Alpha Zero - Building a Smart Rover

If one of our members could take the AlphaZero algo, adapt it for a rover, run it in a simulator, and then transfer it to a physical rover with a RPi, they would have one of the coolest projects ever, and a baseline for a really smart rover from which many could benefit.

The following are some links to get started. The algo is open source.

AlphaZero is a further evolution of the algo that beat everything else at Chess, Go, and many other games. It uses a deep neural net along with some Monte Carlo tree search stuff to make decisions, learn, etc. through trial and error and reinforcement.

I think a simulator would be necessary so it could learn in a reasonable amount of time.

The rover would need some sensors, could be sonar, sharps, camera, etc. An orientation sensor, for detecting pitch and roll would probably be a good idea too.

I would love to hear it if anyone is working on something like this. I could be interested in helping or creating incentives/contest if there was interest.

2 Likes

No direct experience with AlphaZero, but it seems that much of the development focused on Go, chess and shogi. Adapting an AI for general robotic use seems like it might need a team of knowledgeable AI programmers (effectively if Alphabet’s DeepMind decided to try that they stand a chance). There are other open source “learning” platforms available which are more directly compatible with / useful in robotics, for example:

  • Software which learns how to describe images it sees based on learning from an extensive database, which has now been applied to real-time describing of what the program “sees” through a video camera, like “Two people are riding on bicycles”. Combine this with real-time 3D imaging, and a robot may be able to determine what is happening in a scene, where the actions are (i.e. locate a cup), use speech to text to interpret what is being said, then use the cloud to provide an appropriate response (verbally and physically).
    For projects like this to start though, it (normally) takes quite a dedicated team either at a company or university.