What is the appropriate amount of loss for real-world robot training?

[namecq]

In the simulation, my val loss can reach 0.05, but when I use the UR5 robot to collect data, the trained val loss can only reach around 0.3. Has this loss been successfully trained?