>>13928023>to get the model to learnI'll set the reward for it to get to get its head high, which will mean it will need to keep its balance.
I will hang the physical and simulated robot from wires so that its knees are at about a 45 degree angle. Maybe other positions, but the movements will be slow and weak to prevent damage.
I will shift the reward to other things. Like Mountain Car. I'll see how many times I can shift the reward.
>>13928023>> the gap between the physical and simulated robot>? How are you going to collect the data for this step and how will you train it?I'll just have to try different neural net architectures. The 'make this look like / or act like that' model might be a GAN.
Like I said, I will program the Arduino to give the bot a specific sequence of movements, store the encoder and accelerometer readings. Then I will make the simulated robot perfom that sequence.
Then I will have two robots that perfomed the same movements and ended up in different places. The simulation / physical gap.
I will get a neural network to map the differences in the gap. I will deploy this model in the simulation. Now the simulation is moving like the physical robot is moving.
I know this will take many iterations. Hopefully I can get the gap mapped well if the movements don't create too much 'reflexivity' or magnification in subtle movements that result in a too much of a divergence. Maybe I'll hang the upper torso from 4 wires.
>>13928023>GazeboI've sunk a few weeks into learning MuJoCo. Hopefully Deepmind will add some docs and tutorials.
>>13928023>what specifically are you hoping to accomplish?To see how much I can get it do. Stand, walk, open doors, carry things.
Best case would be to get funding or sell something.
