Open Pansamic opened 1 month ago
Sorry for the late reply. There is something wrong with the export script, we already fixed it. Please try again. And please notice that in this implementation, projected_gravity is a unit vector only indicates direction, not similar to (0,0,-9.81). Please also make sure that you do all the pre-process of observations exactly the same as _computeobservations.
Thanks for your fix. My deployment works, but there are still some little problems.
hip_reduction
is set to 0.1.I tried python deployment but it's too slow and only 20Hz control frequency (actually I don't know how to optimize python code). So I export ONNX model as torchcript export
policy training
interation 1500
play policy in simulation
unitree go1 was walking nicely in environment.
step observation order and information
total 45 numbers
model input information
using exported torchscript. model input tensor size : 45 * 6 = 270 6 step observations, from lastest to oldest, the first 45 values is the latest observation.
unexpected policy output
real robot condition: stand on ground and model output was not executed on real robot; velocity commands were all set to 0.
actions
increased very quickly and became very large the following is the observations.