The policy can learn holonomous walking from scratch if not bothered too much. But if you get weird walking on the first run, try reducing the yaw vel command first. TODO other tips
## First run
- no domain rand
- no push
- plane terrain
Stop when you have a nice holonomous walk.
Resume the training on the last checkpoint of the previous run
- Add rough terrain. But not very rough at first, just a little
- Add domain randomization
- Add push