AI Learns to Walk (RL)

AI Learns to Walk, a proposal collaboration.
A good video for reference. and another You can find many other references by searching on YouTube or general search “AI reinforcement learning” “AI learns to walk”, etc.

Idea: Make an AI teach itself to walk using reinforcement learning. The AI would be a sprite (ie: stick figure) that would start off random, it would have 15 or so seconds to reach the other side of the area. Every time it does not make it there in 15 (or so) seconds it will be punished (reset, dropped off an edge, etc) and if it completes the task, it would be rewarded (ie: receiving a variable input perceived as good for instance if it has cookie it is good, if not bad)

I would really love to make this but it requires a team of people familiar with the field.

More rambling:

It would have a generation counter, every time it dies, the generation goes up marking each progression.

The AI has no boundaries on how it walks, it could find a way to squiggle across the project and it would work, eventually it would evolve more advanced, hopefully.

The AI can not:

  • Change the size of its limbs:

In an experiment done telling an AI to get to the other side of a canvas the AI sized it’s limbs very large and fell onto the objective.

ai is fun

yes indeed, wanna help?

# Step 1

Well you have to start off somewhere right?

I think before we do anything, we should create a sprite that is a single line, like this: |
We would then try to get it to “walk” or at least move. The purpose is for the AI:

Gen 1:
The AI would not know how to move.
Gen #:
The AI would do something like turning in circles to get to the end, as the striaght shape moves it.

I think it should have different inputs (move joints) and learn based on which character walked the longest

There is a SAP AI course with :snap: RL model and examples (Donkey Kong, Flappy Bird)