Just run an env and visualize the human mode rendering:
python basic.py
Similar to basic, but allows you to set the env from cmd line, and it enables some keyboard control (ESC - quit, SPACE - pause, RIGHT - step env, 0 - reset env)
python less_basic.py --env=UrchinBall
or
python less_basic.py --env=LuxoBall --lcd_mode=rgb # experimental visualization mode
Simple naive training approach using a Tranformer, where LCD (~16x16) frames are flattened and used directly as tokens. For training and sampling, entire frames are produced at a time and each individual pixel is sampled.
For simplicity, and because I haven't gotten it working well yet, we do not predict or deal with continuous proprioception state information, just images
- Collect data:
python collect.py --env=Bounce --collect_n=10000 # (10k rollouts, should take 5-ish minutes)
- Train a model on that data:
python train.py --env=Bounce --datadir=rollouts/Bounce-10000.npz # trains in a few minutes on GPU, longer for harder tasks
- Visualize training and samples:
tensorboard logs/