met g1_29dof sim2sim problem

[DavidZhaoP]

Hi,I can run sim2sim with the original policy.onnx,which can correctly control the robot to move forward and backward.But when i try to replace the policy.onnx with my own trained tmp_policy.onnx.And I directly change the depoy c++ code, from
env->alg = std::make_uniqueisaaclab::OrtRunner(policy_dir / "exported" / "policy.onnx");

to

env->alg = std::make_uniqueisaaclab::OrtRunner(policy_dir / "exported" / "tmp_policy.onnx");

besides,I did nothing to the source code. But the g1_29dof robot directly crashed.

And the "tmp_policy.onnx file is generated by the play.py of this repo.And in isaacsim the g1_29dof_robot works well! it can normally walk.

And when I change the depoy c++ code, from

env->alg = std::make_uniqueisaaclab::OrtRunner(policy_dir / "exported" / "tmp_policy.onnx");

to

when i try to replace the policy.onnx with my own trained tmp_policy.onnx.And I directly change the depoy c++ code, from
env->alg = std::make_uniqueisaaclab::OrtRunner(policy_dir / "exported" / "policy.onnx");

everything goes right! which means robot can normally walk again in mujoco!

which means the only difference is the onnx file

Can anybody help me? Thanks!!!!!!!

[Agnel-Wang]

have you changed the config in env.py? maybe the deploy.yaml files are different.

[DavidZhaoP]

have you changed the config in env.py? maybe the deploy.yaml files are different.

thanks for your help, and as i know, i did nothing to the env.py. And I just directly use the default rewards and isaaclab config to train the policy for 29dof g1 roboot, which means I totally use the default train code and deploy code of this repo's main branch. Now I try to start from a pure conda env and clone this repo to make sure everything is all the same as the main branch of this repo. And I'm quite curious about whether the main branch's isaaclab train code can pass the deploy c++ code directly or not. Or if i use the default config of 29dof g1, can the onnx generated by play.py directly pass the sim2sim test of this repo. Or I have to use the deploy.yaml generated by the train code