[Question] How do I deal with uncertainty when deploying rl policies (rsl_rl)?

[celestialdr4g0n]

Question

Library: rsl_rl
Robot: FR3; have the same dimensions with Franka
The robot is fixed and the rest parts of the robot was hidden because of annoying policies.

Data pipeline: Isaac Sim (action graphs) <-> ROS + policy
Isaac Sim version 5.0
IsaacLab version 2.2.1

I am controlling an FR3 robot with policy trained using rsl_rl. The policy is obtain from logs/.../export after run the play.py script (policy.pt and policy.onxx)

Even though I use ground truth information and the object position was the same. Sometimes, the robot could not lift the object. I attached a video here.

So, how do I deal with this uncertainty?

policy_deploy_uncertain.mp4

[RandomOakForest]

Thank you for posting this. I'll move this post to our Discussions for follow up. The grasp inconsistency you observe is likely caused by minor discrepancies between the Isaac Lab training environment and the Isaac Sim deployment settings. To mitigate you may try the following:

• Load and apply the exact same normalizer and observation preprocessing as in training.
• Standardize simulation parameters and timestep.
• Use TorchScript models for inference instead of .pt dictionary models.
• Add robustness randomization during training.