Training loss spike in put_bottles_dustbin task

[Fanqyu]

Hi Horizon Robotics Team,

Thank you for your excellent open-source baseline for RoboTwin!

During training on the put_bottles_dustbin task, I noticed that the training loss suddenly increases sharply after reaching a certain step, then oscillates, and finally stabilizes at a relatively large value. I’d like to ask whether you have also observed this phenomenon. If so:

1. Does it affect the performance?
2. Have you figured out the reason behind this phenomenon?

I’ve attached the loss curve below. Looking forward to your response.

Best regards,
Qingyu

[linxuewu]

Hello, the divergence of the loss_depth can significantly affect the performance.

This issue is caused by training instability, and you can try reducing the loss_depth weight.

[Fanqyu]

Thank you for your reply!
Could you please suggest a reasonable range for the weight? The original loss_depth weight is set to 1.

[linxuewu]

0.1~0.3 might be a good choice.

[Fanqyu]

Thanks for your suggestion. I compared the original implementation with the new one where the loss_depth weight is set to 0.1. However, I noticed that the issue still exists — it’s only scaled down by the smaller weight rather than completely resolved.

[linxuewu]

We currently don't have a good solution for this issue, and we're not entirely sure about the exact cause. You can try the following:

1.Reduce grad_max_norm,

2.Further decrease the learning rate and loss weight,

3.Directly use the model before the loss started diverging,

4.Remove the depth loss.

We will later open-source the VLM version of SEM, which should improve training stability. Stay tuned!