We are building DroidOS — the Microsoft for robots meets Ethereum-like incentives— which leads to ever smarter infinite labor for everybody.
To succeed we need a viral loop composed of (1) ease of app development and (2) useful machine-learning and (3) viral incentives.
How do we manage the cold start? How do we build a system that is good enough to ship? How do we get droids on the platform? How do we get users so that companies will make droids?
We start with those who are both the maker and the consumer. In the vein of the Apple M1 and the Oculus we start with a devkit that let’s people make Droids at home. We focus on the robots that people use most — robot arms.
To prime this cold start there are a few ingredients:
- ease of use — making mobile robots at home is nearly impossible, we will make it 1000x easier to make a mobile robot by providing half the equation, the OS to write Droid applications by building the models
- bring your own robot — we sell robots by bundling them with our conversion kit that makes a robot into a droid but you can use your own robot at home
Application Development (DroidsOS)
When you boil down the fundamental constraints of application development it is about skill transfer.
Skill transfer is a function of a great interface. The setup needs to be simple, the cost needs to be low, and the interaction needs to be intuitive. The very best skill transfer interface I’ve seen is Mobile Aloha but it still requires deep technical knowledge to build a leader-follower setup and you have to do so for each robot.
The most approachable robot I’ve seen is this project by Tau Robotics. Even that requires a build step, a 3D printer, and knowledge to run ML.
There is a few other pre-requisites:
- Getting a robot is a function of cost and design barriers. Off the shelf robots are expensive and rolling your own is an epic project. Not to mention sourcing the right robot for the job.
- Robot-to-Droid assimilation. Once we have the hardware we have to make it smart. This will require standard interfacing between the DroidOS and robot. Similar to how drivers have allowed operating systems to adapt to any hardware in a plug and play manner we need a similar adapter pattern for robotics.
The alchemy of these three areas together will create an explosive rise in everyday robotics.
Federated Self-Learning (DroidAcademy)
As with all things, the constraint on learning is rapid and high signal feedback.
Models simply do not get high signal feedback when they learn. It’s like throwing a baby into a jungle and having it learn how to build a space station from trial and error.
No wonder training ChatGPT takes $100 million to train! The feedback loop is maximized and it cannot self-correct.
As with all great inventions we aim to solve this by re-framing the problem. There is nothing that provides faster or higher signal feedback than reality. We see all the elements that we need to create the Academy for DroidOS available in existing research today — all that’s left is synthesis.
Viral Incentives
In order to drive federated tutoring from the community we need to incentive it. There are a few ideas here but we want to reward people when they contribute similar to a blockchain-esque models. When you contribute training data that creates learning you are awarded either free usage or perhaps instead accrue karma on a crypto network to provide an exchange rate from the network to shares in the company towards a capped percentage on a logarithmic curve toward X%. That way we crowdsource remote employee and honor our customers for their contributions. We may not use crypto since it comes with overhead but a similar idea.
Closing Words
Of course, underpinning this is not only the “simple” matter of the above which are daunting tasks in and of themselves but also demonstrating universal assimilation of robots into Droids.