unitree_rl_gym icon indicating copy to clipboard operation
unitree_rl_gym copied to clipboard

Deploy policy on Jetson Orin NX with poor performence

Open Greatsjk opened this issue 7 months ago • 2 comments

I have tried the deploy_real.py on my pc and Jetson Orin with the same code. However, the performance is hugely different. When I run this on the Orin, G1 joints are shaking and can not maintain balance. And I calculated the time of network inference, which is 0.01s in the range of control_dt (0.02s). I also changed the sleep code to
time_until_next_step = self.config.control_dt - (time.time() - step_start) if time_until_next_step > 0: time.sleep(time_until_next_step) I wonder about the cause and solution. Thank you!

Greatsjk avatar Jun 05 '25 04:06 Greatsjk

for the shaking problem, I think you might need to close the sport mode (from my experience on Go2)

ShaoshuSu avatar Jun 09 '25 17:06 ShaoshuSu

Jetson NX comes with the Tensor RT tool, which you can use to accelerate inference

M0M0ljh avatar Jul 16 '25 07:07 M0M0ljh