search

Armandpl / dreamerv3

DreamerV3 + gSDE, using pytorch, on a real robot
1 stars 0 forks source link

implement RSSM in PyTorch and train it on pendulum data #2

Closed Armandpl closed 7 months ago

Armandpl commented 7 months ago

start with simulated data then move on to captured data on the pendulum to capture data on the real pendulum we could just feed it sin waves, square waves (?), sawtooth waves (?) at different frequencies and gather the observations

Armandpl commented 7 months ago

add gsde to collect_data?

Armandpl commented 7 months ago

Ok so I collected 50k steps with the MCAPWrapper on the simulated pendulum. At each step we have at_minus_1, obs, reward I trained a RSSM on it: plot

And this is how it looks like at inference time:

https://github.com/Armandpl/minidream/assets/14967758/4665139a-8617-4e3f-b8d4-e00e73ee3a32