Anybody have any ideas why policies that are doing well in the sample config tests are not scoring well on submission leader board? I am working on randomizing my domain more, I am guessing that is the issue, but wondering if anyone else had any thoughts?
Yes, I am facing this issue. I asked for the logs from the aic_model container.
Same here, there are 3 randomization ranges that are not explicitly stated by the organizers:
- robot spawn position and joint configuration
- task board x, y, yaw
- number of distractor components
My guess is we will need more variety across these
For the robot spawn position, this thread is relevant:
For the joint configuration, I’d also like to know more. What we know is that we’ll see the target port from the robot’s initial TCP pose. But it would be very useful to know the randomization range of that initial TCP pose, if any. Because we’d have to consider seeing the taskboard from a range of angles (e.g. from the side) at startup. So, we’d also like to know: will the TCP start pose be randomized?
1 Like