Training on 1 task leads to 0% success rate while 10-task training works fine

[sj-li]

Hi

I was using the provided pi0 code for training. When I train the model on 10 tasks together, the results look normal and the success rates are reasonable.

However, when I train the model on only 1 task, the final success rate drops to 0%.

Multi-task (10 tasks): ✅ success, results look fine

Single-task (1 task): ❌ always 0% success

I also tried to train the model on 5 and 3 tasks, the results are reasonable.

Is this expected behavior, or could it be an issue with the implementation / hyperparameters when training on a single task? Any suggestions on how to debug this would be very helpful.

Thanks!

[JiantongChen]

We are still troubleshooting this issue, but it may not be fixed soon. If you have more information, please share it with us.

[sj-li]

Hi Jiantong,

Thanks for your quick response. Does this mean your team is also encountering the same issue?

Thanks!

[JiantongChen]

As far as I know, no one has done the same thing. I remember doing single-task training and testing on gr00t before (but not on the data for this iros competition), and there was no problem.

[sj-li]

Got it, if you can produce normal outcome, please let me know the correct command and setting to run the code.

Thanks!