need to be a mechanical engineer, an electrical engineer, and have a little bit of software knowledge.

And then you can become an industrial automation expert and go work for Toyota automating their next assembly line, or Tesla, maybe a little bit cooler.

There's the research path, which is probably more interesting for the intellectual folks who like to think about problems that haven't been solved yet.

And then there's the industrial automation path, which is a little bit more practical, a little bit more like classical engineer.

And so, Daniel, what's come across your desk with the AI grants related to robotics?

Yeah, so just in back color, AI Grant is a kind of decentralized research lab where people can fill out an application in about 5 minutes.

And within a couple of weeks, we let them know if we're giving them a grant for what that has both cash and $20,000 in GPU training credits.

And we've gotten almost a thousand applications for the second round of AI grant.

Many of them focus on, of the ones involved in robotics, they're almost all entirely in simulation.

The advantage to simulation is you could do this from your pajamas at home.

And they're all kind of very similar to the work OpenAI is trying to do where you want to give an agent a video feed like a human would get and try to get it to learn to do something.

You know, all of the approaches are kind of, as far as I could tell, somewhat similar to what general research is doing.

So, if you had to come across something incredibly different and weird but I'm eager to find it and help that person get their paper published if they have, if they're listening to this podcast now and have a really weird idea.

Well, what are the approaches people are using?

Because I don't think it's, I mean, AI is obviously of interest to many people, but robotics specifically within AI, like how training in a simulation even works.

Could you just walk through that for people who want to understand?

Yeah, and I'm sure there are folks that are significantly more expert in this than me, but the rough idea is even beyond a simulation, in reality, what you're doing is you set up a camera and you use about the 2014, 15 technology that we've uncovered as a species, which is we've managed to teach computers how to look inside images to get a general understanding of what the agent is looking at.

And then you'll traditionally, you'll have, ideally a human perform a task.

So, you're moving a ball from 1 hand to another, or you're moving a ball from 1 point to another.

And then there's kind of 2 scenarios here.

There's 1 where totally on a free form basis, the robot arm will try to do the same exact thing.

And it's just trying to emulate what the human hand saw.

There's another approach, which is that the robot arm has some type of goal function that it is looking to maximize.

And so it kind of flails around and realizes, okay, like randomly moving around in space, bad idea.

Because that's not increasing my goal function, you know, touching this ball seemingly slightly better, you know, moving this ball across the coordinates of the right place even better.

And so in particular, this is done with things that where you could measure very clearly the ejected function, like unscrewing a cap from a bottle.

And then you basically try to mash those 2 things together into an end-to-end system.

As far as I can tell, this truly end-to-end system, you know, that works in the physical world, I don't think has been demonstrated by anyone yet, but I may be wrong.

There was 1 gnarly, unnamed research lab that had just a goal maximizer robot arm, I think with a bottle cap.

And this thing was, it wasn't even given any image prior so it was just told try to unscrew the cap and so the arm which is quite strong is flailing around trying to unscrew the cap and it's it's not smart it's just trying to unscrew the cap you know don't don't blame it and then apparently the story so the story is told it grabs onto a researcher's arm and it's like, oh, let me try squeezing.

But I think it's an interesting example to how AI, true AI could be actually like not evil, pretty harmless.

It's just doing this other thing where it just wants to get more paper clips and then by accident destroys us all.

Yeah, hopefully the goal function isn't like take over the world, launch nuclear missiles.

I guess my point is even with a simple goal function like get more energy to run faster you could accidentally kind of cause a lot of damage.

I mean who's putting in the bounds right now?

Obviously, OpenAI is doing a ton of research around this, but yeah, I mean, when you hit a level of intelligence, maybe slightly past unscrewing someone's arm, what are the constraints people are trying to create in AI right now?

Boy, it's an active area of research and tweeting of Elon Musk.

So I mean, I think there's a lot of people working on safe AI.

I'm personally pretty skeptical on the regulation approach that people keep on proposing because it's not clear to me how you can make it global.

And I don't think that, you know, all the smart people are in America.

So I think regulation would just mean that America's last to the game.

I think there's folks working on, you know, kind of fail safe or dead man switches, something that would stop an algorithm before it starts hurting other humans.

There's people working on algorithms that are cooperative in their style.

So OpenAI and DeepMind actually published a paper on this where the training is done with a human in the loop who's actively making suggestions.

And the hypothesis, I believe, is that, well, if that would always happen, at least the human could direct the AI to be like, oh, it's bad.

And I'm sure there's plenty other approaches that I'm not even aware of.

I am personally, I haven't heard, it's quite fascinating when there's a area like this, there's an open problem, which is the AGI problem, which we're all aware of, we're all talking about, you know, over dinner, oh my gosh, what's gonna happen?

And no one's really thought of a clean answer.

And it's kind of made me wonder if there is 1 at all.

Because it's 1 of those things where I feel like if there was an obvious 1, It would have come up by now.

There's a lot of smart people thinking about it And so I've kind of resigned myself to believe that from AGI, you know killing us It's not clear to me that there is a smart answer I mean Depressingly if you go online you read stories about what we do to other species we consider less intelligent,

We're not that moral Like there's a great slash very sad Reddit thread about how people abuse characters from the Sims.

You know you like build a house or you put them in a pool and then you take out the yeah the ladder and we all thought it was hilarious but you think about it and you're like well that's kind of the same thing if you superimpose it onto a super smart AI to us.

Obviously we crush ants without giving it another thought.

I'm skeptical that we will find a way to fix that.

I am optimistic that that's not going to be what kills us as a species though.

Because I think there are far more threatening things that are closer.

For example, the use of AI by a human to create some type of nefarious damage.

So what's happening is there's this AI hype right now.

And so more and more industries are moving their infrastructure to kind of non-deterministic machine learning based stuff.

Cars are gonna be using machine learning to drive.

The electricity grid will manage itself using machine learning.

And so as a result, it'll be more susceptible to attacks from that same thing, as opposed to being manned by a human.

So the power grid is now constantly being regulated by an algorithm instead of a human.

Could that algorithm be interrupted by another smarter algorithm in a way that would have been easier than to interrupt the human?

So I suspect war will actually not be from AGI.

War and damage and carnage will be from weaponized dumb AI.

Well, to kind of like concretely connect the dots, Your robots currently right now are fixed to the ground, but they're going to start moving at some point, right?

Yeah, they're going to be put on mobile bases, not dissimilar from like an Amazon warehouse.

And So have you begun that training process?

No, we've gotten a few demos, but the hardware is actually really early.

When Amazon acquired Kiva, they just took that out of the market.

And so right around now is when people are starting to release mobile bases again, but They're all in beta.

I think there are a few that you can actually buy.

So no, we haven't gotten started on the moving problem.

So Amazon's acquisition of Kiva, you think, kind of slowed down innovation

It took a long time for everyone else to catch up.

I mean, have you been to a shipping container port ever?

There are basically trailers that are autonomous.

I don't know who is producing those, but I don't think it was Kiva.

Which is all like pick and play, pick and pack and stuff like that.

I think that's also, the shipyard example is an interesting, it's an interesting example that shows how robotics scale in a way that humans don't.

You could build a little mini shipping yard and train on that or do it digitally, right?

Do it totally as a simulation and then roll it out and start moving around like multi-ton objects very quickly.

That's just a process that doesn't scale like, it scales way more efficiently than a human process does, which is a little bit scary.

What challenges do you guys see if you wanted to scale to like

a hundred thousand square foot warehouse.

You know for us, it's actually, so right now what we're building is kind of like this footprint unit factory which has 160 printers and we'll have like 2 arms moving around doing machine tending.

And then as parts come off of the printers, they have to be post-processed, QC-ed, packed and shipped out.

I think when we automate each of those steps, they each scale very easily.

Most of the challenge is getting the first part of it automated.

Everything beyond that is kind of like, can you buy the hardware quickly enough?

See all Y Combinator transcripts on Youtube

Robotics and 3D Printing with Voodoo Manufacturing