ZiwenZhuang / parkour

[CoRL 2023] Robot Parkour Learning
https://robot-parkour.github.io
MIT License
522 stars 92 forks source link

Distillation loss in Nov02_16-18-16_674k_distill_crawljumpjumpleap #48

Open Ekaterinador opened 2 months ago

Ekaterinador commented 2 months ago

Hello! I've been testing out a policy that I trained with your repository on the Go2 robot in real life and I've found that only crawl works really well. The good news is that the rest of the skills work in Isaac Gym (except tilt), but unfortunately they don't work in real life. I think the distillation might have been undertrained because I ran it on 32 num_envs instead of the 256 specified in the repository due to graphics card limitations. I've attached a graph of distillation steps 4 and 7 (purple and blue, respectively) below. I'd be really grateful if you could help me figure out what the distillation loss should be at the end of training for a robust launch in real life.

image

meng-zha commented 1 week ago

Hello, have you solved the problem? I also want to figure out the distillation loss should be at the end of training. I trained a policy just as the instruction, but the policy don't work well even in the isaac gym. My distillation loss is here. image