Thank you for open-sourcing such excellent work!
I noticed that in the RoboTwin benchmark, you represent actions as end-effector increments rather than the joint angles used by default in the RoboTwin benchmark.
Could I ask if this design choice helps improve performance?