Migration to 6.1 after PolicyController.py was dropped for RobotPolicyRunner .

[driven-by-code]

Hello Everyone ,
I have a quick question on how to implement/create an IO descriptors file or if it can ignored for a policy that was not trained using Isaaclab 3.0+. On 6.0.1 , I was able to override class functions and implement custom observations , actions and forward functions.
How could that be done in isaacsim 6.1 ? Would a custom class need to be implemented and not use RobotPolicyRunner ?
Right now , the solution was to copy over the respective scripts from 6.0.1 source code and inherit that in my code.
I did refer to this and it doesn't give a clear idea if the above can be done . I also checked out the labs doc for IO descriptors.
Thank you !

[PeterL-NV]

Hi, thanks for raising this. You do not need to fabricate an Isaac Lab IO_descriptors.yaml for a policy trained outside Isaac Lab. In 6.1, the deployment interface must instead be supplied by either that descriptor or an explicit PolicySpec.binding. The PolicyArtifact.from_files() API makes descriptor_path optional.
RobotPolicyRunner is intended to manage deployment and the simulation-step lifecycle; it is not a direct replacement for subclassing PolicyController and overriding its observation, action, and forward() methods. If your custom policy fits the runner’s model and configuration requirements, use PolicyArtifact and PolicySpec with an explicit binding that reproduces the exact trained observation order, preprocessing, action order, scaling, and joint mapping. If observations require state outside the robot articulation, also provide a task_state_provider.
If those hooks cannot express your existing forward() workflow, keeping a custom controller and inference loop is reasonable; you need not force it through RobotPolicyRunner.

[driven-by-code]

Hello,
Thank you for your reply !
Are there any examples or documentation to refer to creating an explicit binding for the trained observations and preprocessing for the same along with action order and joint mapping ?
Quoting from the documentation here,the PolicySpec.binding and a task_state_provider to RobotPolicyRunner is what you are speaking of ?
For an example workflow, a spot/go2 with a policy trained in mujuco lab.
Thank you once again !