Robot Not a Dev? Buy Now
Not a Dev?

Misty Community Forum

Hardware based ROS?

This seems to be a novel concept, I am curious what the community thinks of trying to standardize components for ‘plug and play’ performance at the hardware level.

The ‘Bio-Inspired’ component seems a little too close to ‘… and then profit!’ Does anyone have a meaningful example of successfully training robotic behavior in a non-constrained environment? EG Something a hobbyist or non-researcher might be able to do?


I like the idea, but also see where it can become problematic. Modularity does have it’s benefits in terms of just connecting things and getting to implementation faster but not all sensors or actuators or HW vendors can get things to “talk” on a universal bus with out latency issues and the like.
Almost seems like i2c could handle a lot of that if each item had it’s own address etc. but I think the main issue is that there are SO many vendors and they all have diff customers that use their products in different ways based on constraints from who knows what that it wouldn’t seem like a benefit for them to support or make products specifically for that.
I think TurtleBot3 is was a closer step in this direction but more in the modularity of the 3D printable body connectors (but that isn’t the electronic hardware), and the hardware can still be whatever as long has it has a ROS package.
The other issue could potentially be the connector types, if they do have differing comm protocols how do you then create a generic connector that supports them all without then making the robot almost have a minimum size based on those requirements?
If I remember correctly, Google’s project Aurora ran into these issues which ultimately led to it’s abandonment (could be other factors, I know they had a kind of working prototype but not sure what the real cause was there) yet for the most part it was indeed a cool idea.
I think it would be great to have components like this (especially for rapid prototypes) but it seems it would have to come from a hardware vendor that can really get behind it to make is successful.

As far as I’ve seen the “train it and forget it” model isn’t quite there, it’s still very much a hybrid of the meticulous and traditional control system algorithms along side some newer ML stuff. this recent paper has some interesting results based on “observing” then doing, but it’s not really at the hobbyist level. this video is along the lines of unconstrained reenforcement learning, but the environment is still pretty controlled.

1 Like

Yeah, I find it odd that this article talks about neural nets / bio-inspired techniques for learning when their main point is that we need hardware inter-compatibility. The two are entirely disjoint. On learning, if by ‘non-constrained’ you mean out-of-the-lab, then yes, there’s lots of work that take learning systems into the field and train them. But they almost always then operate under close supervision (cause you never know what they’re going to do).

And on the hardware side, the other option to modularity is just being able to buy a robust, integrated platform. Thinking about computers, while it is possible (or at least it was when I was younger) to go out and buy a backplane, motherboard, graphics card, etc etc and put them all together (thanks to standard interfaces and connectors) to build your own computer, most people just buy a fully assembled system.

If only there were a company out there that build a robust, integrated, fully functioning robot…