Flexion Robotics raises $50M from DST Global and Nvidia to build simulation-trained intelligence layer for robots

experts.

It needs an RM.

He definitely and some chrome hearts.

Definitely needs a reshard mill. Why not? Why not? Uh uh next up we got Nikita from Flexion. Excited for this one.

Thank you so much for taking time to talk to us today. Thanks for waiting. Good to meet you. [screaming] How you doing?

Hi. I'm really excited to be here.

Um so

I'm the CEO and co-founder of Flexion.

Cool.

Uh where we're building the intelligence layer or the brain that powers all kinds of robots from humanoids to mobile manipulators.

Yeah. I mean fantastic. We were just talking about humanoid robotics. uh how do you see the the market playing out?

Okay, so so so you uh hopefully you caught at least the end of our of our conversation with with Sunday robotics. But something that I was thinking about like a real challenge is when Sunday gets good enough at uh picking up, you know, manipulating objects. Uh what happens if Sunday walks up to our table here after the show? We typically have lunch and Sunday needs to figure out what's trash and what's what like what should be taken and thrown away and what's actually should just stay there, right? Because that's actually like somewhat of a like it requires some memory. It's like okay, this is an item that that is uh it needs to be able to identify objects, figure out uh what what is like what is something that is worthy of just throwing away? What is something that like I don't want thrown thrown away? And if it gets thrown away, I'll be frustrated. So, I feel like there's like a lot deeper uh uh more levels of complexity to a lot of these robotic tasks than than just like object manipulation and kind of understanding uh understanding the general environment and and really having like intelligence around the environments that it operates. And I feel like that might be something that you're solving, but uh tell me if I'm wrong or correct.

Absolutely. Uh let me just say that Sunday is amazing. I think their videos are really, really impressive. probably the most impressive uh demos I've seen so far. So, just start let's just start with that. Um I don't know if you're doing it on purpose, but it's a great reference to the video we released this morning where we have a robot walking around and picking up trash and bring it to to a garbage can.

Yeah.

And and the way we're doing that is actually splitting the problem into two parts. The first one has nothing to do with robotics. It's about um common sense and understanding. And the for that part, we don't really need to train a a specific model ourselves because that knowledge is already contained in in large language models. Think of it as GPT5 or all of these models. If you take a picture of that table in front of you and you ask GPT what is garbage, what is not, those models are already really good at understanding that. And once you have that, then the next part is actually the object manipulation. Um, which we're also solving in a slightly different way compared to Sunday. We we bet that the vast majority of data needed to train those models will come from simulation. Uh great. You can have a look at the video.

Uh yeah. So, so say more or maybe even just narrate the video.

Sure. So, let me just quickly come back. We're bad on on simulations. We train robots using reinforcement learning not to humiliate humans but to solve specific tasks and just through trial and error. So we have robots trying millions and millions of times and get tens or hundreds of years of simulated data and then they come up with very specific ways on how to walk across complex terrains but also use their whole body to manipulate objects.

But problem Yeah. Yeah. Isn't there a little bit of a problem there where uh to perfectly simulate that forest uh path uh requires incredible you know uh just like CGI just to I mean you need like Unreal Engine cranked to max on every physics calculation because yes you can model it like a video game like it's all just one smooth surface but it's not actually that in reality there's tons of different blades of grass there might be slightly more friction over here on this blade grass versus that one. You have to simulate all of that to actually recreate the real world. Is there not a gap?

Yeah, absolutely. That's a great point. Usually we call the sim to real gap.

Yes.

And once you train in simulation, the whole challenge is to cross that sim to real gap.

Mhm.

For example, here in this video, everything is trained in simulation. And we were actually not even thinking about forests or mountains, but we were training the robot.

Mhm. So you don't need to simulate every single piece of grass or every single rock. As long as you train on general enough scenarios, but somewhat encompass what is happening here.

Mhm.

Then you can deploy and and the other thing is that we're not training our policies directly from RGB camera inputs. Otherwise, you would actually need to simulate exactly how a forest looks. Mhm.

So we're doing some processing on top once again using some other models but we're actually trained on internet scale data.

Okay.

A good example I think is if you want to train a robot to open a door.

Yeah.

Um either you have to simulate every single possible door that exists in the world with all the textures, the lighting, etc.

Sure.

Or you can use a model like segment anything. Uh and then you paint the door in red and the handle in let's say green. Then all doors kind of started to look the same.

Look the same. Interesting. And then you're b you're basically training the motion against the segment anything version of the door of the world.

Yeah. Something like that.

Okay. Yeah. Uh uh what what technologies are are you most excited about across uh these generative world models uh these Gaussian splats uh just Unreal Engine getting better like traditional 3D workflows uh Houdini and Cinema 4D. uh are are of those tools which ones will be most useful to you uh in the future or or is everything kind of bespoke in its own world for you?

All of this is super important. It's all about the time frame.

So today we're using physics based simulators just like Unreal Engine.

And that's actually my take this is enough for for way more than what most people think. we can go a long way with with just those simulators

and and the logical

and and explain explain that is it is it that uh if you have a physics simulation that's running fine and let's say Unreal Engine you might use something else but Unreal Engine is do you think we're on a scaling curve where if you had a million GPUs running a million instances of of uh Unreal Engine generating simulated data that that would actually result in better progress on the on the robotics side on the actual decision-making and planning side.

Yeah, exactly. That mixed with one more thing which is generative models that can create assets for simulation.

Okay, got it.

That you don't need have humans coming up with a million different versions of all the things that the robot needs to interact with.

Yeah. So, previously that was uh programmatically like like to try and get to something with a varied world like that where there's, you know, a little hill over here and a and a rock out of place that the robot might trip over. You would have to do all that programmatically uh maybe through some nodebased workflow in Houdini or just kind of uh or or just uh just inject just randomness just random number generators and then rotate this rock over here change the geometry etc etc but you're saying that generative AI can can create even more variation is that the idea

yeah exactly you actually have two ways to add more variation [clears throat] easily one is something like gshian splats where you go outside You collect real world data.

Sure.

And suddenly you have a lot of assets. And the the second version is you ask journi to to do it for you.

Yeah. Yeah. That makes sense. That's cool. Uh

any news?

Yeah. What you got?

Give us.

Yeah. So we announced this morning that for the first time that we raised 50 million uh just

congratulations.

Uh who participated?

Thanks.

A bunch of masters. Um DST Global. Okay. Nvidia Adventures, Versipated, Proas, First Moonfire.

Awesome. And then where where are you building where are you building the company? You're uh in Europe or have you moved uh over to the West Coast?

So, right now we're all in Zurich in Switzerland.

This is why you see the robot walking in our nice Alps.

Oh, yeah.

But actually, right now, I'm I'm in San Francisco right now for a few days and I'm here to to find the right team to start a second office here. Okay, that's great. Well, good luck.

Yeah, I would, if I were you, I wouldn't. It'd be tough to leave. Uh,

pretty nice.

Switzerland completely. Second favorite country in the world for me after America. So,

uh hopefully uh next time next time I'm in uh Switzerland, I'll I'll definitely would love to stop by the office and and and meet.

Um but uh congratulations on the milestone. Super exciting. And if you