[quanto]

Their video:

https://www.youtube.com/watch?v=H7dhwFcuUn0

Taken at face value, it is a very cool project. With pneumatic muscles, the biggest challenge is actually not strength but control. Fine strength/distance control is difficult. Pneumatic muscles are more for bang-bang control (on/off). The company seems to address this issue by having many "myofibers" and recruiting the fibers one by one. However, in the video, the movement is still very jerky. My limbic robots were very jerky too.

Nevertheless, very cool. The claimed 500w pump, if really done, is quite a feat. Assuming 80% efficiency, 400w output for muscles would put the robot on par with a competitive cyclist activating quadriceps, the largest muscles in the body.

[Joel_Mckay]

Indeed, the valve/line/pump mass will greatly dominate design cost.

The reason no one uses pneumatic actuators for repeatable positioning systems is gas is compressible, adiabatic thermal problems, and dangers when something fails.

Never saw automatons as particularly practical either, as the motion control planner balloons in complexity. Makes creepy looking sculptures though... =3

[Xmd5a]

>repeatable positioning systems

Typical in CNC, but is it really useful in closed-loop systems that use computer vision to adjust to an uncontrolled environment ? I've been wondering if I could reproduce something like mobile aloha but with sg90 servos instead of expansive dynamixel servos (in this case, I'm not looking at applications that require torque obviously).

[Joel_Mckay]

In general, remote operated equipment still suffers latency, and thus visual servoing doesn't work well even in ideal controlled lighting conditions.

One of the lowest-cost commercial UR5 options with force-feedback is usually packaged with a traditional turtle-bot research platform. These are safer to use around people without a cage, as you can set impact safety thresholds:

https://www.universal-robots.com/products/ur5e/

https://github.com/UniversalRobots/Universal_Robots_ROS_Driv...

The hardware is a solved problem, and there are 3D printed 5-axis and 7-axis community kits around too.

Have fun =3

[Xmd5a]

It's between 20k and 30k dollars. Maybe your profession is suffering from "industrial bias": it focuses only on high-end robotics because it's the only thing that makes sense economically for now and anything super-cheap is seen as laughable ? Wasn't it the case for FDM 3D printer too before they went mainstream in the 2010s ?

I just want to build a toy and explore what's possible with this technology for non-serious applications. The servos may wear too quickly ? I'm fine with that. However I have trouble figuring out how cheap I can go. If I can carry out a task using the "control arm" and the "working arm" performs ok, does that mean Aloha will be able to learn it ? What are the limits then ? What if my arms are really flexible (as flexible as a 10 cm section of a PLA filament for instance), but I manage to carry out the task, can something based on policy diffusion learn to handle the task with "bones" that flex and oscillate ?

[Joel_Mckay]

There are inexpensive educational products:

https://www.youtube.com/watch?v=4xD9QCBkxAs

https://www.anninrobotics.com/product-page/ar2-aluminum-and-...

https://www.omc-stepperonline.com/upgraded-ar4-robot-complet...

https://github.com/shoiland/ar4_ros_driver

YMMV as I haven't tried it myself, it looks like a solid aluminum frame with geared-stepper drivers. This means many of the axis will have minimal backlash behavior from the gears, but if you keep the driver power minimal you are less likely to strip off teeth when (not if) you make a programming error.

I also like that it is open design, as making it do something useful gets easier. You might be able to get away with a fully 3D printed plastic arm if your use case is not demanding. However, when stating your project goal in one sentence you will have to decide if you are A. building robots or B. solving some lab task.

Best of luck, one may find academic sponsorship/funding is easier than trying to recreate the wheel. Best of luck =3

[quanto]

Your intuition is quite correct that closed-loop means open-loop repeatability is not critical, but controlability (non-stickness, smallest possible controlled movement) still matters. Motors are well behaved and continuous, so even with some backlashes, you will be fine. There are also ways to compensate for backlashes from software (e.g. same approach angle for end effector).

Aloha is a fascinating project and would love to hear more about what you are thinking. Dynamixel indeed is too expensive

[Xmd5a]

First I'd like to reproduce the Push-T task of the policy diffusion paper (video here: https://diffusion-policy.cs.columbia.edu/)

Afterwards I'd like to tackle useful tasks related to gardening/botanical experiments: uprooting weeds, handling pests, harvesting small fruits. What's interesting is that you can develop new approaches to these problems. Uprooting stuff is difficult to do for a machine I guess. Maybe just cut the weeds with scissors every day, that'll teach them a lesson. Or remove aphids "by hand".

Another interesting thing is to do more scientific tasks such as handling a lot of tedious tasks on many, many plants. Example: creating polyploid plants is a lot of manual labor, what I'm talking about here is basically lab automation (doing flow cytometry on dozen or even hundred of samples).

Another aspect to explore in this space is continuous measurements (measuring photosynthesis efficiency for instance). I'm not a botanist but it seems that measuring devices either come in the form of a box you put the plant in, and you can get quasi-continuous measurements, or they are hand-held and you can only do punctual measurements (typical example: chlorophyll fluorometry). Also plants grow and change shape so putting a measuring device on a plant is in fact rather difficult. I think something like Aloha (even without the "Mobile" extension) could help tackle these situations.

[Xmd5a]

Stumbled upon this paper when exploring the topic of "visual servoing"

https://arxiv.org/abs/2208.11538

Visual Servoing in Orchard Settings

We present a general framework for accurate positioning of sensors and end effectors in farm settings using a camera mounted on a robotic manipulator. Our main contribution is a visual servoing approach based on a new and robust feature tracking algorithm. Results from field experiments performed at an apple orchard demonstrate that our approach converges to a given termination criterion even under environmental influences such as strong winds, varying illumination conditions and partial occlusion of the target object. Further, we show experimentally that the system converges to the desired view for a wide range of initial conditions. This approach opens possibilities for new applications such as automated fruit inspection, fruit picking or precise pesticide application.

[Joel_Mckay]

Or you could just try a classic library yourself:

https://visp.inria.fr/

http://wiki.ros.org/visp_ros

https://github.com/lagadic/visp

example:

https://github.com/afonsocastro/larcc_interface

YMMV =3

[Joel_Mckay]

"closed-loop means open-loop repeatability is not critical"

Hobby servos have limited repeatability especially as they wear, and when swinging an arm with shifting center of mass the errors will be quite obvious.

One may want to look at the UR5 platform design before wasting time and money. Best of luck =3

[quanto]

> Hobby servos have limited repeatability especially as they wear, and when swinging an arm with shifting center of mass the errors will be quite obvious.

Your statement is correct by itself, but it is not completely applicable to what the GP comment (not my comment) said. Aloha uses optical encoders (which themselves are far better than hobby-grade potentiometer servo encoders) and in fact also compensates for inaccuracies in the optical encoder.

I have no experience with UR5 but looks interesting. Any reason why you recommend this?

[Joel_Mckay]

"Any reason why you recommend this?"

Pre-built 3D simulation models for easier software development, official ROS support from the company, and better physical safety with advanced force-sensing motion systems.

"optical encoders" have their own set of issues... usually require quadrature encoder tracking FPGA or dedicated ASIC features in the mcu. i.e. you now have 2 problems to solve, and higher costs.

Best of luck, =3

[semi-extrinsic]

I can imagine the noise coming out of a 500w air pump has to be quite something. Imagine having a vacuum cleaner running at random intervals 24/7.

[culopatin]

And beyond that, what does maintenance look like? A few fibers get, is it done? Is it modular?