>be me
>ruing that i didn't do mech engineering as i like robotics
>boston dynamics now being sold by google
dodged a bullet, robotics is kill
>>53633420
google is siding with the humans, truly the google overlord cares for us
>>53633420
I was initially relived too, but then I realized the obvious: they just move all the research underground because it's too creepy for normies to handle.
>>53633555
They b uy a company, form a new division inside Alphabet and then they sell the company, minus the people they actually wanted to hire to do their bidding.
Same shit happened with Motorola.
Robotics is not kill. BD is. Turns out they are just a meme company. They are also stuck in the past. They are dead set on using hydraulics, even though one of google's other companies made them obsolete
Why would anyone ever buy a company "from" google. That seems like clearly indication that it isn't worth it.
>>53633755
you are getting a brand on a discount. Like Lenovo did. You don't get the top engineers or at least not the ones making interesting stuff but you get to keep the brand.
And if you are Lenovo you go and fuck that up too.
>>53633731
What company/group, what tech? I will bump this thread until I know. I've been looking for anything that can provide the power/weight requirements that hydraulics can and I have been left with nothing. It doesn't matter what kind of servo motor you're talking about none of them compete and that is incredibly sad to me.
If you prove me wrong it will be one of the happiest moments of my life. Srs, do it.
>>53634047
bumping. this is actually important.
>>53634047
SCHAFT. Water cooled ultracapacitor driven brushless motors.
What are you making?
>>53634047
Brushless motors have all the power to weight that you need. Is 10 KW/kg good enough?
https://en.m.wikipedia.org/wiki/Power-to-weight_ratio
>>53634424
A tennis playing humanoid robot. I figure with wireless power transfer or a companion cord carrier bot-chain I wouldn't have to worry about electrical requirements to power the robot.
The problem is that I need stupid amounts of torque, speed with low weight. I've already found a solution to the locomotion issues assuming enough training time with multiple bots, ball tracking algorithms and plenty of training material for tracking humanoid motion given a 2d image by using normal tv broadcasts. I was trying to make Federer essentially. Because everyone else seems to be giving up on it.
I'll actually ask some people about making this thing.
thanks.
>>53634693
Ok look, you are trying to do something really hard and are biting off more than you can chew.
Building a humanoid that can just barely walk is hard, building one that can play tennis is pretty insane.
Second, learning from TV broadcasts has never been done and is not an easy problem. Heck, even getting the data that you'd need to do this from motion capture data of tennis players is hard. Following the motion is not enough, you need forces. Getting this from mo cap data is hard.
I would recommend that you don't make a humanoid, an omnidirectional platform with a racket on top should be enough.
Oh and get it working in simulation before you buy stuff irl.
>>53634916
>>53634693
lol, anon is out to create master thesis
>>53634916
I'll get the simulation working first, I've already started learning ROS for this so it's not a big deal.
There are a couple of white papers on calculating ball trajectories and human motion based on 2D images. If you take into consideration the dimensions of the players and the court itself it follows that you could calculate the rest of the information you need based on the angle of their movement. Approximations are what I'm looking for.
With the new kinds of motion tracking that exists, like linked below it should be easy enough to use machine learning from the various open source projects out there to make a robot mimic human motion through large amounts of training and multiple bots. There must be someone out there that has thought of combining these things together
https://neuronmocap.com/
As for biting off more than I can chew, well that's how I learn stuff. I ended up learning everything from UE4 to C++ to assembly fucking with the operating system to make my virtual reality desktop after not knowing anything:
https://www.youtube.com/watch?v=YbyUZk3TGZ8
Nearly done and releasable btw.
Anyways, thanks for the info.
>>53635665
I'm not sure who would grant any money for such a thing.
>>53636125
Literally only universities grant money for this, and only if you use it as thesis material for them. Running bipeds are still in their fledgling years, and you wanna add lunge/jump to it? It's cool and all, but you're gonna need some serious simulation on the controlability of it all.
Source: Mechatronic control engineer.
>>53633420
boston dynaimcs did cool shit way before google bought them, and will continue to do so. nothing is changing except the parent company.
>>53636125
>>There are a couple of white papers on calculating ball trajectories and human motion based on 2D images.
that is completely and totally do able. You can probably figure out exactly where the ball is and it's speed with just the right sensor and use simple ballistics calculations to figure out where it will go(computers were originally made for this).
>> it should be easy enough to use machine learning from the various open source projects out there to make a robot mimic human motion
It is not. What you really need out are forces, not motion. Motion capture only captures motions, so you have to guess what they are from the motions.
It has been done, and it is harder than it looks:
https://www.disneyresearch.com/project/controlling-humanoid-robots-with-motion-capture-data/
Optimizing to match human motion is a bad way to go. You want a robot that can play tennis, not mimic human motion(as the paper above is concerned with). The kinematics and dynamics of a robot leg are going to be different than the kinematics and dynamics of a human leg. The response of robot actuators are going to be different than human muscles. So if you want a robot that's really good at moving around fast, you should not try to copy human motion but rather find what is best for the robot.
There is a reason we don't use machine learning to teach bipeds to run and walk. Bipeds tend to break when they fall over and when you are learning to walk you fall a lot. Oh well you'll just start on smaller bipeds that won't break if they fall over. Doesn't work, because now the dynamics are different and what you learn doesn't generalize to bigger bipeds. Oh well I'll just put cushions on my bipeds. Of course, this changes the dynamics too. Do it in simulation! Again simulators aren't perfect and what you learn won't be applicable to the real world.
just make something like pic related with a racket attached
Any 1 have info on how to build EMPs for the upcoming robot war?
>>53637915
Yeah, first build a nuclear weapon. Do that and I'll tell you step two.
