Our earlier policies were failing on the portal with no err and stdout logs, while running perfectly fine on the local verification test. Highlighting the issues so that anyone else facing the same problems doesnt waste days trying to debug this. Huge thanks to Senne_Bogaerts for helping out:
-
task.time_limitis enforced on the sim clock, not the wall clock. If your policy usestime.time()/time.sleep()it might work fine on local but have problems in the portal -
Module import time counts against the 30-second discovery budget. If your policy has heavy imports (torch etc) move them inside
__init__. Policy-class__init__runs later inon_configure, which has a 60-second budget.
after making these changes the submission ran correctly.