also, “mapping and obstacle-avoidance AI entirely on chip” is all pure marketing built on top of the current LLM hype, it’s not remotely impressive anymore for a high end university. every autonomous drone (and your roomba) already does all of that. they didn’t develop something novel.
unless he meant that perdue actually built a chip specifically to do it with no support hardware which i doubt, since everything needed for this can be built with fairly cheap off the shelf components.
also, i googled it. they did not build their own silicon. several articles call it “from scratch.” it runs on a regular old intel nuc. so, no, it’s not even in the same ballpark as “from scratch” like they claim.
this is all so clearly some marketing bullshit built around a drone meant to hunt people.
Your Roomba and the majority of autonomous robots rely on pre mapping the environment to navigate reliably. That’s why you have to let your Roomba map out your house the first time you set it up. It’s also obviously much slower than a drone, and I don’t think they’re using any kind of AI for motion planning anyway. They also don’t keep track of a moving target. So no, not “every autonomous drone” does this.
Of course they didn’t make their own silicon. That’s an insane and stupid standard to qualify as “from scratch”. If a bakery claims their bread is made from scratch, do you ask where their wheat fields are? They’re obviously claiming the software and AI model they trained are their own original work. (I have no idea if that’s true or not, it seems like they may have “taken inspiration” from an existing open source Chinese drone.)
Running an AI model (which is not the same thing as an LLM) to do object avoidance at speed on the device in an unmapped environment is impressive and difficult.
also, “mapping and obstacle-avoidance AI entirely on chip” is all pure marketing built on top of the current LLM hype, it’s not remotely impressive anymore for a high end university. every autonomous drone (and your roomba)
Are you seriously comparing this with a roomba? Have you even seen a roomba working? They move so slowly you could implement their obstacle avoidance in Python. And they bump into things all the time! Doing this in real time, in 3D, at the speed of a runner is a completely different problem. Doing this on a drone’s own chip is actually quite impressive.
All LIDAR does is map an environment. You still have to actually do something with that data to avoid obstacles. Usually this was done with traditional control software but writing that software for something like this is really really difficult, and it wasn’t really practical to run an AI model on device like this until relatively recently.
also, “mapping and obstacle-avoidance AI entirely on chip” is all pure marketing built on top of the current LLM hype, it’s not remotely impressive anymore for a high end university. every autonomous drone (and your roomba) already does all of that. they didn’t develop something novel.
unless he meant that perdue actually built a chip specifically to do it with no support hardware which i doubt, since everything needed for this can be built with fairly cheap off the shelf components.
also, i googled it. they did not build their own silicon. several articles call it “from scratch.” it runs on a regular old intel nuc. so, no, it’s not even in the same ballpark as “from scratch” like they claim.
this is all so clearly some marketing bullshit built around a drone meant to hunt people.
Your Roomba and the majority of autonomous robots rely on pre mapping the environment to navigate reliably. That’s why you have to let your Roomba map out your house the first time you set it up. It’s also obviously much slower than a drone, and I don’t think they’re using any kind of AI for motion planning anyway. They also don’t keep track of a moving target. So no, not “every autonomous drone” does this.
Of course they didn’t make their own silicon. That’s an insane and stupid standard to qualify as “from scratch”. If a bakery claims their bread is made from scratch, do you ask where their wheat fields are? They’re obviously claiming the software and AI model they trained are their own original work. (I have no idea if that’s true or not, it seems like they may have “taken inspiration” from an existing open source Chinese drone.)
Running an AI model (which is not the same thing as an LLM) to do object avoidance at speed on the device in an unmapped environment is impressive and difficult.
Are you seriously comparing this with a roomba? Have you even seen a roomba working? They move so slowly you could implement their obstacle avoidance in Python. And they bump into things all the time! Doing this in real time, in 3D, at the speed of a runner is a completely different problem. Doing this on a drone’s own chip is actually quite impressive.
Your roomba never chased you through a forest?
The “snail that will kill you” scenario is getting more plausible every day.
Doomba
I mean… if they have LIDAR… why would they need an obstacle avoidance stochastic parrot?
All LIDAR does is map an environment. You still have to actually do something with that data to avoid obstacles. Usually this was done with traditional control software but writing that software for something like this is really really difficult, and it wasn’t really practical to run an AI model on device like this until relatively recently.