Simple Neural Network with hardcoded positions for walk optimisation

Question

I'm building a quadrupedal robot that will learn how to walk. From the responses I got from asking if its possible to run a NN on a micro controller I realised I needed to think of a clever system that wouldn't take 1000 years to be effective and would still be able to demonstrate onboard learning. I've designed a system but I'm not sure how effective it will be.

Firstly I hardcode 5-20 positions for the legs.

I set up a (simple) neural network where each node is a different set of positions for the legs, which I will write.

The robot moves from one node to another and the weight of the joint is determined by how far forward the robot moves.

Eventually there will be strong connections between the best nodes/positions and the robot will have found a pattern of moves that are most successful in walking.

How effective would this be in learning to walk?

Note: instead of positions I could write short gaits and the process would work out which sets work best when combined.

Answer 1

This reminds me of QWOP (seriously). You need to plan a sequence of button presses to move forward. For every situation the runner finds himself in, we need to know what button to press, how long to hold it, etc. When you play that, you probably look a lot like this video I stumbled across (pun alert): NN training bi-pedal walkers

What you need to recognize is what configuration is my robot in and you need to plan what leg should move to where? So you need a topology like this

Leg encoder values → NN → Leg desired encoder values

The difference is you need actual input in the form of sensor values, and output in the form of actions.

Note, if you had other sensors (IMU would be helpful) it would be included in the inputs, and if you had force-controlled joints, those would be your outputs.

Answer 2

Really this is a question to be answered by experiments. Will it work? It seems like it could. Two things that will be important to look at are:

1. Training time - You are still using a neural network and they take time to train. Whether this formulation would reduce the number of rounds required for training is really to be seen. It will of course change with the number of connections in the net as your agent will need to test each multiple times.
2. Training method - Based on your description it seems you are planning to use a Recurrent Neural Network (RNN) which, if memory serves, makes training more computationally intense.

There is a bit of literature on this topic already. For example a quick google reveals a paper titled A neural network model for quadruped gait generation and transitions. It may be worth looking to see what has already been tried. But then sometimes it is just fun to run the experiments.