All the tests below are conducted on a desktop with RTX 3080 10GB graphics memory.
For a go2 walking on the plane task with 4096 envs, the training speed is approximately 1.3x compared to Isaac Gym.
While the graphics memory usage is roughly 1/2 compared to IsaacGym.
With this smaller memory usage, it's possible to run more parallel environments, which can further improve the training speed.
-
Simulation
For a go2 walking on the plane task, training a policy with 10000 envs for 600 ites(which is 144M steps) takes about 12mins. The play result is as below:
-
Real Robot
Also for a go2 walking on the plane task, training policy+explicit estimator with 10000 envs for 1k ites takes about 23mins. Deployment result is as below:
embedded terrain can't specify difficulty, not practical to use.
Compilation takes 2min 45s, with the below params:
Parameter | Value |
---|---|
task | go2 |
headless | False |
num_envs | 100 |
horizontal_scale | 0.1 |
vertical_scale | 0.005 |
terrain_length | 6.0 |
terrain_width | 6.0 |
border_size | 5.0 |
num_rows | 4 |
num_cols | 4 |
for headless=True with other params the same, it takes 2min 30s.
Maybe because that Genesis needs to first compile then execute, it speeds less graphics memory but takes longer time to compile.