General BC only needs current observations as input and outputs the actions.
Several RL/IL works will process the provided demonstrations to generate task embeddings/goal embeddings.
Is it allowed to design such models in this competition? That is, accessing provided demonstrations (download via mirerl) during inference is available?
Yes, this is allowed. You should not download it via MineRL. The demonstrations will already be present on the evaluation / inference machines in the data directory, just as they are during training.
If you need data that is generated during training (e.g. human-generated comparisons), you can save it in the same directory as the trained models. (Please don’t make any such data too large, we did not design our evaluation server to be capable of handling large datasets.)
Also remember the following part of the rules:
In addition to the provided dataset, participants may include additional small datasets in the source file submissions, whose total size should not exceed 30 MB. Pretrained models are only permitted if they were publicly available on June 4, 2021.
During the evaluation of submitted code, the individual containers will not have access to any external network in order to avoid any information leak. Relevant exceptions are added to ensure participants can download and use the pre-trained models included in popular frameworks like PyTorch and TensorFlow. Participants can request to add network exceptions for any other publicly available pre-trained models, which will be validated by AICrowd on a case-by-case basis.
All submitted code repositories will be scrubbed to remove files larger than 30MB. (It is against the rules to split a custom >30 MB dataset into multiple smaller files. This will be detected during manual review.)
Pretrained models are not allowed to have been trained on MineRL or any related or unrelated Minecraft data. The intent of this rule is to allow participants to use models which are, for example, trained on ImageNet or similar datasets.
