This is Stringman. An open source room scale CDPR compatible with LeRobot and designed for picking up laundry. @IlirAliu_
$1235 assembled at https://t.co/UY0T5oTlwZ
If you do that in grade school they call you ADD and if you do it when you have a family they call you a narcissist.
I'm as much of a fan of markets as anything else, but I think property is best understood as a system with a benefit coordination but an overhead cost in bookkeeping.
When things are not managed by prices, it actually isn't because it would be impossible or because the thing is too abundant, But because the overhead cost of partitioning the good and tracking ownership of it exceeds the value gained by allocating it more efficiently.
Yep, power is coming from one of the support lines.
Definitely considered the battery and a charging system and I just didn't feel confident that it would be simple and reliable and cheap enough.
I went through about 15 different wires that could have been used for the tether looking for something that was low weight, low visibility, low friction, low memory, and had enough tensile strength and current capacity for the application.
Most of the original wires were composites that I twisted together with braided fishing line, but I eventually determined that a durable flouro-polymer jacket is essential.
$8k - I'd buy the parts to build 10 more robots and send them to people who could promote them.
$12k - I'd rent a medium sized room in a coworking center for the rest of the year, install a robot in it, and let the internet drive it around for testing, promotion, and data collection purposes.
$20k I'd hire an intern for the rest of the year to help me design build and train robots.
$5k video ad promotions
$5k five more 3D printers
$5k plane tickets to present the product at two or more trade shows
Everything else would probably go to paying for groceries and stuff so I don't have to get another job.
I tried a battery powered fully self contained design a while back, but it really cuts into payload capacity.
I never tried power from the wall but motors on the platform though, but I think it would still be too much of a payload penalty.
The payload capacity of this system right now is about 4kg, and that's about the weight of all the motors, so there wouldn't be any headroom left.
I think you also benefit from having an overhead camera or two for target finding.
@chrisbolas If the motion controller is running, it will detect the tension overload at about 16 Newtons and relax the lines. If nothing is connected (motors just holding position) the knots will pull out at about 50 newtons.
Sounds a little bit like a light wind in a canyon. I don't have a quantitative answer. about the same as a computer fan.
there's no fan on it, it's just the noise of the FOC controller holding position and conducting those small vibrations into the frame of the robot.
It can definitely be improved too by isolating the anchor from the wall with a little foam tape, and I think most actuators with reducers could be made quieter as well by damping and isolation
I think there's a couple of overlapping types of failure here.
First, in order to make the model reactive to the cameras, I lowered the time horizon to about 1 second. And within a 1 second of video a lot of the tasks are indistinguishable. just moving towards something mostly.
Second, the action space includes gripper velocity in the frame of reference of the gripper camera. So naturally, it's easy to learn to move towards things in the gripper camera, but hard to learn to move towards them in the overhead camera.
If I had made the action space velocity in the overhead cameras frame of reference, it would have been the reverse I think.
I have tried making the network learn to predict velocities in both frames of reference but then I get too lousy signals instead of one good one.
I think I'll just have to rotate and warp one of the images for the model so that they're in the same world frame.
I tried to put too much variety in the dataset again. Saw a little success with DiT and thought "this model can handle everything!" So I trained it on 250 eps of all kinds of random tasks. and just hoped for the best.
The model learned to ignore the prompt and generate average movement examples. It also failed to make any cross room traversals.
I'm going to go back to training a single cam grasp only model again because that worked.
This is Stringman. An open source room scale CDPR compatible with LeRobot and designed for picking up laundry. @IlirAliu_
$1235 assembled at https://t.co/UY0T5oTlwZ